{
  "version": "2026-09-29",
  "description": "Inference precision and runtime support for every EdgeAIStack catalog platform, sourced from the vendors’ own documentation. For each platform: the precisions and runtimes it is listed as running, in catalog order, and every decided value with its evidence class — class A where a vendor document settles it, supported or not supported, with the document, its url, the date it was accessed and a verbatim quote; class E where no vendor document settles it (or the vendor’s documents conflict), with the gap stated and no source. Every engine precision is decided for every platform. A listed INT4 means the platform runs 4-bit weight quantization (W4A16 or A8W4); a precision the hardware reaches only in emulation is not listed. The lists are per platform, not per model family. Compute Fit’s precision and runtime checks and the Hardware Selector’s precision filter read these lists and cite these sources.",
  "license": {
    "name": "CC BY 4.0",
    "url": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "EdgeAIStack (https://edgeaistack.ai/datasets/)",
    "scope": "Covers EdgeAIStack's compilation and the fields EdgeAIStack wrote. Text quoted from vendor documents remains its publisher's; each quote names its source."
  },
  "last_updated": "2026-09-29",
  "confidence": "high",
  "evidence_class": "A",
  "generated_from": "edgeaistack-api/src/v2/data/precision_support_data.js by scripts/precompute/export-precision-support-dataset.mjs",
  "schema": {
    "platforms.<id>": "{ platform_id, label, summary, precisions { supported[], list_class, decided.<precision> }, runtimes { supported[], list_class, decided.<runtime> } }",
    "decided.<value>": "class A: { supported, class: \"A\", sources[] } — a vendor document says the platform runs the value, or says it does not; class E: { supported: false, class: \"E\", note } — no vendor document settles it",
    "sources[]": "{ publisher, document (title), doc_id, revision, section, url, retrieved (date accessed), verified, quote (verbatim, at most 25 words), class }",
    "documents.<key>": "each vendor document once, with the same fields and backs[] — the platform facts it is cited for (\"<platform> <precision|runtime> <value> supported|not supported\")",
    "list_class": "A when every listed value is vendor-documented, else E"
  },
  "vocabulary": {
    "precisions": [
      "fp32",
      "fp16",
      "int8",
      "bf16",
      "fp8",
      "int4",
      "fp4"
    ],
    "runtimes": [
      "tensorrt",
      "pytorch",
      "onnx",
      "llama_cpp",
      "edge_tpu",
      "hailo_sdk",
      "rknn",
      "openvino"
    ]
  },
  "coverage": {
    "platforms": 20,
    "decisions": 200,
    "class_a": 182,
    "class_a_supported": 116,
    "class_a_not_supported": 66,
    "class_e": 18,
    "documents": 65,
    "citations": 405
  },
  "platforms": {
    "jetson_orin_nano": {
      "platform_id": "jetson_orin_nano",
      "label": "Jetson Orin Nano 8GB",
      "summary": "Jetson Orin Nano 8GB: precisions FP32, FP16, INT8 and INT4; runtimes TensorRT, PyTorch, ONNX Runtime and llama.cpp — every listed value documented by its vendor (class A: NVIDIA, Microsoft, ggml-org). Documented as not supported (class A): precisions FP8 and FP4. Not listed, and no vendor document settles it (class E): precision BF16.",
      "precisions": {
        "supported": [
          "fp32",
          "fp16",
          "int8",
          "int4"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": false,
            "class": "E",
            "note": "Contested between NVIDIA documents: the TensorRT support matrix marks BF16 N/A on compute capability 8.7, while the Jetson Orin NX data sheet lists bfloat16 among the Tensor Core precisions and TensorRT Edge-LLM lists a BF16 build of one model for Orin. No vendor statement settles BF16 model deployment on Orin through the catalog runtimes, so the exclusion stays a catalog assertion."
          },
          "fp8": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "FP8 | 2x reduction | Not available | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "INT4 AWQ | 4x reduction | Supported | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "fp4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVFP4 | 4x reduction | Not available | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "tensorrt",
          "pytorch",
          "onnx",
          "llama_cpp"
        ],
        "list_class": "A",
        "decided": {
          "tensorrt": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "JetPack 6.2.1 Release Notes",
                "doc_id": "jetpack-6.2.1-release-notes",
                "revision": "JetPack 6.2.1",
                "section": "Additional Release Details",
                "url": "https://docs.nvidia.com/jetson/jetpack/release-notes/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Compute stack with CUDA 12.6, TensorRT 10.3, cuDNN 9.3, VPI 3.2, DLA 3.1, and DLFW 24.0.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "JetPack 6.2.1 Release Notes",
                "doc_id": "jetpack-6.2.1-release-notes",
                "revision": "JetPack 6.2.1",
                "section": "New Features",
                "url": "https://docs.nvidia.com/jetson/jetpack/release-notes/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVIDIA® Jetson™ Linux 36.4.4 supports all NVIDIA Jetson Orin modules and developer kits",
                "class": "A"
              }
            ]
          },
          "pytorch": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "Installing PyTorch for Jetson Platform",
                "doc_id": "nvidia-install-pytorch-jetson",
                "revision": "page updated 2026-09-08",
                "section": "1. Overview > Jetson AGX Orin",
                "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The NVIDIA Jetson AGX Orin Developer Kit includes a high-performance, power-efficient Jetson AGX Orin module, and can emulate the other Jetson modules.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "PyTorch for Jetson Platform — release notes, compatibility table",
                "doc_id": "nvidia-pytorch-jetson-release-notes",
                "revision": "page updated 2026-09-08",
                "section": "Compatibility (PyTorch version | NVIDIA container | … | JetPack)",
                "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform-release-notes/pytorch-jetson-rel.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "2.8.0a0+5228986c39 | 25.06 | - | 6.2 |",
                "class": "A"
              }
            ]
          },
          "onnx": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Microsoft",
                "document": "ONNX Runtime — Build with different execution providers",
                "doc_id": "onnxruntime-build-eps",
                "section": "TensorRT execution provider: NVIDIA Jetson",
                "url": "https://onnxruntime.ai/docs/build/eps.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVIDIA Jetson TX1/TX2/Nano/Xavier/Orin",
                "class": "A"
              }
            ]
          },
          "llama_cpp": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "Gemma 4 on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-gemma4",
                "section": "Overview",
                "url": "https://www.jetson-ai-lab.com/tutorials/gemma4-on-jetson",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The full family is supported on Jetson through both vLLM and llama.cpp. All of the models are supported on Orin and Thor,",
                "class": "A"
              },
              {
                "publisher": "ggml-org",
                "document": "llama.cpp documentation — MobileVLM",
                "doc_id": "llama-cpp-docs-mobilevlm",
                "revision": "master, 2026-09-29",
                "section": "Orin compile and run",
                "url": "https://raw.githubusercontent.com/ggml-org/llama.cpp/master/docs/multimodal/MobileVLM.md",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "- [x] run MobileVLM on `Jetson Orin`",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "jetson_orin_nano_super": {
      "platform_id": "jetson_orin_nano_super",
      "label": "Jetson Orin Nano Super",
      "summary": "Jetson Orin Nano Super: precisions FP32, FP16, INT8 and INT4; runtimes TensorRT, PyTorch, ONNX Runtime and llama.cpp — every listed value documented by its vendor (class A: NVIDIA, Microsoft, ggml-org). Documented as not supported (class A): precisions FP8 and FP4. Not listed, and no vendor document settles it (class E): precision BF16.",
      "precisions": {
        "supported": [
          "fp32",
          "fp16",
          "int8",
          "int4"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Orin Nano Developer Kit Gets a Super Boost",
                "doc_id": "nvidia-blog-orin-nano-super",
                "revision": "Dec 17, 2024",
                "section": "Introduction",
                "url": "https://developer.nvidia.com/blog/nvidia-jetson-orin-nano-developer-kit-gets-a-super-boost/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "With the same hardware architecture, this performance boost is enabled by a new power mode which increases the GPU, memory, and CPU clocks.",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Orin Nano Developer Kit Gets a Super Boost",
                "doc_id": "nvidia-blog-orin-nano-super",
                "revision": "Dec 17, 2024",
                "section": "Introduction",
                "url": "https://developer.nvidia.com/blog/nvidia-jetson-orin-nano-developer-kit-gets-a-super-boost/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "With the same hardware architecture, this performance boost is enabled by a new power mode which increases the GPU, memory, and CPU clocks.",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Orin Nano Developer Kit Gets a Super Boost",
                "doc_id": "nvidia-blog-orin-nano-super",
                "revision": "Dec 17, 2024",
                "section": "Introduction",
                "url": "https://developer.nvidia.com/blog/nvidia-jetson-orin-nano-developer-kit-gets-a-super-boost/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "With the same hardware architecture, this performance boost is enabled by a new power mode which increases the GPU, memory, and CPU clocks.",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": false,
            "class": "E",
            "note": "Contested between NVIDIA documents: the TensorRT support matrix marks BF16 N/A on compute capability 8.7, while the Jetson Orin NX data sheet lists bfloat16 among the Tensor Core precisions and TensorRT Edge-LLM lists a BF16 build of one model for Orin. No vendor statement settles BF16 model deployment on Orin through the catalog runtimes, so the exclusion stays a catalog assertion."
          },
          "fp8": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "FP8 | 2x reduction | Not available | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Orin Nano Developer Kit Gets a Super Boost",
                "doc_id": "nvidia-blog-orin-nano-super",
                "revision": "Dec 17, 2024",
                "section": "Introduction",
                "url": "https://developer.nvidia.com/blog/nvidia-jetson-orin-nano-developer-kit-gets-a-super-boost/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "With the same hardware architecture, this performance boost is enabled by a new power mode which increases the GPU, memory, and CPU clocks.",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "INT4 AWQ | 4x reduction | Supported | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Orin Nano Developer Kit Gets a Super Boost",
                "doc_id": "nvidia-blog-orin-nano-super",
                "revision": "Dec 17, 2024",
                "section": "Introduction",
                "url": "https://developer.nvidia.com/blog/nvidia-jetson-orin-nano-developer-kit-gets-a-super-boost/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "With the same hardware architecture, this performance boost is enabled by a new power mode which increases the GPU, memory, and CPU clocks.",
                "class": "A"
              }
            ]
          },
          "fp4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVFP4 | 4x reduction | Not available | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Orin Nano Developer Kit Gets a Super Boost",
                "doc_id": "nvidia-blog-orin-nano-super",
                "revision": "Dec 17, 2024",
                "section": "Introduction",
                "url": "https://developer.nvidia.com/blog/nvidia-jetson-orin-nano-developer-kit-gets-a-super-boost/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "With the same hardware architecture, this performance boost is enabled by a new power mode which increases the GPU, memory, and CPU clocks.",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "tensorrt",
          "pytorch",
          "onnx",
          "llama_cpp"
        ],
        "list_class": "A",
        "decided": {
          "tensorrt": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "JetPack 6.2.1 Release Notes",
                "doc_id": "jetpack-6.2.1-release-notes",
                "revision": "JetPack 6.2.1",
                "section": "Additional Release Details",
                "url": "https://docs.nvidia.com/jetson/jetpack/release-notes/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Compute stack with CUDA 12.6, TensorRT 10.3, cuDNN 9.3, VPI 3.2, DLA 3.1, and DLFW 24.0.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "JetPack 6.2.1 Release Notes",
                "doc_id": "jetpack-6.2.1-release-notes",
                "revision": "JetPack 6.2.1",
                "section": "New Features",
                "url": "https://docs.nvidia.com/jetson/jetpack/release-notes/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVIDIA® Jetson™ Linux 36.4.4 supports all NVIDIA Jetson Orin modules and developer kits",
                "class": "A"
              }
            ]
          },
          "pytorch": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "Installing PyTorch for Jetson Platform",
                "doc_id": "nvidia-install-pytorch-jetson",
                "revision": "page updated 2026-09-08",
                "section": "1. Overview > Jetson AGX Orin",
                "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The NVIDIA Jetson AGX Orin Developer Kit includes a high-performance, power-efficient Jetson AGX Orin module, and can emulate the other Jetson modules.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "PyTorch for Jetson Platform — release notes, compatibility table",
                "doc_id": "nvidia-pytorch-jetson-release-notes",
                "revision": "page updated 2026-09-08",
                "section": "Compatibility (PyTorch version | NVIDIA container | … | JetPack)",
                "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform-release-notes/pytorch-jetson-rel.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "2.8.0a0+5228986c39 | 25.06 | - | 6.2 |",
                "class": "A"
              }
            ]
          },
          "onnx": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Microsoft",
                "document": "ONNX Runtime — Build with different execution providers",
                "doc_id": "onnxruntime-build-eps",
                "section": "TensorRT execution provider: NVIDIA Jetson",
                "url": "https://onnxruntime.ai/docs/build/eps.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVIDIA Jetson TX1/TX2/Nano/Xavier/Orin",
                "class": "A"
              }
            ]
          },
          "llama_cpp": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "Gemma 4 on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-gemma4",
                "section": "Overview",
                "url": "https://www.jetson-ai-lab.com/tutorials/gemma4-on-jetson",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The full family is supported on Jetson through both vLLM and llama.cpp. All of the models are supported on Orin and Thor,",
                "class": "A"
              },
              {
                "publisher": "ggml-org",
                "document": "llama.cpp documentation — MobileVLM",
                "doc_id": "llama-cpp-docs-mobilevlm",
                "revision": "master, 2026-09-29",
                "section": "Orin compile and run",
                "url": "https://raw.githubusercontent.com/ggml-org/llama.cpp/master/docs/multimodal/MobileVLM.md",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "- [x] run MobileVLM on `Jetson Orin`",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "jetson_orin_nx_8gb": {
      "platform_id": "jetson_orin_nx_8gb",
      "label": "Jetson Orin NX 8GB",
      "summary": "Jetson Orin NX 8GB: precisions FP32, FP16, INT8 and INT4; runtimes TensorRT, PyTorch, ONNX Runtime and llama.cpp — every listed value documented by its vendor (class A: NVIDIA, Microsoft, ggml-org). Documented as not supported (class A): precisions FP8 and FP4. Not listed, and no vendor document settles it (class E): precision BF16.",
      "precisions": {
        "supported": [
          "fp32",
          "fp16",
          "int8",
          "int4"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": false,
            "class": "E",
            "note": "Contested between NVIDIA documents: the TensorRT support matrix marks BF16 N/A on compute capability 8.7, while the Jetson Orin NX data sheet lists bfloat16 among the Tensor Core precisions and TensorRT Edge-LLM lists a BF16 build of one model for Orin. No vendor statement settles BF16 model deployment on Orin through the catalog runtimes, so the exclusion stays a catalog assertion."
          },
          "fp8": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "FP8 | 2x reduction | Not available | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "INT4 AWQ | 4x reduction | Supported | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "fp4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVFP4 | 4x reduction | Not available | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "tensorrt",
          "pytorch",
          "onnx",
          "llama_cpp"
        ],
        "list_class": "A",
        "decided": {
          "tensorrt": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "JetPack 6.2.1 Release Notes",
                "doc_id": "jetpack-6.2.1-release-notes",
                "revision": "JetPack 6.2.1",
                "section": "Additional Release Details",
                "url": "https://docs.nvidia.com/jetson/jetpack/release-notes/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Compute stack with CUDA 12.6, TensorRT 10.3, cuDNN 9.3, VPI 3.2, DLA 3.1, and DLFW 24.0.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "JetPack 6.2.1 Release Notes",
                "doc_id": "jetpack-6.2.1-release-notes",
                "revision": "JetPack 6.2.1",
                "section": "New Features",
                "url": "https://docs.nvidia.com/jetson/jetpack/release-notes/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVIDIA® Jetson™ Linux 36.4.4 supports all NVIDIA Jetson Orin modules and developer kits",
                "class": "A"
              }
            ]
          },
          "pytorch": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "Installing PyTorch for Jetson Platform",
                "doc_id": "nvidia-install-pytorch-jetson",
                "revision": "page updated 2026-09-08",
                "section": "1. Overview > Jetson AGX Orin",
                "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The NVIDIA Jetson AGX Orin Developer Kit includes a high-performance, power-efficient Jetson AGX Orin module, and can emulate the other Jetson modules.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "PyTorch for Jetson Platform — release notes, compatibility table",
                "doc_id": "nvidia-pytorch-jetson-release-notes",
                "revision": "page updated 2026-09-08",
                "section": "Compatibility (PyTorch version | NVIDIA container | … | JetPack)",
                "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform-release-notes/pytorch-jetson-rel.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "2.8.0a0+5228986c39 | 25.06 | - | 6.2 |",
                "class": "A"
              }
            ]
          },
          "onnx": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Microsoft",
                "document": "ONNX Runtime — Build with different execution providers",
                "doc_id": "onnxruntime-build-eps",
                "section": "TensorRT execution provider: NVIDIA Jetson",
                "url": "https://onnxruntime.ai/docs/build/eps.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVIDIA Jetson TX1/TX2/Nano/Xavier/Orin",
                "class": "A"
              }
            ]
          },
          "llama_cpp": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "Gemma 4 on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-gemma4",
                "section": "Overview",
                "url": "https://www.jetson-ai-lab.com/tutorials/gemma4-on-jetson",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The full family is supported on Jetson through both vLLM and llama.cpp. All of the models are supported on Orin and Thor,",
                "class": "A"
              },
              {
                "publisher": "ggml-org",
                "document": "llama.cpp documentation — MobileVLM",
                "doc_id": "llama-cpp-docs-mobilevlm",
                "revision": "master, 2026-09-29",
                "section": "Orin compile and run",
                "url": "https://raw.githubusercontent.com/ggml-org/llama.cpp/master/docs/multimodal/MobileVLM.md",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "- [x] run MobileVLM on `Jetson Orin`",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "jetson_orin_nx": {
      "platform_id": "jetson_orin_nx",
      "label": "Jetson Orin NX 16GB",
      "summary": "Jetson Orin NX 16GB: precisions FP32, FP16, INT8 and INT4; runtimes TensorRT, PyTorch, ONNX Runtime and llama.cpp — every listed value documented by its vendor (class A: NVIDIA, Microsoft, ggml-org). Documented as not supported (class A): precisions FP8 and FP4. Not listed, and no vendor document settles it (class E): precision BF16.",
      "precisions": {
        "supported": [
          "fp32",
          "fp16",
          "int8",
          "int4"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": false,
            "class": "E",
            "note": "Contested between NVIDIA documents: the TensorRT support matrix marks BF16 N/A on compute capability 8.7, while the Jetson Orin NX data sheet lists bfloat16 among the Tensor Core precisions and TensorRT Edge-LLM lists a BF16 build of one model for Orin. No vendor statement settles BF16 model deployment on Orin through the catalog runtimes, so the exclusion stays a catalog assertion."
          },
          "fp8": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "FP8 | 2x reduction | Not available | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "INT4 AWQ | 4x reduction | Supported | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "fp4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVFP4 | 4x reduction | Not available | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "tensorrt",
          "pytorch",
          "onnx",
          "llama_cpp"
        ],
        "list_class": "A",
        "decided": {
          "tensorrt": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "JetPack 6.2.1 Release Notes",
                "doc_id": "jetpack-6.2.1-release-notes",
                "revision": "JetPack 6.2.1",
                "section": "Additional Release Details",
                "url": "https://docs.nvidia.com/jetson/jetpack/release-notes/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Compute stack with CUDA 12.6, TensorRT 10.3, cuDNN 9.3, VPI 3.2, DLA 3.1, and DLFW 24.0.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "JetPack 6.2.1 Release Notes",
                "doc_id": "jetpack-6.2.1-release-notes",
                "revision": "JetPack 6.2.1",
                "section": "New Features",
                "url": "https://docs.nvidia.com/jetson/jetpack/release-notes/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVIDIA® Jetson™ Linux 36.4.4 supports all NVIDIA Jetson Orin modules and developer kits",
                "class": "A"
              }
            ]
          },
          "pytorch": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "Installing PyTorch for Jetson Platform",
                "doc_id": "nvidia-install-pytorch-jetson",
                "revision": "page updated 2026-09-08",
                "section": "1. Overview > Jetson AGX Orin",
                "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The NVIDIA Jetson AGX Orin Developer Kit includes a high-performance, power-efficient Jetson AGX Orin module, and can emulate the other Jetson modules.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "PyTorch for Jetson Platform — release notes, compatibility table",
                "doc_id": "nvidia-pytorch-jetson-release-notes",
                "revision": "page updated 2026-09-08",
                "section": "Compatibility (PyTorch version | NVIDIA container | … | JetPack)",
                "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform-release-notes/pytorch-jetson-rel.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "2.8.0a0+5228986c39 | 25.06 | - | 6.2 |",
                "class": "A"
              }
            ]
          },
          "onnx": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Microsoft",
                "document": "ONNX Runtime — Build with different execution providers",
                "doc_id": "onnxruntime-build-eps",
                "section": "TensorRT execution provider: NVIDIA Jetson",
                "url": "https://onnxruntime.ai/docs/build/eps.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVIDIA Jetson TX1/TX2/Nano/Xavier/Orin",
                "class": "A"
              }
            ]
          },
          "llama_cpp": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "Gemma 4 on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-gemma4",
                "section": "Overview",
                "url": "https://www.jetson-ai-lab.com/tutorials/gemma4-on-jetson",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The full family is supported on Jetson through both vLLM and llama.cpp. All of the models are supported on Orin and Thor,",
                "class": "A"
              },
              {
                "publisher": "ggml-org",
                "document": "llama.cpp documentation — MobileVLM",
                "doc_id": "llama-cpp-docs-mobilevlm",
                "revision": "master, 2026-09-29",
                "section": "Orin compile and run",
                "url": "https://raw.githubusercontent.com/ggml-org/llama.cpp/master/docs/multimodal/MobileVLM.md",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "- [x] run MobileVLM on `Jetson Orin`",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "jetson_agx_orin_32gb": {
      "platform_id": "jetson_agx_orin_32gb",
      "label": "Jetson AGX Orin 32GB",
      "summary": "Jetson AGX Orin 32GB: precisions FP32, FP16, INT8 and INT4; runtimes TensorRT, PyTorch, ONNX Runtime and llama.cpp — every listed value documented by its vendor (class A: NVIDIA, Microsoft, ggml-org). Documented as not supported (class A): precisions FP8 and FP4. Not listed, and no vendor document settles it (class E): precision BF16.",
      "precisions": {
        "supported": [
          "fp32",
          "fp16",
          "int8",
          "int4"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": false,
            "class": "E",
            "note": "Contested between NVIDIA documents: the TensorRT support matrix marks BF16 N/A on compute capability 8.7, while the Jetson Orin NX data sheet lists bfloat16 among the Tensor Core precisions and TensorRT Edge-LLM lists a BF16 build of one model for Orin. No vendor statement settles BF16 model deployment on Orin through the catalog runtimes, so the exclusion stays a catalog assertion."
          },
          "fp8": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "FP8 | 2x reduction | Not available | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "INT4 AWQ | 4x reduction | Supported | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "fp4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVFP4 | 4x reduction | Not available | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "tensorrt",
          "pytorch",
          "onnx",
          "llama_cpp"
        ],
        "list_class": "A",
        "decided": {
          "tensorrt": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "JetPack 6.2.1 Release Notes",
                "doc_id": "jetpack-6.2.1-release-notes",
                "revision": "JetPack 6.2.1",
                "section": "Additional Release Details",
                "url": "https://docs.nvidia.com/jetson/jetpack/release-notes/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Compute stack with CUDA 12.6, TensorRT 10.3, cuDNN 9.3, VPI 3.2, DLA 3.1, and DLFW 24.0.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "JetPack 6.2.1 Release Notes",
                "doc_id": "jetpack-6.2.1-release-notes",
                "revision": "JetPack 6.2.1",
                "section": "New Features",
                "url": "https://docs.nvidia.com/jetson/jetpack/release-notes/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVIDIA® Jetson™ Linux 36.4.4 supports all NVIDIA Jetson Orin modules and developer kits",
                "class": "A"
              }
            ]
          },
          "pytorch": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "Installing PyTorch for Jetson Platform",
                "doc_id": "nvidia-install-pytorch-jetson",
                "revision": "page updated 2026-09-08",
                "section": "1. Overview > Jetson AGX Orin",
                "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The NVIDIA Jetson AGX Orin Developer Kit includes a high-performance, power-efficient Jetson AGX Orin module, and can emulate the other Jetson modules.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "PyTorch for Jetson Platform — release notes, compatibility table",
                "doc_id": "nvidia-pytorch-jetson-release-notes",
                "revision": "page updated 2026-09-08",
                "section": "Compatibility (PyTorch version | NVIDIA container | … | JetPack)",
                "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform-release-notes/pytorch-jetson-rel.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "2.8.0a0+5228986c39 | 25.06 | - | 6.2 |",
                "class": "A"
              }
            ]
          },
          "onnx": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Microsoft",
                "document": "ONNX Runtime — Build with different execution providers",
                "doc_id": "onnxruntime-build-eps",
                "section": "TensorRT execution provider: NVIDIA Jetson",
                "url": "https://onnxruntime.ai/docs/build/eps.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVIDIA Jetson TX1/TX2/Nano/Xavier/Orin",
                "class": "A"
              }
            ]
          },
          "llama_cpp": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "Gemma 4 on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-gemma4",
                "section": "Overview",
                "url": "https://www.jetson-ai-lab.com/tutorials/gemma4-on-jetson",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The full family is supported on Jetson through both vLLM and llama.cpp. All of the models are supported on Orin and Thor,",
                "class": "A"
              },
              {
                "publisher": "ggml-org",
                "document": "llama.cpp documentation — MobileVLM",
                "doc_id": "llama-cpp-docs-mobilevlm",
                "revision": "master, 2026-09-29",
                "section": "Orin compile and run",
                "url": "https://raw.githubusercontent.com/ggml-org/llama.cpp/master/docs/multimodal/MobileVLM.md",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "- [x] run MobileVLM on `Jetson Orin`",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "jetson_agx_orin": {
      "platform_id": "jetson_agx_orin",
      "label": "Jetson AGX Orin 64GB",
      "summary": "Jetson AGX Orin 64GB: precisions FP32, FP16, INT8 and INT4; runtimes TensorRT, PyTorch, ONNX Runtime and llama.cpp — every listed value documented by its vendor (class A: NVIDIA, Microsoft, ggml-org). Documented as not supported (class A): precisions FP8 and FP4. Not listed, and no vendor document settles it (class E): precision BF16.",
      "precisions": {
        "supported": [
          "fp32",
          "fp16",
          "int8",
          "int4"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": false,
            "class": "E",
            "note": "Contested between NVIDIA documents: the TensorRT support matrix marks BF16 N/A on compute capability 8.7, while the Jetson Orin NX data sheet lists bfloat16 among the Tensor Core precisions and TensorRT Edge-LLM lists a BF16 build of one model for Orin. No vendor statement settles BF16 model deployment on Orin through the catalog runtimes, so the exclusion stays a catalog assertion."
          },
          "fp8": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "FP8 | 2x reduction | Not available | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "INT4 AWQ | 4x reduction | Supported | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          },
          "fp4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.7",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVFP4 | 4x reduction | Not available | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.7 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "tensorrt",
          "pytorch",
          "onnx",
          "llama_cpp"
        ],
        "list_class": "A",
        "decided": {
          "tensorrt": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "JetPack 6.2.1 Release Notes",
                "doc_id": "jetpack-6.2.1-release-notes",
                "revision": "JetPack 6.2.1",
                "section": "Additional Release Details",
                "url": "https://docs.nvidia.com/jetson/jetpack/release-notes/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Compute stack with CUDA 12.6, TensorRT 10.3, cuDNN 9.3, VPI 3.2, DLA 3.1, and DLFW 24.0.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "JetPack 6.2.1 Release Notes",
                "doc_id": "jetpack-6.2.1-release-notes",
                "revision": "JetPack 6.2.1",
                "section": "New Features",
                "url": "https://docs.nvidia.com/jetson/jetpack/release-notes/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVIDIA® Jetson™ Linux 36.4.4 supports all NVIDIA Jetson Orin modules and developer kits",
                "class": "A"
              }
            ]
          },
          "pytorch": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "Installing PyTorch for Jetson Platform",
                "doc_id": "nvidia-install-pytorch-jetson",
                "revision": "page updated 2026-09-08",
                "section": "1. Overview > Jetson AGX Orin",
                "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform/index.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The NVIDIA Jetson AGX Orin Developer Kit includes a high-performance, power-efficient Jetson AGX Orin module, and can emulate the other Jetson modules.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "PyTorch for Jetson Platform — release notes, compatibility table",
                "doc_id": "nvidia-pytorch-jetson-release-notes",
                "revision": "page updated 2026-09-08",
                "section": "Compatibility (PyTorch version | NVIDIA container | … | JetPack)",
                "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform-release-notes/pytorch-jetson-rel.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "2.8.0a0+5228986c39 | 25.06 | - | 6.2 |",
                "class": "A"
              }
            ]
          },
          "onnx": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Microsoft",
                "document": "ONNX Runtime — Build with different execution providers",
                "doc_id": "onnxruntime-build-eps",
                "section": "TensorRT execution provider: NVIDIA Jetson",
                "url": "https://onnxruntime.ai/docs/build/eps.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVIDIA Jetson TX1/TX2/Nano/Xavier/Orin",
                "class": "A"
              }
            ]
          },
          "llama_cpp": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "Gemma 4 on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-gemma4",
                "section": "Overview",
                "url": "https://www.jetson-ai-lab.com/tutorials/gemma4-on-jetson",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The full family is supported on Jetson through both vLLM and llama.cpp. All of the models are supported on Orin and Thor,",
                "class": "A"
              },
              {
                "publisher": "ggml-org",
                "document": "llama.cpp documentation — MobileVLM",
                "doc_id": "llama-cpp-docs-mobilevlm",
                "revision": "master, 2026-09-29",
                "section": "Orin compile and run",
                "url": "https://raw.githubusercontent.com/ggml-org/llama.cpp/master/docs/multimodal/MobileVLM.md",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "- [x] run MobileVLM on `Jetson Orin`",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "jetson_thor_t5000": {
      "platform_id": "jetson_thor_t5000",
      "label": "Jetson Thor T5000",
      "summary": "Jetson Thor T5000: precisions FP32, FP16, INT8, BF16, FP8, INT4 and FP4; runtimes TensorRT, llama.cpp and PyTorch — every listed value documented by its vendor (class A: NVIDIA). Not listed, and no vendor document settles it (class E): runtime ONNX Runtime.",
      "precisions": {
        "supported": [
          "fp32",
          "fp16",
          "int8",
          "bf16",
          "fp8",
          "int4",
          "fp4"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 11.0",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '11.0', devices: ['NVIDIA Jetson AGX Thor'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Modules Data Sheet",
                "doc_id": "DS-11945-001",
                "revision": "v1.5, June 2026",
                "section": "2.1.4 Compute Features",
                "url": "https://developer.nvidia.com/downloads/assets/embedded/secure/jetson/thor/docs/jetson-thor-series-modules-datasheet_ds-11945-001.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Blackwell introduces fifth-generation NVIDIA Tensor Cores which offer a wider range of precisions including TensorFloat-32 (TF32), bfloat16, FP16, FP8, FP4, and INT8,",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 11.0",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '11.0', devices: ['NVIDIA Jetson AGX Thor'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Modules Data Sheet",
                "doc_id": "DS-11945-001",
                "revision": "v1.5, June 2026",
                "section": "2.1.4 Compute Features",
                "url": "https://developer.nvidia.com/downloads/assets/embedded/secure/jetson/thor/docs/jetson-thor-series-modules-datasheet_ds-11945-001.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Blackwell introduces fifth-generation NVIDIA Tensor Cores which offer a wider range of precisions including TensorFloat-32 (TF32), bfloat16, FP16, FP8, FP4, and INT8,",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 11.0",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '11.0', devices: ['NVIDIA Jetson AGX Thor'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Modules Data Sheet",
                "doc_id": "DS-11945-001",
                "revision": "v1.5, June 2026",
                "section": "2.1.4 Compute Features",
                "url": "https://developer.nvidia.com/downloads/assets/embedded/secure/jetson/thor/docs/jetson-thor-series-modules-datasheet_ds-11945-001.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Blackwell introduces fifth-generation NVIDIA Tensor Cores which offer a wider range of precisions including TensorFloat-32 (TF32), bfloat16, FP16, FP8, FP4, and INT8,",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 11.0",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '11.0', devices: ['NVIDIA Jetson AGX Thor'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Modules Data Sheet",
                "doc_id": "DS-11945-001",
                "revision": "v1.5, June 2026",
                "section": "2.1.4 Compute Features",
                "url": "https://developer.nvidia.com/downloads/assets/embedded/secure/jetson/thor/docs/jetson-thor-series-modules-datasheet_ds-11945-001.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Blackwell introduces fifth-generation NVIDIA Tensor Cores which offer a wider range of precisions including TensorFloat-32 (TF32), bfloat16, FP16, FP8, FP4, and INT8,",
                "class": "A"
              }
            ]
          },
          "fp8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 11.0",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '11.0', devices: ['NVIDIA Jetson AGX Thor'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Modules Data Sheet",
                "doc_id": "DS-11945-001",
                "revision": "v1.5, June 2026",
                "section": "2.1.4 Compute Features",
                "url": "https://developer.nvidia.com/downloads/assets/embedded/secure/jetson/thor/docs/jetson-thor-series-modules-datasheet_ds-11945-001.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Blackwell introduces fifth-generation NVIDIA Tensor Cores which offer a wider range of precisions including TensorFloat-32 (TF32), bfloat16, FP16, FP8, FP4, and INT8,",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "FP8 | 2x reduction | Not available | Supported |",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "INT4 AWQ | 4x reduction | Supported | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              }
            ]
          },
          "fp4": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 11.0",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '11.0', devices: ['NVIDIA Jetson AGX Thor'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Modules Data Sheet",
                "doc_id": "DS-11945-001",
                "revision": "v1.5, June 2026",
                "section": "2.1.4 Compute Features",
                "url": "https://developer.nvidia.com/downloads/assets/embedded/secure/jetson/thor/docs/jetson-thor-series-modules-datasheet_ds-11945-001.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Blackwell introduces fifth-generation NVIDIA Tensor Cores which offer a wider range of precisions including TensorFloat-32 (TF32), bfloat16, FP16, FP8, FP4, and INT8,",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVFP4 | 4x reduction | Not available | Supported |",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "tensorrt",
          "llama_cpp",
          "pytorch"
        ],
        "list_class": "A",
        "decided": {
          "tensorrt": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "JetPack SDK Downloads and Notes",
                "doc_id": "jetpack-downloads",
                "revision": "JetPack 7.2.1 / Jetson Linux 39.2.1 (8/11/2026)",
                "section": "Release information, JetPack 7.2.1",
                "url": "https://developer.nvidia.com/embedded/jetpack/downloads",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVIDIA JetPack 7.2.1 with Jetson Linux 39.2.1, CUDA 13.2.2, TensorRT 10.16.2,",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA JetPack SDK",
                "doc_id": "jetpack-sdk",
                "section": "JetPack 7",
                "url": "https://developer.nvidia.com/embedded/jetpack",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "JetPack 7 gives you full support for the NVIDIA Jetson Orin™ and Thor™ platform,",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 11.0",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '11.0', devices: ['NVIDIA Jetson AGX Thor'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              }
            ]
          },
          "llama_cpp": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "Gemma 4 on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-gemma4",
                "section": "Overview",
                "url": "https://www.jetson-ai-lab.com/tutorials/gemma4-on-jetson",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The full family is supported on Jetson through both vLLM and llama.cpp. All of the models are supported on Orin and Thor,",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Technical Brief",
                "doc_id": "TB-12572-001",
                "revision": "v1.0, August 29, 2025",
                "section": "JetPack 7 (p.28)",
                "url": "https://dam-cdn.nvd.orangelogic.com/AssetLink/v686u306w26ftuk863xjeimilimtx55e.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "It supports leading generative AI serving frameworks like vLLM, SGLang, TensorRT, TensorRT- LLM, PyTorch, Llama.cpp, and MLC for efficient real-time model deployment.",
                "class": "A"
              }
            ]
          },
          "pytorch": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "PyTorch for Jetson Platform — release notes, compatibility table",
                "doc_id": "nvidia-pytorch-jetson-release-notes",
                "revision": "page updated 2026-09-08",
                "section": "Compatibility (PyTorch version | NVIDIA container | … | JetPack)",
                "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform-release-notes/pytorch-jetson-rel.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "2.14.0a0+4fdf77b940 | 26.08 | - | 7.1 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Technical Brief",
                "doc_id": "TB-12572-001",
                "revision": "v1.0, August 29, 2025",
                "section": "JetPack 7 (p.28)",
                "url": "https://dam-cdn.nvd.orangelogic.com/AssetLink/v686u306w26ftuk863xjeimilimtx55e.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "It supports leading generative AI serving frameworks like vLLM, SGLang, TensorRT, TensorRT- LLM, PyTorch, Llama.cpp, and MLC for efficient real-time model deployment.",
                "class": "A"
              }
            ]
          },
          "onnx": {
            "supported": false,
            "class": "E",
            "note": "ONNX Runtime documents Jetson builds up to Jetson Orin; no ONNX Runtime or NVIDIA document found names Jetson Thor, so ONNX Runtime is not listed for Thor and the exclusion is a catalog assertion."
          }
        }
      }
    },
    "jetson_thor_t4000": {
      "platform_id": "jetson_thor_t4000",
      "label": "Jetson T4000",
      "summary": "Jetson T4000: precisions FP32, FP16, INT8, BF16, FP8, INT4 and FP4; runtimes TensorRT, llama.cpp and PyTorch — every listed value documented by its vendor (class A: NVIDIA). Not listed, and no vendor document settles it (class E): runtime ONNX Runtime.",
      "precisions": {
        "supported": [
          "fp32",
          "fp16",
          "int8",
          "bf16",
          "fp8",
          "int4",
          "fp4"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 11.0",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '11.0', devices: ['NVIDIA Jetson AGX Thor'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Modules Data Sheet",
                "doc_id": "DS-11945-001",
                "revision": "v1.5, June 2026",
                "section": "2.1.4 Compute Features",
                "url": "https://developer.nvidia.com/downloads/assets/embedded/secure/jetson/thor/docs/jetson-thor-series-modules-datasheet_ds-11945-001.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Blackwell introduces fifth-generation NVIDIA Tensor Cores which offer a wider range of precisions including TensorFloat-32 (TF32), bfloat16, FP16, FP8, FP4, and INT8,",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 11.0",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '11.0', devices: ['NVIDIA Jetson AGX Thor'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Modules Data Sheet",
                "doc_id": "DS-11945-001",
                "revision": "v1.5, June 2026",
                "section": "2.1.4 Compute Features",
                "url": "https://developer.nvidia.com/downloads/assets/embedded/secure/jetson/thor/docs/jetson-thor-series-modules-datasheet_ds-11945-001.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Blackwell introduces fifth-generation NVIDIA Tensor Cores which offer a wider range of precisions including TensorFloat-32 (TF32), bfloat16, FP16, FP8, FP4, and INT8,",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 11.0",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '11.0', devices: ['NVIDIA Jetson AGX Thor'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Modules Data Sheet",
                "doc_id": "DS-11945-001",
                "revision": "v1.5, June 2026",
                "section": "2.1.4 Compute Features",
                "url": "https://developer.nvidia.com/downloads/assets/embedded/secure/jetson/thor/docs/jetson-thor-series-modules-datasheet_ds-11945-001.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Blackwell introduces fifth-generation NVIDIA Tensor Cores which offer a wider range of precisions including TensorFloat-32 (TF32), bfloat16, FP16, FP8, FP4, and INT8,",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 11.0",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '11.0', devices: ['NVIDIA Jetson AGX Thor'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Modules Data Sheet",
                "doc_id": "DS-11945-001",
                "revision": "v1.5, June 2026",
                "section": "2.1.4 Compute Features",
                "url": "https://developer.nvidia.com/downloads/assets/embedded/secure/jetson/thor/docs/jetson-thor-series-modules-datasheet_ds-11945-001.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Blackwell introduces fifth-generation NVIDIA Tensor Cores which offer a wider range of precisions including TensorFloat-32 (TF32), bfloat16, FP16, FP8, FP4, and INT8,",
                "class": "A"
              }
            ]
          },
          "fp8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 11.0",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '11.0', devices: ['NVIDIA Jetson AGX Thor'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Modules Data Sheet",
                "doc_id": "DS-11945-001",
                "revision": "v1.5, June 2026",
                "section": "2.1.4 Compute Features",
                "url": "https://developer.nvidia.com/downloads/assets/embedded/secure/jetson/thor/docs/jetson-thor-series-modules-datasheet_ds-11945-001.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Blackwell introduces fifth-generation NVIDIA Tensor Cores which offer a wider range of precisions including TensorFloat-32 (TF32), bfloat16, FP16, FP8, FP4, and INT8,",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "FP8 | 2x reduction | Not available | Supported |",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "INT4 AWQ | 4x reduction | Supported | Supported |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              }
            ]
          },
          "fp4": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 11.0",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '11.0', devices: ['NVIDIA Jetson AGX Thor'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Modules Data Sheet",
                "doc_id": "DS-11945-001",
                "revision": "v1.5, June 2026",
                "section": "2.1.4 Compute Features",
                "url": "https://developer.nvidia.com/downloads/assets/embedded/secure/jetson/thor/docs/jetson-thor-series-modules-datasheet_ds-11945-001.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Blackwell introduces fifth-generation NVIDIA Tensor Cores which offer a wider range of precisions including TensorFloat-32 (TF32), bfloat16, FP16, FP8, FP4, and INT8,",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
                "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
                "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVFP4 | 4x reduction | Not available | Supported |",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "tensorrt",
          "llama_cpp",
          "pytorch"
        ],
        "list_class": "A",
        "decided": {
          "tensorrt": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "JetPack SDK Downloads and Notes",
                "doc_id": "jetpack-downloads",
                "revision": "JetPack 7.2.1 / Jetson Linux 39.2.1 (8/11/2026)",
                "section": "Release information, JetPack 7.2.1",
                "url": "https://developer.nvidia.com/embedded/jetpack/downloads",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "NVIDIA JetPack 7.2.1 with Jetson Linux 39.2.1, CUDA 13.2.2, TensorRT 10.16.2,",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA JetPack SDK",
                "doc_id": "jetpack-sdk",
                "section": "JetPack 7",
                "url": "https://developer.nvidia.com/embedded/jetpack",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "JetPack 7 gives you full support for the NVIDIA Jetson Orin™ and Thor™ platform,",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 11.0",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '11.0', devices: ['NVIDIA Jetson AGX Thor'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 11.0 (Jetson)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
                "class": "A"
              }
            ]
          },
          "llama_cpp": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "Gemma 4 on Jetson (Jetson AI Lab)",
                "doc_id": "jetson-ai-lab-gemma4",
                "section": "Overview",
                "url": "https://www.jetson-ai-lab.com/tutorials/gemma4-on-jetson",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The full family is supported on Jetson through both vLLM and llama.cpp. All of the models are supported on Orin and Thor,",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Technical Brief",
                "doc_id": "TB-12572-001",
                "revision": "v1.0, August 29, 2025",
                "section": "JetPack 7 (p.28)",
                "url": "https://dam-cdn.nvd.orangelogic.com/AssetLink/v686u306w26ftuk863xjeimilimtx55e.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "It supports leading generative AI serving frameworks like vLLM, SGLang, TensorRT, TensorRT- LLM, PyTorch, Llama.cpp, and MLC for efficient real-time model deployment.",
                "class": "A"
              }
            ]
          },
          "pytorch": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "PyTorch for Jetson Platform — release notes, compatibility table",
                "doc_id": "nvidia-pytorch-jetson-release-notes",
                "revision": "page updated 2026-09-08",
                "section": "Compatibility (PyTorch version | NVIDIA container | … | JetPack)",
                "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform-release-notes/pytorch-jetson-rel.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "2.14.0a0+4fdf77b940 | 26.08 | - | 7.1 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA Jetson Thor Series Technical Brief",
                "doc_id": "TB-12572-001",
                "revision": "v1.0, August 29, 2025",
                "section": "JetPack 7 (p.28)",
                "url": "https://dam-cdn.nvd.orangelogic.com/AssetLink/v686u306w26ftuk863xjeimilimtx55e.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "It supports leading generative AI serving frameworks like vLLM, SGLang, TensorRT, TensorRT- LLM, PyTorch, Llama.cpp, and MLC for efficient real-time model deployment.",
                "class": "A"
              }
            ]
          },
          "onnx": {
            "supported": false,
            "class": "E",
            "note": "ONNX Runtime documents Jetson builds up to Jetson Orin; no ONNX Runtime or NVIDIA document found names Jetson Thor, so ONNX Runtime is not listed for Thor and the exclusion is a catalog assertion."
          }
        }
      }
    },
    "coral_tpu": {
      "platform_id": "coral_tpu",
      "label": "Google Coral Edge TPU",
      "summary": "Google Coral Edge TPU: precision INT8; runtime the Edge TPU runtime — every listed value documented by its vendor (class A: Google). Documented as not supported (class A): precisions FP32, FP16, BF16, FP8, INT4 and FP4; runtimes TensorRT, PyTorch, ONNX Runtime, llama.cpp, HailoRT, RKNN and OpenVINO.",
      "precisions": {
        "supported": [
          "int8"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Google",
                "document": "TensorFlow models on the Edge TPU",
                "doc_id": "coral-edgetpu-models-intro",
                "section": "Compatibility overview",
                "url": "https://coral.ai/docs/edgetpu/models-intro/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "It supports only TensorFlow Lite models that are fully 8-bit quantized and then compiled specifically for the Edge TPU.",
                "class": "A"
              },
              {
                "publisher": "Google",
                "document": "TensorFlow models on the Edge TPU",
                "doc_id": "coral-edgetpu-models-intro",
                "section": "Quantization",
                "url": "https://coral.ai/docs/edgetpu/models-intro/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "By default, the TensorFlow Lite converter leaves those operations in their float format, which is not compatible with the Edge TPU.",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Google",
                "document": "TensorFlow models on the Edge TPU",
                "doc_id": "coral-edgetpu-models-intro",
                "section": "Compatibility overview",
                "url": "https://coral.ai/docs/edgetpu/models-intro/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "It supports only TensorFlow Lite models that are fully 8-bit quantized and then compiled specifically for the Edge TPU.",
                "class": "A"
              },
              {
                "publisher": "Google",
                "document": "TensorFlow models on the Edge TPU",
                "doc_id": "coral-edgetpu-models-intro",
                "section": "Quantization",
                "url": "https://coral.ai/docs/edgetpu/models-intro/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "By default, the TensorFlow Lite converter leaves those operations in their float format, which is not compatible with the Edge TPU.",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Google",
                "document": "TensorFlow models on the Edge TPU",
                "doc_id": "coral-edgetpu-models-intro",
                "section": "Compatibility overview",
                "url": "https://coral.ai/docs/edgetpu/models-intro/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "It supports only TensorFlow Lite models that are fully 8-bit quantized and then compiled specifically for the Edge TPU.",
                "class": "A"
              },
              {
                "publisher": "Google",
                "document": "TensorFlow models on the Edge TPU",
                "doc_id": "coral-edgetpu-models-intro",
                "section": "Model requirements",
                "url": "https://coral.ai/docs/edgetpu/models-intro/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Tensor parameters are quantized (8-bit fixed-point numbers; int8 or uint8).",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Google",
                "document": "TensorFlow models on the Edge TPU",
                "doc_id": "coral-edgetpu-models-intro",
                "section": "Compatibility overview",
                "url": "https://coral.ai/docs/edgetpu/models-intro/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "It supports only TensorFlow Lite models that are fully 8-bit quantized and then compiled specifically for the Edge TPU.",
                "class": "A"
              },
              {
                "publisher": "Google",
                "document": "TensorFlow models on the Edge TPU",
                "doc_id": "coral-edgetpu-models-intro",
                "section": "Quantization",
                "url": "https://coral.ai/docs/edgetpu/models-intro/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "By default, the TensorFlow Lite converter leaves those operations in their float format, which is not compatible with the Edge TPU.",
                "class": "A"
              }
            ]
          },
          "fp8": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Google",
                "document": "TensorFlow models on the Edge TPU",
                "doc_id": "coral-edgetpu-models-intro",
                "section": "Compatibility overview",
                "url": "https://coral.ai/docs/edgetpu/models-intro/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "It supports only TensorFlow Lite models that are fully 8-bit quantized and then compiled specifically for the Edge TPU.",
                "class": "A"
              },
              {
                "publisher": "Google",
                "document": "TensorFlow models on the Edge TPU",
                "doc_id": "coral-edgetpu-models-intro",
                "section": "Quantization",
                "url": "https://coral.ai/docs/edgetpu/models-intro/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "By default, the TensorFlow Lite converter leaves those operations in their float format, which is not compatible with the Edge TPU.",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Google",
                "document": "TensorFlow models on the Edge TPU",
                "doc_id": "coral-edgetpu-models-intro",
                "section": "Compatibility overview",
                "url": "https://coral.ai/docs/edgetpu/models-intro/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "It supports only TensorFlow Lite models that are fully 8-bit quantized and then compiled specifically for the Edge TPU.",
                "class": "A"
              }
            ]
          },
          "fp4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Google",
                "document": "TensorFlow models on the Edge TPU",
                "doc_id": "coral-edgetpu-models-intro",
                "section": "Compatibility overview",
                "url": "https://coral.ai/docs/edgetpu/models-intro/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "It supports only TensorFlow Lite models that are fully 8-bit quantized and then compiled specifically for the Edge TPU.",
                "class": "A"
              },
              {
                "publisher": "Google",
                "document": "TensorFlow models on the Edge TPU",
                "doc_id": "coral-edgetpu-models-intro",
                "section": "Quantization",
                "url": "https://coral.ai/docs/edgetpu/models-intro/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "By default, the TensorFlow Lite converter leaves those operations in their float format, which is not compatible with the Edge TPU.",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "edge_tpu"
        ],
        "list_class": "A",
        "decided": {
          "edge_tpu": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Google",
                "document": "Edge TPU inferencing overview",
                "doc_id": "coral-edgetpu-inference",
                "section": "Intro + note",
                "url": "https://coral.ai/docs/edgetpu/inference/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "All inferencing with the Edge TPU is executed with TensorFlow Lite libraries.",
                "class": "A"
              },
              {
                "publisher": "Google",
                "document": "Edge TPU inferencing overview",
                "doc_id": "coral-edgetpu-inference",
                "section": "General-purpose operating systems",
                "url": "https://coral.ai/docs/edgetpu/inference/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Regardless of the language you choose, you need to install the Edge TPU runtime (libedgetpu.so), as documented in the setup for each Coral device.",
                "class": "A"
              }
            ]
          },
          "tensorrt": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Google",
                "document": "Edge TPU frequently asked questions",
                "doc_id": "coral-edgetpu-faq",
                "section": "What machine learning frameworks does the Edge TPU support?",
                "url": "https://coral.ai/docs/edgetpu/faq/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorFlow Lite only.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorRT supports NVIDIA hardware with compute capability SM 7.5 or higher.",
                "class": "A"
              }
            ]
          },
          "pytorch": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Google",
                "document": "Edge TPU frequently asked questions",
                "doc_id": "coral-edgetpu-faq",
                "section": "What machine learning frameworks does the Edge TPU support?",
                "url": "https://coral.ai/docs/edgetpu/faq/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorFlow Lite only.",
                "class": "A"
              }
            ]
          },
          "onnx": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Google",
                "document": "Edge TPU frequently asked questions",
                "doc_id": "coral-edgetpu-faq",
                "section": "What machine learning frameworks does the Edge TPU support?",
                "url": "https://coral.ai/docs/edgetpu/faq/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorFlow Lite only.",
                "class": "A"
              }
            ]
          },
          "llama_cpp": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Google",
                "document": "Edge TPU frequently asked questions",
                "doc_id": "coral-edgetpu-faq",
                "section": "What machine learning frameworks does the Edge TPU support?",
                "url": "https://coral.ai/docs/edgetpu/faq/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorFlow Lite only.",
                "class": "A"
              }
            ]
          },
          "hailo_sdk": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Google",
                "document": "Edge TPU frequently asked questions",
                "doc_id": "coral-edgetpu-faq",
                "section": "What machine learning frameworks does the Edge TPU support?",
                "url": "https://coral.ai/docs/edgetpu/faq/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorFlow Lite only.",
                "class": "A"
              }
            ]
          },
          "rknn": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Google",
                "document": "Edge TPU frequently asked questions",
                "doc_id": "coral-edgetpu-faq",
                "section": "What machine learning frameworks does the Edge TPU support?",
                "url": "https://coral.ai/docs/edgetpu/faq/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorFlow Lite only.",
                "class": "A"
              }
            ]
          },
          "openvino": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Google",
                "document": "Edge TPU frequently asked questions",
                "doc_id": "coral-edgetpu-faq",
                "section": "What machine learning frameworks does the Edge TPU support?",
                "url": "https://coral.ai/docs/edgetpu/faq/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorFlow Lite only.",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "hailo_8": {
      "platform_id": "hailo_8",
      "label": "Hailo-8",
      "summary": "Hailo-8: precisions INT8 and INT4; runtime HailoRT — every listed value documented by its vendor (class A: Hailo). Documented as not supported (class A): precisions FP32, FP16, BF16, FP8 and FP4; runtime TensorRT.",
      "precisions": {
        "supported": [
          "int8",
          "int4"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-8 AI Processor Datasheet",
                "doc_id": "hailo-8-datasheet",
                "revision": "Rev 1.9.2, April 2026",
                "section": "5.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-8-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8/16-bit precision for both weights and activations, configurable at layer granularity.",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Model Optimization",
                "doc_id": "hailo-model-zoo-optimization",
                "revision": "master a20777477e (2025-09-28); same in v2.19.1",
                "section": "Optimization Workflow",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Getting Started",
                "doc_id": "hailo-model-zoo-getting-started",
                "revision": "master 757435a9b6 (2026-03-30)",
                "section": "Optimization",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Full-precision / emulation runs on GPU.",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-8 AI Processor Datasheet",
                "doc_id": "hailo-8-datasheet",
                "revision": "Rev 1.9.2, April 2026",
                "section": "5.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-8-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8/16-bit precision for both weights and activations, configurable at layer granularity.",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Model Optimization",
                "doc_id": "hailo-model-zoo-optimization",
                "revision": "master a20777477e (2025-09-28); same in v2.19.1",
                "section": "Optimization Workflow",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Getting Started",
                "doc_id": "hailo-model-zoo-getting-started",
                "revision": "master 757435a9b6 (2026-03-30)",
                "section": "Optimization",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Full-precision / emulation runs on GPU.",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-8 AI Processor Datasheet",
                "doc_id": "hailo-8-datasheet",
                "revision": "Rev 1.9.2, April 2026",
                "section": "5.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-8-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8/16-bit precision for both weights and activations, configurable at layer granularity.",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo-8 M.2 Key M ET Module Data Sheet",
                "doc_id": "hailo-8-m2-key-m-et-datasheet",
                "revision": "Rev 2.5, November 2025",
                "section": "Table 1 Key properties",
                "url": "https://hailo.ai/hailo-files/hailo-8-m-2-key-m-et-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Peak Performance (INT8) 26 TOPS",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-8 AI Processor Datasheet",
                "doc_id": "hailo-8-datasheet",
                "revision": "Rev 1.9.2, April 2026",
                "section": "5.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-8-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8/16-bit precision for both weights and activations, configurable at layer granularity.",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Model Optimization",
                "doc_id": "hailo-model-zoo-optimization",
                "revision": "master a20777477e (2025-09-28); same in v2.19.1",
                "section": "Optimization Workflow",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Getting Started",
                "doc_id": "hailo-model-zoo-getting-started",
                "revision": "master 757435a9b6 (2026-03-30)",
                "section": "Optimization",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Full-precision / emulation runs on GPU.",
                "class": "A"
              }
            ]
          },
          "fp8": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-8 AI Processor Datasheet",
                "doc_id": "hailo-8-datasheet",
                "revision": "Rev 1.9.2, April 2026",
                "section": "5.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-8-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8/16-bit precision for both weights and activations, configurable at layer granularity.",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Model Optimization",
                "doc_id": "hailo-model-zoo-optimization",
                "revision": "master a20777477e (2025-09-28); same in v2.19.1",
                "section": "Optimization Workflow",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Getting Started",
                "doc_id": "hailo-model-zoo-getting-started",
                "revision": "master 757435a9b6 (2026-03-30)",
                "section": "Optimization",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Full-precision / emulation runs on GPU.",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-8 AI Processor Datasheet",
                "doc_id": "hailo-8-datasheet",
                "revision": "Rev 1.9.2, April 2026",
                "section": "5.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-8-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "4-bit precision for weights, configurable at layer granularity.",
                "class": "A"
              }
            ]
          },
          "fp4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-8 AI Processor Datasheet",
                "doc_id": "hailo-8-datasheet",
                "revision": "Rev 1.9.2, April 2026",
                "section": "5.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-8-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8/16-bit precision for both weights and activations, configurable at layer granularity.",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Model Optimization",
                "doc_id": "hailo-model-zoo-optimization",
                "revision": "master a20777477e (2025-09-28); same in v2.19.1",
                "section": "Optimization Workflow",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Getting Started",
                "doc_id": "hailo-model-zoo-getting-started",
                "revision": "master 757435a9b6 (2026-03-30)",
                "section": "Optimization",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Full-precision / emulation runs on GPU.",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "hailo_sdk"
        ],
        "list_class": "A",
        "decided": {
          "hailo_sdk": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "HailoRT README",
                "doc_id": "hailort-readme",
                "revision": "master f51959034a (2026-08-16), release v5.4.0",
                "section": "Supported devices",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailort/master/README.md",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Note: The `master` branch supports only the Hailo-10 and Hailo-15 device families. For Hailo-8, Hailo-8R, and Hailo-8L devices, please use the `hailo8` branch.",
                "class": "A"
              }
            ]
          },
          "tensorrt": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorRT supports NVIDIA hardware with compute capability SM 7.5 or higher.",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "hailo_8l": {
      "platform_id": "hailo_8l",
      "label": "Hailo-8L",
      "summary": "Hailo-8L: precisions INT8 and INT4; runtime HailoRT — every listed value documented by its vendor (class A: Hailo). Documented as not supported (class A): precisions FP32, FP16, BF16, FP8 and FP4; runtime TensorRT.",
      "precisions": {
        "supported": [
          "int8",
          "int4"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-8L AI Processor Datasheet (industrial)",
                "doc_id": "hailo-8l-industrial-datasheet",
                "revision": "Rev 1.9.3, April 2026",
                "section": "5.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-8l-industrial-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8/16-bit precision for both weights and activations, configurable at layer granularity.",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Model Optimization",
                "doc_id": "hailo-model-zoo-optimization",
                "revision": "master a20777477e (2025-09-28); same in v2.19.1",
                "section": "Optimization Workflow",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Getting Started",
                "doc_id": "hailo-model-zoo-getting-started",
                "revision": "master 757435a9b6 (2026-03-30)",
                "section": "Optimization",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Full-precision / emulation runs on GPU.",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-8L AI Processor Datasheet (industrial)",
                "doc_id": "hailo-8l-industrial-datasheet",
                "revision": "Rev 1.9.3, April 2026",
                "section": "5.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-8l-industrial-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8/16-bit precision for both weights and activations, configurable at layer granularity.",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Model Optimization",
                "doc_id": "hailo-model-zoo-optimization",
                "revision": "master a20777477e (2025-09-28); same in v2.19.1",
                "section": "Optimization Workflow",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Getting Started",
                "doc_id": "hailo-model-zoo-getting-started",
                "revision": "master 757435a9b6 (2026-03-30)",
                "section": "Optimization",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Full-precision / emulation runs on GPU.",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-8L AI Processor Datasheet (industrial)",
                "doc_id": "hailo-8l-industrial-datasheet",
                "revision": "Rev 1.9.3, April 2026",
                "section": "5.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-8l-industrial-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8/16-bit precision for both weights and activations, configurable at layer granularity.",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo-8L M.2 Key B+M ET Module Data Sheet",
                "doc_id": "hailo-8l-m2-key-bm-et-datasheet",
                "revision": "Rev 4.0, November 2025",
                "section": "Table 1 Key properties",
                "url": "https://hailo.ai/hailo-files/hailo-8l-m-2-key-b-m-et-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Peak Performance (INT8) 13 TOPS",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-8L AI Processor Datasheet (industrial)",
                "doc_id": "hailo-8l-industrial-datasheet",
                "revision": "Rev 1.9.3, April 2026",
                "section": "5.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-8l-industrial-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8/16-bit precision for both weights and activations, configurable at layer granularity.",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Model Optimization",
                "doc_id": "hailo-model-zoo-optimization",
                "revision": "master a20777477e (2025-09-28); same in v2.19.1",
                "section": "Optimization Workflow",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Getting Started",
                "doc_id": "hailo-model-zoo-getting-started",
                "revision": "master 757435a9b6 (2026-03-30)",
                "section": "Optimization",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Full-precision / emulation runs on GPU.",
                "class": "A"
              }
            ]
          },
          "fp8": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-8L AI Processor Datasheet (industrial)",
                "doc_id": "hailo-8l-industrial-datasheet",
                "revision": "Rev 1.9.3, April 2026",
                "section": "5.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-8l-industrial-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8/16-bit precision for both weights and activations, configurable at layer granularity.",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Model Optimization",
                "doc_id": "hailo-model-zoo-optimization",
                "revision": "master a20777477e (2025-09-28); same in v2.19.1",
                "section": "Optimization Workflow",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Getting Started",
                "doc_id": "hailo-model-zoo-getting-started",
                "revision": "master 757435a9b6 (2026-03-30)",
                "section": "Optimization",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Full-precision / emulation runs on GPU.",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-8L AI Processor Datasheet (industrial)",
                "doc_id": "hailo-8l-industrial-datasheet",
                "revision": "Rev 1.9.3, April 2026",
                "section": "5.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-8l-industrial-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "4-bit precision for weights, configurable at layer granularity.",
                "class": "A"
              }
            ]
          },
          "fp4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-8L AI Processor Datasheet (industrial)",
                "doc_id": "hailo-8l-industrial-datasheet",
                "revision": "Rev 1.9.3, April 2026",
                "section": "5.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-8l-industrial-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8/16-bit precision for both weights and activations, configurable at layer granularity.",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Model Optimization",
                "doc_id": "hailo-model-zoo-optimization",
                "revision": "master a20777477e (2025-09-28); same in v2.19.1",
                "section": "Optimization Workflow",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Getting Started",
                "doc_id": "hailo-model-zoo-getting-started",
                "revision": "master 757435a9b6 (2026-03-30)",
                "section": "Optimization",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Full-precision / emulation runs on GPU.",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "hailo_sdk"
        ],
        "list_class": "A",
        "decided": {
          "hailo_sdk": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "HailoRT README",
                "doc_id": "hailort-readme",
                "revision": "master f51959034a (2026-08-16), release v5.4.0",
                "section": "Supported devices",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailort/master/README.md",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Note: The `master` branch supports only the Hailo-10 and Hailo-15 device families. For Hailo-8, Hailo-8R, and Hailo-8L devices, please use the `hailo8` branch.",
                "class": "A"
              }
            ]
          },
          "tensorrt": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorRT supports NVIDIA hardware with compute capability SM 7.5 or higher.",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "hailo_10h": {
      "platform_id": "hailo_10h",
      "label": "Hailo-10H",
      "summary": "Hailo-10H: precisions INT8 and INT4; runtime HailoRT — every listed value documented by its vendor (class A: Hailo). Documented as not supported (class A): precisions FP32, FP16, BF16, FP8 and FP4; runtime TensorRT.",
      "precisions": {
        "supported": [
          "int8",
          "int4"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-10H AI Accelerator Datasheet",
                "doc_id": "hailo-10h-datasheet",
                "revision": "Rev 1.0, July 2025",
                "section": "4.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-10h-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Precision configurability, support 4/8/16-bit precision for weights and activations",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Model Optimization",
                "doc_id": "hailo-model-zoo-optimization",
                "revision": "master a20777477e (2025-09-28); same in v2.19.1",
                "section": "Optimization Workflow",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Getting Started",
                "doc_id": "hailo-model-zoo-getting-started",
                "revision": "master 757435a9b6 (2026-03-30)",
                "section": "Optimization",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Full-precision / emulation runs on GPU.",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-10H AI Accelerator Datasheet",
                "doc_id": "hailo-10h-datasheet",
                "revision": "Rev 1.0, July 2025",
                "section": "4.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-10h-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Precision configurability, support 4/8/16-bit precision for weights and activations",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Model Optimization",
                "doc_id": "hailo-model-zoo-optimization",
                "revision": "master a20777477e (2025-09-28); same in v2.19.1",
                "section": "Optimization Workflow",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Getting Started",
                "doc_id": "hailo-model-zoo-getting-started",
                "revision": "master 757435a9b6 (2026-03-30)",
                "section": "Optimization",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Full-precision / emulation runs on GPU.",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-10H AI Accelerator Product Brief",
                "doc_id": "hailo-10h-product-brief",
                "revision": "7/25",
                "section": "Key Features & Benefits",
                "url": "https://hailo.ai/hailo-files/hailo-10h-product-brief-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "40|20 TOPS (INT4|8)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo-10H AI Accelerator Datasheet",
                "doc_id": "hailo-10h-datasheet",
                "revision": "Rev 1.0, July 2025",
                "section": "4.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-10h-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Precision configurability, support 4/8/16-bit precision for weights and activations",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-10H AI Accelerator Datasheet",
                "doc_id": "hailo-10h-datasheet",
                "revision": "Rev 1.0, July 2025",
                "section": "4.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-10h-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Precision configurability, support 4/8/16-bit precision for weights and activations",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Model Optimization",
                "doc_id": "hailo-model-zoo-optimization",
                "revision": "master a20777477e (2025-09-28); same in v2.19.1",
                "section": "Optimization Workflow",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Getting Started",
                "doc_id": "hailo-model-zoo-getting-started",
                "revision": "master 757435a9b6 (2026-03-30)",
                "section": "Optimization",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Full-precision / emulation runs on GPU.",
                "class": "A"
              }
            ]
          },
          "fp8": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-10H AI Accelerator Datasheet",
                "doc_id": "hailo-10h-datasheet",
                "revision": "Rev 1.0, July 2025",
                "section": "4.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-10h-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Precision configurability, support 4/8/16-bit precision for weights and activations",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Model Optimization",
                "doc_id": "hailo-model-zoo-optimization",
                "revision": "master a20777477e (2025-09-28); same in v2.19.1",
                "section": "Optimization Workflow",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Getting Started",
                "doc_id": "hailo-model-zoo-getting-started",
                "revision": "master 757435a9b6 (2026-03-30)",
                "section": "Optimization",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Full-precision / emulation runs on GPU.",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-10H AI Accelerator Datasheet",
                "doc_id": "hailo-10h-datasheet",
                "revision": "Rev 1.0, July 2025",
                "section": "4.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-10h-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Precision configurability, support 4/8/16-bit precision for weights and activations",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo-10H AI Accelerator Product Brief",
                "doc_id": "hailo-10h-product-brief",
                "revision": "7/25",
                "section": "Key Features & Benefits",
                "url": "https://hailo.ai/hailo-files/hailo-10h-product-brief-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "40|20 TOPS (INT4|8)",
                "class": "A"
              }
            ]
          },
          "fp4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "Hailo-10H AI Accelerator Datasheet",
                "doc_id": "hailo-10h-datasheet",
                "revision": "Rev 1.0, July 2025",
                "section": "4.3 NN Core",
                "url": "https://hailo.ai/hailo-files/hailo-10h-datasheet-en/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Precision configurability, support 4/8/16-bit precision for weights and activations",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Model Optimization",
                "doc_id": "hailo-model-zoo-optimization",
                "revision": "master a20777477e (2025-09-28); same in v2.19.1",
                "section": "Optimization Workflow",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "Hailo Model Zoo — Getting Started",
                "doc_id": "hailo-model-zoo-getting-started",
                "revision": "master 757435a9b6 (2026-03-30)",
                "section": "Optimization",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Full-precision / emulation runs on GPU.",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "hailo_sdk"
        ],
        "list_class": "A",
        "decided": {
          "hailo_sdk": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Hailo",
                "document": "HailoRT README",
                "doc_id": "hailort-readme",
                "revision": "master f51959034a (2026-08-16), release v5.4.0",
                "section": "Supported devices",
                "url": "https://raw.githubusercontent.com/hailo-ai/hailort/master/README.md",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Note: The `master` branch supports only the Hailo-10 and Hailo-15 device families. For Hailo-8, Hailo-8R, and Hailo-8L devices, please use the `hailo8` branch.",
                "class": "A"
              },
              {
                "publisher": "Hailo",
                "document": "HailoRT 5.4.0 User Guide — Installation Inside a Docker Container",
                "doc_id": "hailort-5.4.0-user-guide",
                "revision": "HailoRT 5.4.0",
                "section": "Running Hailo-10H and Hailo-8 Side by Side",
                "url": "https://hailo.ai/docs/31771-hailort_5.4.0_sphinx-2/html/install/docker.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Start a container that contains the Hailo-10H HailoRT library (v5.x) and expose only the Hailo-10H device to it.",
                "class": "A"
              }
            ]
          },
          "tensorrt": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorRT supports NVIDIA hardware with compute capability SM 7.5 or higher.",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "rk3588_npu": {
      "platform_id": "rk3588_npu",
      "label": "Rockchip RK3588 NPU",
      "summary": "Rockchip RK3588 NPU: precisions FP16 and INT8; runtime RKNN — every listed value documented by its vendor (class A: Rockchip). Documented as not supported (class A): precisions FP32, BF16, FP8, INT4 and FP4; runtime TensorRT.",
      "precisions": {
        "supported": [
          "fp16",
          "int8"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "7.1.1 Simulator FP16 Accuracy",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "RKNPU currently does not support FP32 calculations, so the simulator defaults to FP16 computation when quantization is not enabled.",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, float_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "float_dtype: Used to specify the data type of floating in the non-quantized case, the data types currently supported are float16.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "7.1.1 Simulator FP16 Accuracy",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "RKNPU currently does not support FP32 calculations, so the simulator defaults to FP16 computation when quantization is not enabled.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, target_platform",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "'rk3562', 'rk3566', 'rk3568', 'rk3576' and 'rk3588' are currently supported.",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, quantized_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "- w8a8: The weight is 8bit asymmetric quantitative accuracy, and the activation value is 8bit asymmetric quantitative accuracy.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, target_platform",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "'rk3562', 'rk3566', 'rk3568', 'rk3576' and 'rk3588' are currently supported.",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, float_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "float_dtype: Used to specify the data type of floating in the non-quantized case, the data types currently supported are float16.",
                "class": "A"
              }
            ]
          },
          "fp8": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, float_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "float_dtype: Used to specify the data type of floating in the non-quantized case, the data types currently supported are float16.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, quantized_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "quantized_dtype: Quantization type, the quantization types currently supported are \"w8a8\", \"w4a16\", \"w8a16\", \"w4a8\", \"w16a16i\" and \"w16a16i_dfp\".",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, quantized_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "- w4a16: The weight is 4bit asymmetric quantitative accuracy, the activation value is 16bit floating point accuracy. (Only RK3576/RV1126B supported)",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKLLM SDK User Guide",
                "doc_id": "rkllm-sdk-user-guide",
                "revision": "V1.3.1, 2026-09-28",
                "section": "3.1.3 build, quantized_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-llm/main/doc/Rockchip_RKLLM_SDK_EN_1.3.1.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The RK3588 platform supports four quantization types: \"w8a8\", \"w8a8_g128\", \"w8a8_g256\", and \"w8a8_g512\".",
                "class": "A"
              }
            ]
          },
          "fp4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, float_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "float_dtype: Used to specify the data type of floating in the non-quantized case, the data types currently supported are float16.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, quantized_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "quantized_dtype: Quantization type, the quantization types currently supported are \"w8a8\", \"w4a16\", \"w8a16\", \"w4a8\", \"w16a16i\" and \"w16a16i_dfp\".",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "rknn"
        ],
        "list_class": "A",
        "decided": {
          "rknn": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "2.2 RKNPU2 Runtime Library",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "librknnrt.so: The Runtime library for RK3562/RK3566/RK3568/RK3576/RK3588/RV1126B development boards.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, target_platform",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "'rk3562', 'rk3566', 'rk3568', 'rk3576' and 'rk3588' are currently supported.",
                "class": "A"
              }
            ]
          },
          "tensorrt": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorRT supports NVIDIA hardware with compute capability SM 7.5 or higher.",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "rk3588s": {
      "platform_id": "rk3588s",
      "label": "Rockchip RK3588S",
      "summary": "Rockchip RK3588S: precisions FP16 and INT8; runtime RKNN — every listed value documented by its vendor (class A: Rockchip). Documented as not supported (class A): precisions FP32, BF16, FP8, INT4 and FP4; runtime TensorRT.",
      "precisions": {
        "supported": [
          "fp16",
          "int8"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "7.1.1 Simulator FP16 Accuracy",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "RKNPU currently does not support FP32 calculations, so the simulator defaults to FP16 computation when quantization is not enabled.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "1.3 Supported Hardware Platforms",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "RK3588 is used to indicate RK3588/RK3588S series in this document.",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, float_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "float_dtype: Used to specify the data type of floating in the non-quantized case, the data types currently supported are float16.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "7.1.1 Simulator FP16 Accuracy",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "RKNPU currently does not support FP32 calculations, so the simulator defaults to FP16 computation when quantization is not enabled.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, target_platform",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "'rk3562', 'rk3566', 'rk3568', 'rk3576' and 'rk3588' are currently supported.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "1.3 Supported Hardware Platforms",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "RK3588 is used to indicate RK3588/RK3588S series in this document.",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, quantized_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "- w8a8: The weight is 8bit asymmetric quantitative accuracy, and the activation value is 8bit asymmetric quantitative accuracy.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, target_platform",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "'rk3562', 'rk3566', 'rk3568', 'rk3576' and 'rk3588' are currently supported.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "1.3 Supported Hardware Platforms",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "RK3588 is used to indicate RK3588/RK3588S series in this document.",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, float_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "float_dtype: Used to specify the data type of floating in the non-quantized case, the data types currently supported are float16.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "1.3 Supported Hardware Platforms",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "RK3588 is used to indicate RK3588/RK3588S series in this document.",
                "class": "A"
              }
            ]
          },
          "fp8": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, float_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "float_dtype: Used to specify the data type of floating in the non-quantized case, the data types currently supported are float16.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, quantized_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "quantized_dtype: Quantization type, the quantization types currently supported are \"w8a8\", \"w4a16\", \"w8a16\", \"w4a8\", \"w16a16i\" and \"w16a16i_dfp\".",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "1.3 Supported Hardware Platforms",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "RK3588 is used to indicate RK3588/RK3588S series in this document.",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, quantized_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "- w4a16: The weight is 4bit asymmetric quantitative accuracy, the activation value is 16bit floating point accuracy. (Only RK3576/RV1126B supported)",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKLLM SDK User Guide",
                "doc_id": "rkllm-sdk-user-guide",
                "revision": "V1.3.1, 2026-09-28",
                "section": "3.1.3 build, quantized_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-llm/main/doc/Rockchip_RKLLM_SDK_EN_1.3.1.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "The RK3588 platform supports four quantization types: \"w8a8\", \"w8a8_g128\", \"w8a8_g256\", and \"w8a8_g512\".",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "1.3 Supported Hardware Platforms",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "RK3588 is used to indicate RK3588/RK3588S series in this document.",
                "class": "A"
              }
            ]
          },
          "fp4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, float_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "float_dtype: Used to specify the data type of floating in the non-quantized case, the data types currently supported are float16.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, quantized_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "quantized_dtype: Quantization type, the quantization types currently supported are \"w8a8\", \"w4a16\", \"w8a16\", \"w4a8\", \"w16a16i\" and \"w16a16i_dfp\".",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "1.3 Supported Hardware Platforms",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "RK3588 is used to indicate RK3588/RK3588S series in this document.",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "rknn"
        ],
        "list_class": "A",
        "decided": {
          "rknn": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "2.2 RKNPU2 Runtime Library",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "librknnrt.so: The Runtime library for RK3562/RK3566/RK3568/RK3576/RK3588/RV1126B development boards.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, target_platform",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "'rk3562', 'rk3566', 'rk3568', 'rk3576' and 'rk3588' are currently supported.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "1.3 Supported Hardware Platforms",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "RK3588 is used to indicate RK3588/RK3588S series in this document.",
                "class": "A"
              }
            ]
          },
          "tensorrt": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorRT supports NVIDIA hardware with compute capability SM 7.5 or higher.",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "rk3568": {
      "platform_id": "rk3568",
      "label": "Rockchip RK3568",
      "summary": "Rockchip RK3568: precisions FP16 and INT8; runtime RKNN — every listed value documented by its vendor (class A: Rockchip). Documented as not supported (class A): precisions FP32, BF16, FP8, INT4 and FP4; runtime TensorRT.",
      "precisions": {
        "supported": [
          "fp16",
          "int8"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "7.1.1 Simulator FP16 Accuracy",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "RKNPU currently does not support FP32 calculations, so the simulator defaults to FP16 computation when quantization is not enabled.",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, float_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "float_dtype: Used to specify the data type of floating in the non-quantized case, the data types currently supported are float16.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "7.1.1 Simulator FP16 Accuracy",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "RKNPU currently does not support FP32 calculations, so the simulator defaults to FP16 computation when quantization is not enabled.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, target_platform",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "'rk3562', 'rk3566', 'rk3568', 'rk3576' and 'rk3588' are currently supported.",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, quantized_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "- w8a8: The weight is 8bit asymmetric quantitative accuracy, and the activation value is 8bit asymmetric quantitative accuracy.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, target_platform",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "'rk3562', 'rk3566', 'rk3568', 'rk3576' and 'rk3588' are currently supported.",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, float_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "float_dtype: Used to specify the data type of floating in the non-quantized case, the data types currently supported are float16.",
                "class": "A"
              }
            ]
          },
          "fp8": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, float_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "float_dtype: Used to specify the data type of floating in the non-quantized case, the data types currently supported are float16.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, quantized_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "quantized_dtype: Quantization type, the quantization types currently supported are \"w8a8\", \"w4a16\", \"w8a16\", \"w4a8\", \"w16a16i\" and \"w16a16i_dfp\".",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, quantized_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "- w4a16: The weight is 4bit asymmetric quantitative accuracy, the activation value is 16bit floating point accuracy. (Only RK3576/RV1126B supported)",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, target_platform",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "'rk3562', 'rk3566', 'rk3568', 'rk3576' and 'rk3588' are currently supported.",
                "class": "A"
              }
            ]
          },
          "fp4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, float_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "float_dtype: Used to specify the data type of floating in the non-quantized case, the data types currently supported are float16.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, quantized_dtype",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "quantized_dtype: Quantization type, the quantization types currently supported are \"w8a8\", \"w4a16\", \"w8a16\", \"w4a8\", \"w16a16i\" and \"w16a16i_dfp\".",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "rknn"
        ],
        "list_class": "A",
        "decided": {
          "rknn": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Rockchip",
                "document": "RKNN SDK User Guide",
                "doc_id": "RK-KF-YF-417",
                "revision": "V2.3.2, 2025-04-03",
                "section": "2.2 RKNPU2 Runtime Library",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "librknnrt.so: The Runtime library for RK3562/RK3566/RK3568/RK3576/RK3588/RV1126B development boards.",
                "class": "A"
              },
              {
                "publisher": "Rockchip",
                "document": "RKNN-Toolkit2 API Reference",
                "doc_id": "RK-YH-YF-412",
                "revision": "V2.3.2, 2025-04-01",
                "section": "2.2 rknn.config, target_platform",
                "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "'rk3562', 'rk3566', 'rk3568', 'rk3576' and 'rk3588' are currently supported.",
                "class": "A"
              }
            ]
          },
          "tensorrt": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorRT supports NVIDIA hardware with compute capability SM 7.5 or higher.",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "neousys_nuvo9531": {
      "platform_id": "neousys_nuvo9531",
      "label": "Neousys Nuvo-9531 (Intel 12th-14th Gen Core)",
      "summary": "Neousys Nuvo-9531 (Intel 12th-14th Gen Core): precisions FP32, FP16 and INT8; runtimes OpenVINO, PyTorch and ONNX Runtime — every listed value documented by its vendor (class A: Intel, Neousys, PyTorch Foundation, Microsoft). Documented as not supported (class A): runtime TensorRT. Not listed, and no vendor document settles it (class E): precisions BF16, FP8, INT4 and FP4.",
      "precisions": {
        "supported": [
          "fp32",
          "fp16",
          "int8"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Intel",
                "document": "OpenVINO Precision Control",
                "doc_id": "openvino-precision-control",
                "revision": "2026.4",
                "section": "Inference Precision",
                "url": "https://docs.openvino.ai/2026/openvino-workflow/running-inference/optimize-inference/precision-control.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "CPU supports f32 inference precision and bf16 on some platforms, GPU supports f32 and f16",
                "class": "A"
              },
              {
                "publisher": "Intel",
                "document": "OpenVINO System Requirements",
                "doc_id": "openvino-system-requirements",
                "revision": "2026.4",
                "section": "CPU, Supported Hardware",
                "url": "https://docs.openvino.ai/2026/about-openvino/release-notes-openvino/system-requirements.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "6th - 14th generation Intel® Core™ processors",
                "class": "A"
              },
              {
                "publisher": "Neousys",
                "document": "Nuvo-9531 Series datasheet",
                "doc_id": "neousys-nuvo-9531-datasheet",
                "revision": "updated 2026-05-06",
                "section": "System Core, Processor",
                "url": "https://neousys-web-bucket.s3.us-west-1.amazonaws.com/datasheet/nuvo-9531-intel-12th-gen-rugged-compact-fanless-computer.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Intel ® 14 th/13 th/12 th -Gen Core™ 35W/ 65W LGA1700 CPU",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Intel",
                "document": "OpenVINO Precision Control",
                "doc_id": "openvino-precision-control",
                "revision": "2026.4",
                "section": "Inference Precision",
                "url": "https://docs.openvino.ai/2026/openvino-workflow/running-inference/optimize-inference/precision-control.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "CPU supports f32 inference precision and bf16 on some platforms, GPU supports f32 and f16",
                "class": "A"
              },
              {
                "publisher": "Neousys",
                "document": "Nuvo-9531 Series datasheet",
                "doc_id": "neousys-nuvo-9531-datasheet",
                "revision": "updated 2026-05-06",
                "section": "System Core, Graphics",
                "url": "https://neousys-web-bucket.s3.us-west-1.amazonaws.com/datasheet/nuvo-9531-intel-12th-gen-rugged-compact-fanless-computer.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Graphics Integrated Intel ® UHD Graphics 770 (32EU) / 730 (24EU)",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Intel",
                "document": "OpenVINO CPU Device",
                "doc_id": "openvino-cpu-device",
                "revision": "2026.4",
                "section": "Supported Model Precision (CPU plugin data types)",
                "url": "https://docs.openvino.ai/2026/openvino-workflow/running-inference/inference-devices-and-modes/cpu-device.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "INT8 (Intel® x86-64)",
                "class": "A"
              },
              {
                "publisher": "Intel",
                "document": "OpenVINO System Requirements",
                "doc_id": "openvino-system-requirements",
                "revision": "2026.4",
                "section": "CPU, Supported Hardware",
                "url": "https://docs.openvino.ai/2026/about-openvino/release-notes-openvino/system-requirements.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "6th - 14th generation Intel® Core™ processors",
                "class": "A"
              },
              {
                "publisher": "Neousys",
                "document": "Nuvo-9531 Series datasheet",
                "doc_id": "neousys-nuvo-9531-datasheet",
                "revision": "updated 2026-05-06",
                "section": "System Core, Processor",
                "url": "https://neousys-web-bucket.s3.us-west-1.amazonaws.com/datasheet/nuvo-9531-intel-12th-gen-rugged-compact-fanless-computer.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Intel ® 14 th/13 th/12 th -Gen Core™ 35W/ 65W LGA1700 CPU",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": false,
            "class": "E",
            "note": "Inferred, not stated: OpenVINO runs BF16 on a CPU only with native bfloat16 support or AVX-512 simulation, Intel's 12th-14th Gen Core datasheets name AVX2 as the latest vector extension, and OpenVINO's GPU plugin lists f32 and f16 only. No Intel sentence says BF16 is unsupported on these parts, so the exclusion is a catalog assertion."
          },
          "fp8": {
            "supported": false,
            "class": "E",
            "note": "No Intel or Neousys document names FP8 for 12th-14th Gen Core CPUs or their UHD graphics; the exclusion is a catalog assertion."
          },
          "int4": {
            "supported": false,
            "class": "E",
            "note": "OpenVINO documents INT4 weight compression for language models without naming a device; no document states INT4 inference on this system's CPU or UHD graphics, so the exclusion is a catalog assertion."
          },
          "fp4": {
            "supported": false,
            "class": "E",
            "note": "OpenVINO's CPU plugin lists MXFP4 as an internal data type without a processor-generation qualifier; nothing states FP4 inference on 12th-14th Gen Core, so the exclusion is a catalog assertion."
          }
        }
      },
      "runtimes": {
        "supported": [
          "openvino",
          "pytorch",
          "onnx"
        ],
        "list_class": "A",
        "decided": {
          "openvino": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Intel",
                "document": "OpenVINO System Requirements",
                "doc_id": "openvino-system-requirements",
                "revision": "2026.4",
                "section": "CPU, Supported Hardware",
                "url": "https://docs.openvino.ai/2026/about-openvino/release-notes-openvino/system-requirements.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "6th - 14th generation Intel® Core™ processors",
                "class": "A"
              },
              {
                "publisher": "Neousys",
                "document": "Nuvo-9531 Series datasheet",
                "doc_id": "neousys-nuvo-9531-datasheet",
                "revision": "updated 2026-05-06",
                "section": "System Core, Processor",
                "url": "https://neousys-web-bucket.s3.us-west-1.amazonaws.com/datasheet/nuvo-9531-intel-12th-gen-rugged-compact-fanless-computer.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "Intel ® 14 th/13 th/12 th -Gen Core™ 35W/ 65W LGA1700 CPU",
                "class": "A"
              }
            ]
          },
          "pytorch": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "PyTorch Foundation",
                "document": "PyTorch — Start Locally",
                "doc_id": "pytorch-get-started-locally",
                "revision": "stable 2.14",
                "section": "Linux, pip installation",
                "url": "https://pytorch.org/get-started/locally/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "choose OS: Linux, Package: Pip, Language: Python and Compute Platform: CPU.",
                "class": "A"
              }
            ]
          },
          "onnx": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "Microsoft",
                "document": "ONNX Runtime — Get started with Python",
                "doc_id": "onnxruntime-get-started-python",
                "section": "Builds",
                "url": "https://onnxruntime.ai/docs/get-started/with-python.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "onnxruntime | CPU (Release) | Windows (x64), Linux (x64, ARM64), Mac (X64),",
                "class": "A"
              },
              {
                "publisher": "Microsoft",
                "document": "ONNX Runtime — OpenVINO Execution Provider",
                "doc_id": "onnxruntime-openvino-ep",
                "revision": "EP v5.8",
                "section": "Support coverage",
                "url": "https://onnxruntime.ai/docs/execution-providers/OpenVINO-ExecutionProvider.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "CPU refers to Intel ® Atom, Core, and Xeon processors. GPU refers to the Intel Integrated Graphics.",
                "class": "A"
              }
            ]
          },
          "tensorrt": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorRT supports NVIDIA hardware with compute capability SM 7.5 or higher.",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "nvidia_l4_server": {
      "platform_id": "nvidia_l4_server",
      "label": "NVIDIA L4 (24GB)",
      "summary": "NVIDIA L4 (24GB): precisions FP32, FP16, INT8, BF16 and FP8; runtime TensorRT — every listed value documented by its vendor (class A: NVIDIA). Not listed, and no vendor document settles it (class E): precisions INT4 and FP4.",
      "precisions": {
        "supported": [
          "fp32",
          "fp16",
          "int8",
          "bf16",
          "fp8"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.9",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.9', devices: ['NVIDIA L40S'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.9 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.9 | NVIDIA L4; NVIDIA L40; NVIDIA L40S |",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.9",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.9', devices: ['NVIDIA L40S'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.9 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.9 | NVIDIA L4; NVIDIA L40; NVIDIA L40S |",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.9",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.9', devices: ['NVIDIA L40S'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.9 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.9 | NVIDIA L4; NVIDIA L40; NVIDIA L40S |",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.9",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.9', devices: ['NVIDIA L40S'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.9 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.9 | NVIDIA L4; NVIDIA L40; NVIDIA L40S |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA L4 Tensor Core GPU Datasheet",
                "doc_id": "nvidia-l4-datasheet",
                "revision": "3408215. AUG24",
                "section": "Specifications",
                "url": "https://dam-cdn.nvd.orangelogic.com/AssetLink/8f4cjsr458mk07kxxt608b1ct0t570x7.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "BFLOAT16 Tensor Core 242 teraFLOPS* FP8 Tensor Core 485 teraFLOPs*",
                "class": "A"
              }
            ]
          },
          "fp8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.9",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.9', devices: ['NVIDIA L40S'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.9 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.9 | NVIDIA L4; NVIDIA L40; NVIDIA L40S |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA L4 Tensor Core GPU Datasheet",
                "doc_id": "nvidia-l4-datasheet",
                "revision": "3408215. AUG24",
                "section": "Specifications",
                "url": "https://dam-cdn.nvd.orangelogic.com/AssetLink/8f4cjsr458mk07kxxt608b1ct0t570x7.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "BFLOAT16 Tensor Core 242 teraFLOPS* FP8 Tensor Core 485 teraFLOPs*",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": false,
            "class": "E",
            "note": "NVIDIA's datasheet lists INT4 Tensor Core throughput for this GPU, but TensorRT, the runtime the catalog lists for it, documents no INT4 precision for its compute capability (TensorRT documents INT4 weight-only quantization for Hopper, SM 90). No vendor statement settles INT4 model deployment here, so the exclusion is a catalog assertion."
          },
          "fp4": {
            "supported": false,
            "class": "E",
            "note": "TensorRT lists FP4 on compute capability 8.9 only in hardware-emulation mode (the hardware does not accelerate FP4 linear operations), and NVIDIA's Tensor Core input-type table gives 8.9 no FP4. Emulation is not counted as support, and no document states FP4 unsupported, so the exclusion is a catalog assertion."
          }
        }
      },
      "runtimes": {
        "supported": [
          "tensorrt"
        ],
        "list_class": "A",
        "decided": {
          "tensorrt": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorRT supports NVIDIA hardware with compute capability SM 7.5 or higher.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.9 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.9 | NVIDIA L4; NVIDIA L40; NVIDIA L40S |",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "nvidia_l40s_server": {
      "platform_id": "nvidia_l40s_server",
      "label": "NVIDIA L40S (48GB)",
      "summary": "NVIDIA L40S (48GB): precisions FP32, FP16, INT8, BF16 and FP8; runtime TensorRT — every listed value documented by its vendor (class A: NVIDIA). Not listed, and no vendor document settles it (class E): precisions INT4 and FP4.",
      "precisions": {
        "supported": [
          "fp32",
          "fp16",
          "int8",
          "bf16",
          "fp8"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.9",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.9', devices: ['NVIDIA L40S'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.9 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.9 | NVIDIA L4; NVIDIA L40; NVIDIA L40S |",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.9",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.9', devices: ['NVIDIA L40S'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.9 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.9 | NVIDIA L4; NVIDIA L40; NVIDIA L40S |",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.9",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.9', devices: ['NVIDIA L40S'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.9 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.9 | NVIDIA L4; NVIDIA L40; NVIDIA L40S |",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.9",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.9', devices: ['NVIDIA L40S'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.9 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.9 | NVIDIA L4; NVIDIA L40; NVIDIA L40S |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA L40S GPU (product page)",
                "doc_id": "nvidia-l40s-product-page",
                "section": "Specifications",
                "url": "https://www.nvidia.com/en-us/data-center/l40s/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "BFLOAT16 Tensor Core TFLOPS | 362.05 I 733* |",
                "class": "A"
              }
            ]
          },
          "fp8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.9",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.9', devices: ['NVIDIA L40S'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.9 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.9 | NVIDIA L4; NVIDIA L40; NVIDIA L40S |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA L40S GPU (product page)",
                "doc_id": "nvidia-l40s-product-page",
                "section": "Specifications",
                "url": "https://www.nvidia.com/en-us/data-center/l40s/",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "FP8 Tensor Core | 733 I 1,466* |",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": false,
            "class": "E",
            "note": "NVIDIA's datasheet lists INT4 Tensor Core throughput for this GPU, but TensorRT, the runtime the catalog lists for it, documents no INT4 precision for its compute capability (TensorRT documents INT4 weight-only quantization for Hopper, SM 90). No vendor statement settles INT4 model deployment here, so the exclusion is a catalog assertion."
          },
          "fp4": {
            "supported": false,
            "class": "E",
            "note": "TensorRT lists FP4 on compute capability 8.9 only in hardware-emulation mode (the hardware does not accelerate FP4 linear operations), and NVIDIA's Tensor Core input-type table gives 8.9 no FP4. Emulation is not counted as support, and no document states FP4 unsupported, so the exclusion is a catalog assertion."
          }
        }
      },
      "runtimes": {
        "supported": [
          "tensorrt"
        ],
        "list_class": "A",
        "decided": {
          "tensorrt": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorRT supports NVIDIA hardware with compute capability SM 7.5 or higher.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.9 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.9 | NVIDIA L4; NVIDIA L40; NVIDIA L40S |",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "nvidia_a2_server": {
      "platform_id": "nvidia_a2_server",
      "label": "NVIDIA A2 (16GB)",
      "summary": "NVIDIA A2 (16GB): precisions FP32, FP16, INT8 and BF16; runtime TensorRT — every listed value documented by its vendor (class A: NVIDIA). Documented as not supported (class A): precisions FP8 and FP4. Not listed, and no vendor document settles it (class E): precision INT4.",
      "precisions": {
        "supported": [
          "fp32",
          "fp16",
          "int8",
          "bf16"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.6",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.6', devices: ['NVIDIA A10'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.6 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.6 | NVIDIA A40; NVIDIA A10; NVIDIA A16; NVIDIA A2 |",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.6",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.6', devices: ['NVIDIA A10'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.6 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.6 | NVIDIA A40; NVIDIA A10; NVIDIA A16; NVIDIA A2 |",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.6",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.6', devices: ['NVIDIA A10'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.6 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.6 | NVIDIA A40; NVIDIA A10; NVIDIA A16; NVIDIA A2 |",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.6",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.6', devices: ['NVIDIA A10'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.6 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.6 | NVIDIA A40; NVIDIA A10; NVIDIA A16; NVIDIA A2 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA A2 Tensor Core GPU Datasheet",
                "doc_id": "nvidia-a2-datasheet",
                "revision": "MAR22",
                "section": "System Specifications",
                "url": "https://www.nvidia.com/content/dam/en-zz/solutions/data-center/a2/pdf/a2-datasheet.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "BFLOAT16 Tensor Core 18 TF | 36 TF¹",
                "class": "A"
              }
            ]
          },
          "fp8": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.6",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.6', devices: ['NVIDIA A10'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.6 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.6 | NVIDIA A40; NVIDIA A10; NVIDIA A16; NVIDIA A2 |",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": false,
            "class": "E",
            "note": "NVIDIA's datasheet lists INT4 Tensor Core throughput for this GPU, but TensorRT, the runtime the catalog lists for it, documents no INT4 precision for its compute capability (TensorRT documents INT4 weight-only quantization for Hopper, SM 90). No vendor statement settles INT4 model deployment here, so the exclusion is a catalog assertion."
          },
          "fp4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.6",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.6', devices: ['NVIDIA A10'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.6 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.6 | NVIDIA A40; NVIDIA A10; NVIDIA A16; NVIDIA A2 |",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "tensorrt"
        ],
        "list_class": "A",
        "decided": {
          "tensorrt": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorRT supports NVIDIA hardware with compute capability SM 7.5 or higher.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.6 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.6 | NVIDIA A40; NVIDIA A10; NVIDIA A16; NVIDIA A2 |",
                "class": "A"
              }
            ]
          }
        }
      }
    },
    "nvidia_a10_server": {
      "platform_id": "nvidia_a10_server",
      "label": "NVIDIA A10 (24GB)",
      "summary": "NVIDIA A10 (24GB): precisions FP32, FP16, INT8 and BF16; runtime TensorRT — every listed value documented by its vendor (class A: NVIDIA). Documented as not supported (class A): precisions FP8 and FP4. Not listed, and no vendor document settles it (class E): precision INT4.",
      "precisions": {
        "supported": [
          "fp32",
          "fp16",
          "int8",
          "bf16"
        ],
        "list_class": "A",
        "decided": {
          "fp32": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.6",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.6', devices: ['NVIDIA A10'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.6 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.6 | NVIDIA A40; NVIDIA A10; NVIDIA A16; NVIDIA A2 |",
                "class": "A"
              }
            ]
          },
          "fp16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.6",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.6', devices: ['NVIDIA A10'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.6 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.6 | NVIDIA A40; NVIDIA A10; NVIDIA A16; NVIDIA A2 |",
                "class": "A"
              }
            ]
          },
          "int8": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.6",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.6', devices: ['NVIDIA A10'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.6 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.6 | NVIDIA A40; NVIDIA A10; NVIDIA A16; NVIDIA A2 |",
                "class": "A"
              }
            ]
          },
          "bf16": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.6",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.6', devices: ['NVIDIA A10'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.6 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.6 | NVIDIA A40; NVIDIA A10; NVIDIA A16; NVIDIA A2 |",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "NVIDIA A10 Tensor Core GPU Datasheet",
                "doc_id": "nvidia-a10-datasheet",
                "revision": "Mar21",
                "section": "Specifications",
                "url": "https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/a10/pdf/a10-datasheet.pdf",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "BFLOAT16 Tensor Core 125 TF | 250 TF*",
                "class": "A"
              }
            ]
          },
          "fp8": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.6",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.6', devices: ['NVIDIA A10'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.6 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.6 | NVIDIA A40; NVIDIA A10; NVIDIA A16; NVIDIA A2 |",
                "class": "A"
              }
            ]
          },
          "int4": {
            "supported": false,
            "class": "E",
            "note": "NVIDIA's datasheet lists INT4 Tensor Core throughput for this GPU, but TensorRT, the runtime the catalog lists for it, documents no INT4 precision for its compute capability (TensorRT documents INT4 weight-only quantization for Hopper, SM 90). No vendor statement settles INT4 model deployment here, so the exclusion is a catalog assertion."
          },
          "fp4": {
            "supported": false,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support, compute capability 8.6",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "computeCapability: '8.6', devices: ['NVIDIA A10'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'Supported', INT8: 'Supported'",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.6 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.6 | NVIDIA A40; NVIDIA A10; NVIDIA A16; NVIDIA A2 |",
                "class": "A"
              }
            ]
          }
        }
      },
      "runtimes": {
        "supported": [
          "tensorrt"
        ],
        "list_class": "A",
        "decided": {
          "tensorrt": {
            "supported": true,
            "class": "A",
            "sources": [
              {
                "publisher": "NVIDIA",
                "document": "TensorRT Support Matrix",
                "doc_id": "tensorrt-support-matrix",
                "revision": "TensorRT 11.3.0, page updated 2026-09-08",
                "section": "GPU Architecture and Precision Support",
                "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "TensorRT supports NVIDIA hardware with compute capability SM 7.5 or higher.",
                "class": "A"
              },
              {
                "publisher": "NVIDIA",
                "document": "CUDA GPU Compute Capability",
                "doc_id": "nvidia-cuda-gpus",
                "section": "Compute capability 8.6 (data center)",
                "url": "https://developer.nvidia.com/cuda-gpus",
                "retrieved": "2026-09-29",
                "verified": "2026-09-29",
                "quote": "8.6 | NVIDIA A40; NVIDIA A10; NVIDIA A16; NVIDIA A2 |",
                "class": "A"
              }
            ]
          }
        }
      }
    }
  },
  "documents": {
    "trt_cc87": {
      "publisher": "NVIDIA",
      "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
      "doc_id": "tensorrt-support-matrix",
      "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
      "section": "GPU Architecture and Precision Support, compute capability 8.7",
      "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "computeCapability: '8.7', devices: ['NVIDIA Jetson AGX Orin'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'N/A', INT8: 'Supported'",
      "class": "A",
      "backs": [
        "jetson_orin_nano precision fp32 supported",
        "jetson_orin_nano precision fp16 supported",
        "jetson_orin_nano precision int8 supported",
        "jetson_orin_nano precision fp8 not supported",
        "jetson_orin_nano precision fp4 not supported",
        "jetson_orin_nano_super precision fp32 supported",
        "jetson_orin_nano_super precision fp16 supported",
        "jetson_orin_nano_super precision int8 supported",
        "jetson_orin_nano_super precision fp8 not supported",
        "jetson_orin_nano_super precision fp4 not supported",
        "jetson_orin_nx_8gb precision fp32 supported",
        "jetson_orin_nx_8gb precision fp16 supported",
        "jetson_orin_nx_8gb precision int8 supported",
        "jetson_orin_nx_8gb precision fp8 not supported",
        "jetson_orin_nx_8gb precision fp4 not supported",
        "jetson_orin_nx precision fp32 supported",
        "jetson_orin_nx precision fp16 supported",
        "jetson_orin_nx precision int8 supported",
        "jetson_orin_nx precision fp8 not supported",
        "jetson_orin_nx precision fp4 not supported",
        "jetson_agx_orin_32gb precision fp32 supported",
        "jetson_agx_orin_32gb precision fp16 supported",
        "jetson_agx_orin_32gb precision int8 supported",
        "jetson_agx_orin_32gb precision fp8 not supported",
        "jetson_agx_orin_32gb precision fp4 not supported",
        "jetson_agx_orin precision fp32 supported",
        "jetson_agx_orin precision fp16 supported",
        "jetson_agx_orin precision int8 supported",
        "jetson_agx_orin precision fp8 not supported",
        "jetson_agx_orin precision fp4 not supported"
      ]
    },
    "trt_cc110": {
      "publisher": "NVIDIA",
      "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
      "doc_id": "tensorrt-support-matrix",
      "revision": "TensorRT 10.16.1 (JetPack 7.2.1 ships 10.16.2), page updated 2026-09-08",
      "section": "GPU Architecture and Precision Support, compute capability 11.0",
      "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "computeCapability: '11.0', devices: ['NVIDIA Jetson AGX Thor'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
      "class": "A",
      "backs": [
        "jetson_thor_t5000 precision fp32 supported",
        "jetson_thor_t5000 precision fp16 supported",
        "jetson_thor_t5000 precision int8 supported",
        "jetson_thor_t5000 precision bf16 supported",
        "jetson_thor_t5000 precision fp8 supported",
        "jetson_thor_t5000 precision fp4 supported",
        "jetson_thor_t5000 runtime tensorrt supported",
        "jetson_thor_t4000 precision fp32 supported",
        "jetson_thor_t4000 precision fp16 supported",
        "jetson_thor_t4000 precision int8 supported",
        "jetson_thor_t4000 precision bf16 supported",
        "jetson_thor_t4000 precision fp8 supported",
        "jetson_thor_t4000 precision fp4 supported",
        "jetson_thor_t4000 runtime tensorrt supported"
      ]
    },
    "trt_cc89": {
      "publisher": "NVIDIA",
      "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
      "doc_id": "tensorrt-support-matrix",
      "revision": "TensorRT 11.3.0, page updated 2026-09-08",
      "section": "GPU Architecture and Precision Support, compute capability 8.9",
      "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "computeCapability: '8.9', devices: ['NVIDIA L40S'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'Supported', FP4: 'Supported', BF16: 'Supported', INT8: 'Supported'",
      "class": "A",
      "backs": [
        "nvidia_l4_server precision fp32 supported",
        "nvidia_l4_server precision fp16 supported",
        "nvidia_l4_server precision int8 supported",
        "nvidia_l4_server precision bf16 supported",
        "nvidia_l4_server precision fp8 supported",
        "nvidia_l40s_server precision fp32 supported",
        "nvidia_l40s_server precision fp16 supported",
        "nvidia_l40s_server precision int8 supported",
        "nvidia_l40s_server precision bf16 supported",
        "nvidia_l40s_server precision fp8 supported"
      ]
    },
    "trt_cc86": {
      "publisher": "NVIDIA",
      "document": "TensorRT Support Matrix — GPU Architecture and Precision Support",
      "doc_id": "tensorrt-support-matrix",
      "revision": "TensorRT 11.3.0, page updated 2026-09-08",
      "section": "GPU Architecture and Precision Support, compute capability 8.6",
      "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "computeCapability: '8.6', devices: ['NVIDIA A10'], TF32: 'Supported', FP32: 'Supported', FP16: 'Supported', FP8: 'N/A', FP4: 'N/A', BF16: 'Supported', INT8: 'Supported'",
      "class": "A",
      "backs": [
        "nvidia_a2_server precision fp32 supported",
        "nvidia_a2_server precision fp16 supported",
        "nvidia_a2_server precision int8 supported",
        "nvidia_a2_server precision bf16 supported",
        "nvidia_a2_server precision fp8 not supported",
        "nvidia_a2_server precision fp4 not supported",
        "nvidia_a10_server precision fp32 supported",
        "nvidia_a10_server precision fp16 supported",
        "nvidia_a10_server precision int8 supported",
        "nvidia_a10_server precision bf16 supported",
        "nvidia_a10_server precision fp8 not supported",
        "nvidia_a10_server precision fp4 not supported"
      ]
    },
    "trt_sm75": {
      "publisher": "NVIDIA",
      "document": "TensorRT Support Matrix",
      "doc_id": "tensorrt-support-matrix",
      "revision": "TensorRT 11.3.0, page updated 2026-09-08",
      "section": "GPU Architecture and Precision Support",
      "url": "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "TensorRT supports NVIDIA hardware with compute capability SM 7.5 or higher.",
      "class": "A",
      "backs": [
        "coral_tpu runtime tensorrt not supported",
        "hailo_8 runtime tensorrt not supported",
        "hailo_8l runtime tensorrt not supported",
        "hailo_10h runtime tensorrt not supported",
        "rk3588_npu runtime tensorrt not supported",
        "rk3588s runtime tensorrt not supported",
        "rk3568 runtime tensorrt not supported",
        "neousys_nuvo9531 runtime tensorrt not supported",
        "nvidia_l4_server runtime tensorrt supported",
        "nvidia_l40s_server runtime tensorrt supported",
        "nvidia_a2_server runtime tensorrt supported",
        "nvidia_a10_server runtime tensorrt supported"
      ]
    },
    "cuda_cc_orin": {
      "publisher": "NVIDIA",
      "document": "CUDA GPU Compute Capability",
      "doc_id": "nvidia-cuda-gpus",
      "section": "Compute capability 8.7 (Jetson)",
      "url": "https://developer.nvidia.com/cuda-gpus",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "8.7 | | | Jetson AGX Orin; Jetson Orin NX; Jetson Orin Nano |",
      "class": "A",
      "backs": [
        "jetson_orin_nano precision fp32 supported",
        "jetson_orin_nano precision fp16 supported",
        "jetson_orin_nano precision int8 supported",
        "jetson_orin_nano precision fp8 not supported",
        "jetson_orin_nano precision int4 supported",
        "jetson_orin_nano precision fp4 not supported",
        "jetson_orin_nano_super precision fp32 supported",
        "jetson_orin_nano_super precision fp16 supported",
        "jetson_orin_nano_super precision int8 supported",
        "jetson_orin_nano_super precision fp8 not supported",
        "jetson_orin_nano_super precision int4 supported",
        "jetson_orin_nano_super precision fp4 not supported",
        "jetson_orin_nx_8gb precision fp32 supported",
        "jetson_orin_nx_8gb precision fp16 supported",
        "jetson_orin_nx_8gb precision int8 supported",
        "jetson_orin_nx_8gb precision fp8 not supported",
        "jetson_orin_nx_8gb precision int4 supported",
        "jetson_orin_nx_8gb precision fp4 not supported",
        "jetson_orin_nx precision fp32 supported",
        "jetson_orin_nx precision fp16 supported",
        "jetson_orin_nx precision int8 supported",
        "jetson_orin_nx precision fp8 not supported",
        "jetson_orin_nx precision int4 supported",
        "jetson_orin_nx precision fp4 not supported",
        "jetson_agx_orin_32gb precision fp32 supported",
        "jetson_agx_orin_32gb precision fp16 supported",
        "jetson_agx_orin_32gb precision int8 supported",
        "jetson_agx_orin_32gb precision fp8 not supported",
        "jetson_agx_orin_32gb precision int4 supported",
        "jetson_agx_orin_32gb precision fp4 not supported",
        "jetson_agx_orin precision fp32 supported",
        "jetson_agx_orin precision fp16 supported",
        "jetson_agx_orin precision int8 supported",
        "jetson_agx_orin precision fp8 not supported",
        "jetson_agx_orin precision int4 supported",
        "jetson_agx_orin precision fp4 not supported"
      ]
    },
    "cuda_cc_thor": {
      "publisher": "NVIDIA",
      "document": "CUDA GPU Compute Capability",
      "doc_id": "nvidia-cuda-gpus",
      "section": "Compute capability 11.0 (Jetson)",
      "url": "https://developer.nvidia.com/cuda-gpus",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "11.0 | | | Jetson T5000; Jetson T4000 |",
      "class": "A",
      "backs": [
        "jetson_thor_t5000 precision fp32 supported",
        "jetson_thor_t5000 precision fp16 supported",
        "jetson_thor_t5000 precision int8 supported",
        "jetson_thor_t5000 precision bf16 supported",
        "jetson_thor_t5000 precision fp8 supported",
        "jetson_thor_t5000 precision int4 supported",
        "jetson_thor_t5000 precision fp4 supported",
        "jetson_thor_t5000 runtime tensorrt supported",
        "jetson_thor_t4000 precision fp32 supported",
        "jetson_thor_t4000 precision fp16 supported",
        "jetson_thor_t4000 precision int8 supported",
        "jetson_thor_t4000 precision bf16 supported",
        "jetson_thor_t4000 precision fp8 supported",
        "jetson_thor_t4000 precision int4 supported",
        "jetson_thor_t4000 precision fp4 supported",
        "jetson_thor_t4000 runtime tensorrt supported"
      ]
    },
    "cuda_cc_ada": {
      "publisher": "NVIDIA",
      "document": "CUDA GPU Compute Capability",
      "doc_id": "nvidia-cuda-gpus",
      "section": "Compute capability 8.9 (data center)",
      "url": "https://developer.nvidia.com/cuda-gpus",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "8.9 | NVIDIA L4; NVIDIA L40; NVIDIA L40S |",
      "class": "A",
      "backs": [
        "nvidia_l4_server precision fp32 supported",
        "nvidia_l4_server precision fp16 supported",
        "nvidia_l4_server precision int8 supported",
        "nvidia_l4_server precision bf16 supported",
        "nvidia_l4_server precision fp8 supported",
        "nvidia_l4_server runtime tensorrt supported",
        "nvidia_l40s_server precision fp32 supported",
        "nvidia_l40s_server precision fp16 supported",
        "nvidia_l40s_server precision int8 supported",
        "nvidia_l40s_server precision bf16 supported",
        "nvidia_l40s_server precision fp8 supported",
        "nvidia_l40s_server runtime tensorrt supported"
      ]
    },
    "cuda_cc_ampere": {
      "publisher": "NVIDIA",
      "document": "CUDA GPU Compute Capability",
      "doc_id": "nvidia-cuda-gpus",
      "section": "Compute capability 8.6 (data center)",
      "url": "https://developer.nvidia.com/cuda-gpus",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "8.6 | NVIDIA A40; NVIDIA A10; NVIDIA A16; NVIDIA A2 |",
      "class": "A",
      "backs": [
        "nvidia_a2_server precision fp32 supported",
        "nvidia_a2_server precision fp16 supported",
        "nvidia_a2_server precision int8 supported",
        "nvidia_a2_server precision bf16 supported",
        "nvidia_a2_server precision fp8 not supported",
        "nvidia_a2_server precision fp4 not supported",
        "nvidia_a2_server runtime tensorrt supported",
        "nvidia_a10_server precision fp32 supported",
        "nvidia_a10_server precision fp16 supported",
        "nvidia_a10_server precision int8 supported",
        "nvidia_a10_server precision bf16 supported",
        "nvidia_a10_server precision fp8 not supported",
        "nvidia_a10_server precision fp4 not supported",
        "nvidia_a10_server runtime tensorrt supported"
      ]
    },
    "orin_nano_super_same_hw": {
      "publisher": "NVIDIA",
      "document": "NVIDIA Jetson Orin Nano Developer Kit Gets a Super Boost",
      "doc_id": "nvidia-blog-orin-nano-super",
      "revision": "Dec 17, 2024",
      "section": "Introduction",
      "url": "https://developer.nvidia.com/blog/nvidia-jetson-orin-nano-developer-kit-gets-a-super-boost/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "With the same hardware architecture, this performance boost is enabled by a new power mode which increases the GPU, memory, and CPU clocks.",
      "class": "A",
      "backs": [
        "jetson_orin_nano_super precision fp32 supported",
        "jetson_orin_nano_super precision fp16 supported",
        "jetson_orin_nano_super precision int8 supported",
        "jetson_orin_nano_super precision fp8 not supported",
        "jetson_orin_nano_super precision int4 supported",
        "jetson_orin_nano_super precision fp4 not supported"
      ]
    },
    "thor_ds_precisions": {
      "publisher": "NVIDIA",
      "document": "NVIDIA Jetson Thor Series Modules Data Sheet",
      "doc_id": "DS-11945-001",
      "revision": "v1.5, June 2026",
      "section": "2.1.4 Compute Features",
      "url": "https://developer.nvidia.com/downloads/assets/embedded/secure/jetson/thor/docs/jetson-thor-series-modules-datasheet_ds-11945-001.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "Blackwell introduces fifth-generation NVIDIA Tensor Cores which offer a wider range of precisions including TensorFloat-32 (TF32), bfloat16, FP16, FP8, FP4, and INT8,",
      "class": "A",
      "backs": [
        "jetson_thor_t5000 precision fp32 supported",
        "jetson_thor_t5000 precision fp16 supported",
        "jetson_thor_t5000 precision int8 supported",
        "jetson_thor_t5000 precision bf16 supported",
        "jetson_thor_t5000 precision fp8 supported",
        "jetson_thor_t5000 precision fp4 supported",
        "jetson_thor_t4000 precision fp32 supported",
        "jetson_thor_t4000 precision fp16 supported",
        "jetson_thor_t4000 precision int8 supported",
        "jetson_thor_t4000 precision bf16 supported",
        "jetson_thor_t4000 precision fp8 supported",
        "jetson_thor_t4000 precision fp4 supported"
      ]
    },
    "edgellm_int4": {
      "publisher": "NVIDIA",
      "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
      "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
      "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
      "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "INT4 AWQ | 4x reduction | Supported | Supported |",
      "class": "A",
      "backs": [
        "jetson_orin_nano precision int4 supported",
        "jetson_orin_nano_super precision int4 supported",
        "jetson_orin_nx_8gb precision int4 supported",
        "jetson_orin_nx precision int4 supported",
        "jetson_agx_orin_32gb precision int4 supported",
        "jetson_agx_orin precision int4 supported",
        "jetson_thor_t5000 precision int4 supported",
        "jetson_thor_t4000 precision int4 supported"
      ]
    },
    "edgellm_fp8": {
      "publisher": "NVIDIA",
      "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
      "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
      "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
      "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "FP8 | 2x reduction | Not available | Supported |",
      "class": "A",
      "backs": [
        "jetson_orin_nano precision fp8 not supported",
        "jetson_orin_nano_super precision fp8 not supported",
        "jetson_orin_nx_8gb precision fp8 not supported",
        "jetson_orin_nx precision fp8 not supported",
        "jetson_agx_orin_32gb precision fp8 not supported",
        "jetson_agx_orin precision fp8 not supported",
        "jetson_thor_t5000 precision fp8 supported",
        "jetson_thor_t4000 precision fp8 supported"
      ]
    },
    "edgellm_nvfp4": {
      "publisher": "NVIDIA",
      "document": "TensorRT Edge-LLM on Jetson (Jetson AI Lab)",
      "doc_id": "jetson-ai-lab-tensorrt-edge-llm",
      "section": "Quantization and Platform Compatibility (Precision | Memory savings | Jetson Orin, CC 8.7 | Jetson Thor)",
      "url": "https://www.jetson-ai-lab.com/tutorials/tensorrt-edge-llm",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "NVFP4 | 4x reduction | Not available | Supported |",
      "class": "A",
      "backs": [
        "jetson_orin_nano precision fp4 not supported",
        "jetson_orin_nano_super precision fp4 not supported",
        "jetson_orin_nx_8gb precision fp4 not supported",
        "jetson_orin_nx precision fp4 not supported",
        "jetson_agx_orin_32gb precision fp4 not supported",
        "jetson_agx_orin precision fp4 not supported",
        "jetson_thor_t5000 precision fp4 supported",
        "jetson_thor_t4000 precision fp4 supported"
      ]
    },
    "l4_ds": {
      "publisher": "NVIDIA",
      "document": "NVIDIA L4 Tensor Core GPU Datasheet",
      "doc_id": "nvidia-l4-datasheet",
      "revision": "3408215. AUG24",
      "section": "Specifications",
      "url": "https://dam-cdn.nvd.orangelogic.com/AssetLink/8f4cjsr458mk07kxxt608b1ct0t570x7.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "BFLOAT16 Tensor Core 242 teraFLOPS* FP8 Tensor Core 485 teraFLOPs*",
      "class": "A",
      "backs": [
        "nvidia_l4_server precision bf16 supported",
        "nvidia_l4_server precision fp8 supported"
      ]
    },
    "l40s_fp8": {
      "publisher": "NVIDIA",
      "document": "NVIDIA L40S GPU (product page)",
      "doc_id": "nvidia-l40s-product-page",
      "section": "Specifications",
      "url": "https://www.nvidia.com/en-us/data-center/l40s/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "FP8 Tensor Core | 733 I 1,466* |",
      "class": "A",
      "backs": [
        "nvidia_l40s_server precision fp8 supported"
      ]
    },
    "l40s_bf16": {
      "publisher": "NVIDIA",
      "document": "NVIDIA L40S GPU (product page)",
      "doc_id": "nvidia-l40s-product-page",
      "section": "Specifications",
      "url": "https://www.nvidia.com/en-us/data-center/l40s/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "BFLOAT16 Tensor Core TFLOPS | 362.05 I 733* |",
      "class": "A",
      "backs": [
        "nvidia_l40s_server precision bf16 supported"
      ]
    },
    "a2_ds_bf16": {
      "publisher": "NVIDIA",
      "document": "NVIDIA A2 Tensor Core GPU Datasheet",
      "doc_id": "nvidia-a2-datasheet",
      "revision": "MAR22",
      "section": "System Specifications",
      "url": "https://www.nvidia.com/content/dam/en-zz/solutions/data-center/a2/pdf/a2-datasheet.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "BFLOAT16 Tensor Core 18 TF | 36 TF¹",
      "class": "A",
      "backs": [
        "nvidia_a2_server precision bf16 supported"
      ]
    },
    "a10_ds_bf16": {
      "publisher": "NVIDIA",
      "document": "NVIDIA A10 Tensor Core GPU Datasheet",
      "doc_id": "nvidia-a10-datasheet",
      "revision": "Mar21",
      "section": "Specifications",
      "url": "https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/a10/pdf/a10-datasheet.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "BFLOAT16 Tensor Core 125 TF | 250 TF*",
      "class": "A",
      "backs": [
        "nvidia_a10_server precision bf16 supported"
      ]
    },
    "jp621_stack": {
      "publisher": "NVIDIA",
      "document": "JetPack 6.2.1 Release Notes",
      "doc_id": "jetpack-6.2.1-release-notes",
      "revision": "JetPack 6.2.1",
      "section": "Additional Release Details",
      "url": "https://docs.nvidia.com/jetson/jetpack/release-notes/index.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "Compute stack with CUDA 12.6, TensorRT 10.3, cuDNN 9.3, VPI 3.2, DLA 3.1, and DLFW 24.0.",
      "class": "A",
      "backs": [
        "jetson_orin_nano runtime tensorrt supported",
        "jetson_orin_nano_super runtime tensorrt supported",
        "jetson_orin_nx_8gb runtime tensorrt supported",
        "jetson_orin_nx runtime tensorrt supported",
        "jetson_agx_orin_32gb runtime tensorrt supported",
        "jetson_agx_orin runtime tensorrt supported"
      ]
    },
    "jp621_orin": {
      "publisher": "NVIDIA",
      "document": "JetPack 6.2.1 Release Notes",
      "doc_id": "jetpack-6.2.1-release-notes",
      "revision": "JetPack 6.2.1",
      "section": "New Features",
      "url": "https://docs.nvidia.com/jetson/jetpack/release-notes/index.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "NVIDIA® Jetson™ Linux 36.4.4 supports all NVIDIA Jetson Orin modules and developer kits",
      "class": "A",
      "backs": [
        "jetson_orin_nano runtime tensorrt supported",
        "jetson_orin_nano_super runtime tensorrt supported",
        "jetson_orin_nx_8gb runtime tensorrt supported",
        "jetson_orin_nx runtime tensorrt supported",
        "jetson_agx_orin_32gb runtime tensorrt supported",
        "jetson_agx_orin runtime tensorrt supported"
      ]
    },
    "jp721": {
      "publisher": "NVIDIA",
      "document": "JetPack SDK Downloads and Notes",
      "doc_id": "jetpack-downloads",
      "revision": "JetPack 7.2.1 / Jetson Linux 39.2.1 (8/11/2026)",
      "section": "Release information, JetPack 7.2.1",
      "url": "https://developer.nvidia.com/embedded/jetpack/downloads",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "NVIDIA JetPack 7.2.1 with Jetson Linux 39.2.1, CUDA 13.2.2, TensorRT 10.16.2,",
      "class": "A",
      "backs": [
        "jetson_thor_t5000 runtime tensorrt supported",
        "jetson_thor_t4000 runtime tensorrt supported"
      ]
    },
    "jp7_orin_thor": {
      "publisher": "NVIDIA",
      "document": "NVIDIA JetPack SDK",
      "doc_id": "jetpack-sdk",
      "section": "JetPack 7",
      "url": "https://developer.nvidia.com/embedded/jetpack",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "JetPack 7 gives you full support for the NVIDIA Jetson Orin™ and Thor™ platform,",
      "class": "A",
      "backs": [
        "jetson_thor_t5000 runtime tensorrt supported",
        "jetson_thor_t4000 runtime tensorrt supported"
      ]
    },
    "thor_tb_frameworks": {
      "publisher": "NVIDIA",
      "document": "NVIDIA Jetson Thor Series Technical Brief",
      "doc_id": "TB-12572-001",
      "revision": "v1.0, August 29, 2025",
      "section": "JetPack 7 (p.28)",
      "url": "https://dam-cdn.nvd.orangelogic.com/AssetLink/v686u306w26ftuk863xjeimilimtx55e.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "It supports leading generative AI serving frameworks like vLLM, SGLang, TensorRT, TensorRT- LLM, PyTorch, Llama.cpp, and MLC for efficient real-time model deployment.",
      "class": "A",
      "backs": [
        "jetson_thor_t5000 runtime llama_cpp supported",
        "jetson_thor_t5000 runtime pytorch supported",
        "jetson_thor_t4000 runtime llama_cpp supported",
        "jetson_thor_t4000 runtime pytorch supported"
      ]
    },
    "pt_jetson_orin": {
      "publisher": "NVIDIA",
      "document": "Installing PyTorch for Jetson Platform",
      "doc_id": "nvidia-install-pytorch-jetson",
      "revision": "page updated 2026-09-08",
      "section": "1. Overview > Jetson AGX Orin",
      "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform/index.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "The NVIDIA Jetson AGX Orin Developer Kit includes a high-performance, power-efficient Jetson AGX Orin module, and can emulate the other Jetson modules.",
      "class": "A",
      "backs": [
        "jetson_orin_nano runtime pytorch supported",
        "jetson_orin_nano_super runtime pytorch supported",
        "jetson_orin_nx_8gb runtime pytorch supported",
        "jetson_orin_nx runtime pytorch supported",
        "jetson_agx_orin_32gb runtime pytorch supported",
        "jetson_agx_orin runtime pytorch supported"
      ]
    },
    "pt_jetson_jp62": {
      "publisher": "NVIDIA",
      "document": "PyTorch for Jetson Platform — release notes, compatibility table",
      "doc_id": "nvidia-pytorch-jetson-release-notes",
      "revision": "page updated 2026-09-08",
      "section": "Compatibility (PyTorch version | NVIDIA container | … | JetPack)",
      "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform-release-notes/pytorch-jetson-rel.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "2.8.0a0+5228986c39 | 25.06 | - | 6.2 |",
      "class": "A",
      "backs": [
        "jetson_orin_nano runtime pytorch supported",
        "jetson_orin_nano_super runtime pytorch supported",
        "jetson_orin_nx_8gb runtime pytorch supported",
        "jetson_orin_nx runtime pytorch supported",
        "jetson_agx_orin_32gb runtime pytorch supported",
        "jetson_agx_orin runtime pytorch supported"
      ]
    },
    "pt_jetson_jp71": {
      "publisher": "NVIDIA",
      "document": "PyTorch for Jetson Platform — release notes, compatibility table",
      "doc_id": "nvidia-pytorch-jetson-release-notes",
      "revision": "page updated 2026-09-08",
      "section": "Compatibility (PyTorch version | NVIDIA container | … | JetPack)",
      "url": "https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform-release-notes/pytorch-jetson-rel.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "2.14.0a0+4fdf77b940 | 26.08 | - | 7.1 |",
      "class": "A",
      "backs": [
        "jetson_thor_t5000 runtime pytorch supported",
        "jetson_thor_t4000 runtime pytorch supported"
      ]
    },
    "ort_jetson": {
      "publisher": "Microsoft",
      "document": "ONNX Runtime — Build with different execution providers",
      "doc_id": "onnxruntime-build-eps",
      "section": "TensorRT execution provider: NVIDIA Jetson",
      "url": "https://onnxruntime.ai/docs/build/eps.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "NVIDIA Jetson TX1/TX2/Nano/Xavier/Orin",
      "class": "A",
      "backs": [
        "jetson_orin_nano runtime onnx supported",
        "jetson_orin_nano_super runtime onnx supported",
        "jetson_orin_nx_8gb runtime onnx supported",
        "jetson_orin_nx runtime onnx supported",
        "jetson_agx_orin_32gb runtime onnx supported",
        "jetson_agx_orin runtime onnx supported"
      ]
    },
    "ail_gemma4": {
      "publisher": "NVIDIA",
      "document": "Gemma 4 on Jetson (Jetson AI Lab)",
      "doc_id": "jetson-ai-lab-gemma4",
      "section": "Overview",
      "url": "https://www.jetson-ai-lab.com/tutorials/gemma4-on-jetson",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "The full family is supported on Jetson through both vLLM and llama.cpp. All of the models are supported on Orin and Thor,",
      "class": "A",
      "backs": [
        "jetson_orin_nano runtime llama_cpp supported",
        "jetson_orin_nano_super runtime llama_cpp supported",
        "jetson_orin_nx_8gb runtime llama_cpp supported",
        "jetson_orin_nx runtime llama_cpp supported",
        "jetson_agx_orin_32gb runtime llama_cpp supported",
        "jetson_agx_orin runtime llama_cpp supported",
        "jetson_thor_t5000 runtime llama_cpp supported",
        "jetson_thor_t4000 runtime llama_cpp supported"
      ]
    },
    "llamacpp_orin": {
      "publisher": "ggml-org",
      "document": "llama.cpp documentation — MobileVLM",
      "doc_id": "llama-cpp-docs-mobilevlm",
      "revision": "master, 2026-09-29",
      "section": "Orin compile and run",
      "url": "https://raw.githubusercontent.com/ggml-org/llama.cpp/master/docs/multimodal/MobileVLM.md",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "- [x] run MobileVLM on `Jetson Orin`",
      "class": "A",
      "backs": [
        "jetson_orin_nano runtime llama_cpp supported",
        "jetson_orin_nano_super runtime llama_cpp supported",
        "jetson_orin_nx_8gb runtime llama_cpp supported",
        "jetson_orin_nx runtime llama_cpp supported",
        "jetson_agx_orin_32gb runtime llama_cpp supported",
        "jetson_agx_orin runtime llama_cpp supported"
      ]
    },
    "coral_only_int8": {
      "publisher": "Google",
      "document": "TensorFlow models on the Edge TPU",
      "doc_id": "coral-edgetpu-models-intro",
      "section": "Compatibility overview",
      "url": "https://coral.ai/docs/edgetpu/models-intro/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "It supports only TensorFlow Lite models that are fully 8-bit quantized and then compiled specifically for the Edge TPU.",
      "class": "A",
      "backs": [
        "coral_tpu precision fp32 not supported",
        "coral_tpu precision fp16 not supported",
        "coral_tpu precision int8 supported",
        "coral_tpu precision bf16 not supported",
        "coral_tpu precision fp8 not supported",
        "coral_tpu precision int4 not supported",
        "coral_tpu precision fp4 not supported"
      ]
    },
    "coral_int8_params": {
      "publisher": "Google",
      "document": "TensorFlow models on the Edge TPU",
      "doc_id": "coral-edgetpu-models-intro",
      "section": "Model requirements",
      "url": "https://coral.ai/docs/edgetpu/models-intro/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "Tensor parameters are quantized (8-bit fixed-point numbers; int8 or uint8).",
      "class": "A",
      "backs": [
        "coral_tpu precision int8 supported"
      ]
    },
    "coral_float_incompatible": {
      "publisher": "Google",
      "document": "TensorFlow models on the Edge TPU",
      "doc_id": "coral-edgetpu-models-intro",
      "section": "Quantization",
      "url": "https://coral.ai/docs/edgetpu/models-intro/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "By default, the TensorFlow Lite converter leaves those operations in their float format, which is not compatible with the Edge TPU.",
      "class": "A",
      "backs": [
        "coral_tpu precision fp32 not supported",
        "coral_tpu precision fp16 not supported",
        "coral_tpu precision bf16 not supported",
        "coral_tpu precision fp8 not supported",
        "coral_tpu precision fp4 not supported"
      ]
    },
    "coral_tflite_only": {
      "publisher": "Google",
      "document": "Edge TPU frequently asked questions",
      "doc_id": "coral-edgetpu-faq",
      "section": "What machine learning frameworks does the Edge TPU support?",
      "url": "https://coral.ai/docs/edgetpu/faq/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "TensorFlow Lite only.",
      "class": "A",
      "backs": [
        "coral_tpu runtime tensorrt not supported",
        "coral_tpu runtime pytorch not supported",
        "coral_tpu runtime onnx not supported",
        "coral_tpu runtime llama_cpp not supported",
        "coral_tpu runtime hailo_sdk not supported",
        "coral_tpu runtime rknn not supported",
        "coral_tpu runtime openvino not supported"
      ]
    },
    "coral_tflite_exec": {
      "publisher": "Google",
      "document": "Edge TPU inferencing overview",
      "doc_id": "coral-edgetpu-inference",
      "section": "Intro + note",
      "url": "https://coral.ai/docs/edgetpu/inference/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "All inferencing with the Edge TPU is executed with TensorFlow Lite libraries.",
      "class": "A",
      "backs": [
        "coral_tpu runtime edge_tpu supported"
      ]
    },
    "coral_runtime": {
      "publisher": "Google",
      "document": "Edge TPU inferencing overview",
      "doc_id": "coral-edgetpu-inference",
      "section": "General-purpose operating systems",
      "url": "https://coral.ai/docs/edgetpu/inference/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "Regardless of the language you choose, you need to install the Edge TPU runtime (libedgetpu.so), as documented in the setup for each Coral device.",
      "class": "A",
      "backs": [
        "coral_tpu runtime edge_tpu supported"
      ]
    },
    "h8_int_precisions": {
      "publisher": "Hailo",
      "document": "Hailo-8 AI Processor Datasheet",
      "doc_id": "hailo-8-datasheet",
      "revision": "Rev 1.9.2, April 2026",
      "section": "5.3 NN Core",
      "url": "https://hailo.ai/hailo-files/hailo-8-datasheet-en/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "8/16-bit precision for both weights and activations, configurable at layer granularity.",
      "class": "A",
      "backs": [
        "hailo_8 precision fp32 not supported",
        "hailo_8 precision fp16 not supported",
        "hailo_8 precision int8 supported",
        "hailo_8 precision bf16 not supported",
        "hailo_8 precision fp8 not supported",
        "hailo_8 precision fp4 not supported"
      ]
    },
    "h8_w4": {
      "publisher": "Hailo",
      "document": "Hailo-8 AI Processor Datasheet",
      "doc_id": "hailo-8-datasheet",
      "revision": "Rev 1.9.2, April 2026",
      "section": "5.3 NN Core",
      "url": "https://hailo.ai/hailo-files/hailo-8-datasheet-en/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "4-bit precision for weights, configurable at layer granularity.",
      "class": "A",
      "backs": [
        "hailo_8 precision int4 supported"
      ]
    },
    "h8_m2_int8": {
      "publisher": "Hailo",
      "document": "Hailo-8 M.2 Key M ET Module Data Sheet",
      "doc_id": "hailo-8-m2-key-m-et-datasheet",
      "revision": "Rev 2.5, November 2025",
      "section": "Table 1 Key properties",
      "url": "https://hailo.ai/hailo-files/hailo-8-m-2-key-m-et-datasheet-en/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "Peak Performance (INT8) 26 TOPS",
      "class": "A",
      "backs": [
        "hailo_8 precision int8 supported"
      ]
    },
    "h8l_int_precisions": {
      "publisher": "Hailo",
      "document": "Hailo-8L AI Processor Datasheet (industrial)",
      "doc_id": "hailo-8l-industrial-datasheet",
      "revision": "Rev 1.9.3, April 2026",
      "section": "5.3 NN Core",
      "url": "https://hailo.ai/hailo-files/hailo-8l-industrial-datasheet-en/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "8/16-bit precision for both weights and activations, configurable at layer granularity.",
      "class": "A",
      "backs": [
        "hailo_8l precision fp32 not supported",
        "hailo_8l precision fp16 not supported",
        "hailo_8l precision int8 supported",
        "hailo_8l precision bf16 not supported",
        "hailo_8l precision fp8 not supported",
        "hailo_8l precision fp4 not supported"
      ]
    },
    "h8l_w4": {
      "publisher": "Hailo",
      "document": "Hailo-8L AI Processor Datasheet (industrial)",
      "doc_id": "hailo-8l-industrial-datasheet",
      "revision": "Rev 1.9.3, April 2026",
      "section": "5.3 NN Core",
      "url": "https://hailo.ai/hailo-files/hailo-8l-industrial-datasheet-en/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "4-bit precision for weights, configurable at layer granularity.",
      "class": "A",
      "backs": [
        "hailo_8l precision int4 supported"
      ]
    },
    "h8l_m2_int8": {
      "publisher": "Hailo",
      "document": "Hailo-8L M.2 Key B+M ET Module Data Sheet",
      "doc_id": "hailo-8l-m2-key-bm-et-datasheet",
      "revision": "Rev 4.0, November 2025",
      "section": "Table 1 Key properties",
      "url": "https://hailo.ai/hailo-files/hailo-8l-m-2-key-b-m-et-datasheet-en/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "Peak Performance (INT8) 13 TOPS",
      "class": "A",
      "backs": [
        "hailo_8l precision int8 supported"
      ]
    },
    "h10h_precisions": {
      "publisher": "Hailo",
      "document": "Hailo-10H AI Accelerator Datasheet",
      "doc_id": "hailo-10h-datasheet",
      "revision": "Rev 1.0, July 2025",
      "section": "4.3 NN Core",
      "url": "https://hailo.ai/hailo-files/hailo-10h-datasheet-en/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "Precision configurability, support 4/8/16-bit precision for weights and activations",
      "class": "A",
      "backs": [
        "hailo_10h precision fp32 not supported",
        "hailo_10h precision fp16 not supported",
        "hailo_10h precision int8 supported",
        "hailo_10h precision bf16 not supported",
        "hailo_10h precision fp8 not supported",
        "hailo_10h precision int4 supported",
        "hailo_10h precision fp4 not supported"
      ]
    },
    "h10h_tops": {
      "publisher": "Hailo",
      "document": "Hailo-10H AI Accelerator Product Brief",
      "doc_id": "hailo-10h-product-brief",
      "revision": "7/25",
      "section": "Key Features & Benefits",
      "url": "https://hailo.ai/hailo-files/hailo-10h-product-brief-en/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "40|20 TOPS (INT4|8)",
      "class": "A",
      "backs": [
        "hailo_10h precision int8 supported",
        "hailo_10h precision int4 supported"
      ]
    },
    "hmz_integer": {
      "publisher": "Hailo",
      "document": "Hailo Model Zoo — Model Optimization",
      "doc_id": "hailo-model-zoo-optimization",
      "revision": "master a20777477e (2025-09-28); same in v2.19.1",
      "section": "Optimization Workflow",
      "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/OPTIMIZATION.rst",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "Quantization includes compressing the model from floating point to integer representation of the weights and activations (4/8/16 bits)",
      "class": "A",
      "backs": [
        "hailo_8 precision fp32 not supported",
        "hailo_8 precision fp16 not supported",
        "hailo_8 precision bf16 not supported",
        "hailo_8 precision fp8 not supported",
        "hailo_8 precision fp4 not supported",
        "hailo_8l precision fp32 not supported",
        "hailo_8l precision fp16 not supported",
        "hailo_8l precision bf16 not supported",
        "hailo_8l precision fp8 not supported",
        "hailo_8l precision fp4 not supported",
        "hailo_10h precision fp32 not supported",
        "hailo_10h precision fp16 not supported",
        "hailo_10h precision bf16 not supported",
        "hailo_10h precision fp8 not supported",
        "hailo_10h precision fp4 not supported"
      ]
    },
    "hmz_float_emulation": {
      "publisher": "Hailo",
      "document": "Hailo Model Zoo — Getting Started",
      "doc_id": "hailo-model-zoo-getting-started",
      "revision": "master 757435a9b6 (2026-03-30)",
      "section": "Optimization",
      "url": "https://raw.githubusercontent.com/hailo-ai/hailo_model_zoo/master/docs/GETTING_STARTED.rst",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "Full-precision / emulation runs on GPU.",
      "class": "A",
      "backs": [
        "hailo_8 precision fp32 not supported",
        "hailo_8 precision fp16 not supported",
        "hailo_8 precision bf16 not supported",
        "hailo_8 precision fp8 not supported",
        "hailo_8 precision fp4 not supported",
        "hailo_8l precision fp32 not supported",
        "hailo_8l precision fp16 not supported",
        "hailo_8l precision bf16 not supported",
        "hailo_8l precision fp8 not supported",
        "hailo_8l precision fp4 not supported",
        "hailo_10h precision fp32 not supported",
        "hailo_10h precision fp16 not supported",
        "hailo_10h precision bf16 not supported",
        "hailo_10h precision fp8 not supported",
        "hailo_10h precision fp4 not supported"
      ]
    },
    "hailort_devices": {
      "publisher": "Hailo",
      "document": "HailoRT README",
      "doc_id": "hailort-readme",
      "revision": "master f51959034a (2026-08-16), release v5.4.0",
      "section": "Supported devices",
      "url": "https://raw.githubusercontent.com/hailo-ai/hailort/master/README.md",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "Note: The `master` branch supports only the Hailo-10 and Hailo-15 device families. For Hailo-8, Hailo-8R, and Hailo-8L devices, please use the `hailo8` branch.",
      "class": "A",
      "backs": [
        "hailo_8 runtime hailo_sdk supported",
        "hailo_8l runtime hailo_sdk supported",
        "hailo_10h runtime hailo_sdk supported"
      ]
    },
    "hailort_10h": {
      "publisher": "Hailo",
      "document": "HailoRT 5.4.0 User Guide — Installation Inside a Docker Container",
      "doc_id": "hailort-5.4.0-user-guide",
      "revision": "HailoRT 5.4.0",
      "section": "Running Hailo-10H and Hailo-8 Side by Side",
      "url": "https://hailo.ai/docs/31771-hailort_5.4.0_sphinx-2/html/install/docker.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "Start a container that contains the Hailo-10H HailoRT library (v5.x) and expose only the Hailo-10H device to it.",
      "class": "A",
      "backs": [
        "hailo_10h runtime hailo_sdk supported"
      ]
    },
    "rknn_runtime": {
      "publisher": "Rockchip",
      "document": "RKNN SDK User Guide",
      "doc_id": "RK-KF-YF-417",
      "revision": "V2.3.2, 2025-04-03",
      "section": "2.2 RKNPU2 Runtime Library",
      "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "librknnrt.so: The Runtime library for RK3562/RK3566/RK3568/RK3576/RK3588/RV1126B development boards.",
      "class": "A",
      "backs": [
        "rk3588_npu runtime rknn supported",
        "rk3588s runtime rknn supported",
        "rk3568 runtime rknn supported"
      ]
    },
    "rknn_rk3588s_alias": {
      "publisher": "Rockchip",
      "document": "RKNN SDK User Guide",
      "doc_id": "RK-KF-YF-417",
      "revision": "V2.3.2, 2025-04-03",
      "section": "1.3 Supported Hardware Platforms",
      "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "RK3588 is used to indicate RK3588/RK3588S series in this document.",
      "class": "A",
      "backs": [
        "rk3588s precision fp32 not supported",
        "rk3588s precision fp16 supported",
        "rk3588s precision int8 supported",
        "rk3588s precision bf16 not supported",
        "rk3588s precision fp8 not supported",
        "rk3588s precision int4 not supported",
        "rk3588s precision fp4 not supported",
        "rk3588s runtime rknn supported"
      ]
    },
    "rknn_no_fp32": {
      "publisher": "Rockchip",
      "document": "RKNN SDK User Guide",
      "doc_id": "RK-KF-YF-417",
      "revision": "V2.3.2, 2025-04-03",
      "section": "7.1.1 Simulator FP16 Accuracy",
      "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/02_Rockchip_RKNPU_User_Guide_RKNN_SDK_V2.3.2_EN.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "RKNPU currently does not support FP32 calculations, so the simulator defaults to FP16 computation when quantization is not enabled.",
      "class": "A",
      "backs": [
        "rk3588_npu precision fp32 not supported",
        "rk3588_npu precision fp16 supported",
        "rk3588s precision fp32 not supported",
        "rk3588s precision fp16 supported",
        "rk3568 precision fp32 not supported",
        "rk3568 precision fp16 supported"
      ]
    },
    "rknn_w8a8": {
      "publisher": "Rockchip",
      "document": "RKNN-Toolkit2 API Reference",
      "doc_id": "RK-YH-YF-412",
      "revision": "V2.3.2, 2025-04-01",
      "section": "2.2 rknn.config, quantized_dtype",
      "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "- w8a8: The weight is 8bit asymmetric quantitative accuracy, and the activation value is 8bit asymmetric quantitative accuracy.",
      "class": "A",
      "backs": [
        "rk3588_npu precision int8 supported",
        "rk3588s precision int8 supported",
        "rk3568 precision int8 supported"
      ]
    },
    "rknn_qtypes": {
      "publisher": "Rockchip",
      "document": "RKNN-Toolkit2 API Reference",
      "doc_id": "RK-YH-YF-412",
      "revision": "V2.3.2, 2025-04-01",
      "section": "2.2 rknn.config, quantized_dtype",
      "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "quantized_dtype: Quantization type, the quantization types currently supported are \"w8a8\", \"w4a16\", \"w8a16\", \"w4a8\", \"w16a16i\" and \"w16a16i_dfp\".",
      "class": "A",
      "backs": [
        "rk3588_npu precision fp8 not supported",
        "rk3588_npu precision fp4 not supported",
        "rk3588s precision fp8 not supported",
        "rk3588s precision fp4 not supported",
        "rk3568 precision fp8 not supported",
        "rk3568 precision fp4 not supported"
      ]
    },
    "rknn_w4a16": {
      "publisher": "Rockchip",
      "document": "RKNN-Toolkit2 API Reference",
      "doc_id": "RK-YH-YF-412",
      "revision": "V2.3.2, 2025-04-01",
      "section": "2.2 rknn.config, quantized_dtype",
      "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "- w4a16: The weight is 4bit asymmetric quantitative accuracy, the activation value is 16bit floating point accuracy. (Only RK3576/RV1126B supported)",
      "class": "A",
      "backs": [
        "rk3588_npu precision int4 not supported",
        "rk3588s precision int4 not supported",
        "rk3568 precision int4 not supported"
      ]
    },
    "rknn_float16": {
      "publisher": "Rockchip",
      "document": "RKNN-Toolkit2 API Reference",
      "doc_id": "RK-YH-YF-412",
      "revision": "V2.3.2, 2025-04-01",
      "section": "2.2 rknn.config, float_dtype",
      "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "float_dtype: Used to specify the data type of floating in the non-quantized case, the data types currently supported are float16.",
      "class": "A",
      "backs": [
        "rk3588_npu precision fp16 supported",
        "rk3588_npu precision bf16 not supported",
        "rk3588_npu precision fp8 not supported",
        "rk3588_npu precision fp4 not supported",
        "rk3588s precision fp16 supported",
        "rk3588s precision bf16 not supported",
        "rk3588s precision fp8 not supported",
        "rk3588s precision fp4 not supported",
        "rk3568 precision fp16 supported",
        "rk3568 precision bf16 not supported",
        "rk3568 precision fp8 not supported",
        "rk3568 precision fp4 not supported"
      ]
    },
    "rknn_targets": {
      "publisher": "Rockchip",
      "document": "RKNN-Toolkit2 API Reference",
      "doc_id": "RK-YH-YF-412",
      "revision": "V2.3.2, 2025-04-01",
      "section": "2.2 rknn.config, target_platform",
      "url": "https://raw.githubusercontent.com/airockchip/rknn-toolkit2/master/doc/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "'rk3562', 'rk3566', 'rk3568', 'rk3576' and 'rk3588' are currently supported.",
      "class": "A",
      "backs": [
        "rk3588_npu precision fp16 supported",
        "rk3588_npu precision int8 supported",
        "rk3588_npu runtime rknn supported",
        "rk3588s precision fp16 supported",
        "rk3588s precision int8 supported",
        "rk3588s runtime rknn supported",
        "rk3568 precision fp16 supported",
        "rk3568 precision int8 supported",
        "rk3568 precision int4 not supported",
        "rk3568 runtime rknn supported"
      ]
    },
    "rkllm_rk3588": {
      "publisher": "Rockchip",
      "document": "RKLLM SDK User Guide",
      "doc_id": "rkllm-sdk-user-guide",
      "revision": "V1.3.1, 2026-09-28",
      "section": "3.1.3 build, quantized_dtype",
      "url": "https://raw.githubusercontent.com/airockchip/rknn-llm/main/doc/Rockchip_RKLLM_SDK_EN_1.3.1.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "The RK3588 platform supports four quantization types: \"w8a8\", \"w8a8_g128\", \"w8a8_g256\", and \"w8a8_g512\".",
      "class": "A",
      "backs": [
        "rk3588_npu precision int4 not supported",
        "rk3588s precision int4 not supported"
      ]
    },
    "nuvo_cpus": {
      "publisher": "Neousys",
      "document": "Nuvo-9531 Series datasheet",
      "doc_id": "neousys-nuvo-9531-datasheet",
      "revision": "updated 2026-05-06",
      "section": "System Core, Processor",
      "url": "https://neousys-web-bucket.s3.us-west-1.amazonaws.com/datasheet/nuvo-9531-intel-12th-gen-rugged-compact-fanless-computer.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "Intel ® 14 th/13 th/12 th -Gen Core™ 35W/ 65W LGA1700 CPU",
      "class": "A",
      "backs": [
        "neousys_nuvo9531 precision fp32 supported",
        "neousys_nuvo9531 precision int8 supported",
        "neousys_nuvo9531 runtime openvino supported"
      ]
    },
    "nuvo_igpu": {
      "publisher": "Neousys",
      "document": "Nuvo-9531 Series datasheet",
      "doc_id": "neousys-nuvo-9531-datasheet",
      "revision": "updated 2026-05-06",
      "section": "System Core, Graphics",
      "url": "https://neousys-web-bucket.s3.us-west-1.amazonaws.com/datasheet/nuvo-9531-intel-12th-gen-rugged-compact-fanless-computer.pdf",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "Graphics Integrated Intel ® UHD Graphics 770 (32EU) / 730 (24EU)",
      "class": "A",
      "backs": [
        "neousys_nuvo9531 precision fp16 supported"
      ]
    },
    "ov_sysreq": {
      "publisher": "Intel",
      "document": "OpenVINO System Requirements",
      "doc_id": "openvino-system-requirements",
      "revision": "2026.4",
      "section": "CPU, Supported Hardware",
      "url": "https://docs.openvino.ai/2026/about-openvino/release-notes-openvino/system-requirements.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "6th - 14th generation Intel® Core™ processors",
      "class": "A",
      "backs": [
        "neousys_nuvo9531 precision fp32 supported",
        "neousys_nuvo9531 precision int8 supported",
        "neousys_nuvo9531 runtime openvino supported"
      ]
    },
    "ov_precisions": {
      "publisher": "Intel",
      "document": "OpenVINO Precision Control",
      "doc_id": "openvino-precision-control",
      "revision": "2026.4",
      "section": "Inference Precision",
      "url": "https://docs.openvino.ai/2026/openvino-workflow/running-inference/optimize-inference/precision-control.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "CPU supports f32 inference precision and bf16 on some platforms, GPU supports f32 and f16",
      "class": "A",
      "backs": [
        "neousys_nuvo9531 precision fp32 supported",
        "neousys_nuvo9531 precision fp16 supported"
      ]
    },
    "ov_cpu_int8": {
      "publisher": "Intel",
      "document": "OpenVINO CPU Device",
      "doc_id": "openvino-cpu-device",
      "revision": "2026.4",
      "section": "Supported Model Precision (CPU plugin data types)",
      "url": "https://docs.openvino.ai/2026/openvino-workflow/running-inference/inference-devices-and-modes/cpu-device.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "INT8 (Intel® x86-64)",
      "class": "A",
      "backs": [
        "neousys_nuvo9531 precision int8 supported"
      ]
    },
    "pytorch_cpu": {
      "publisher": "PyTorch Foundation",
      "document": "PyTorch — Start Locally",
      "doc_id": "pytorch-get-started-locally",
      "revision": "stable 2.14",
      "section": "Linux, pip installation",
      "url": "https://pytorch.org/get-started/locally/",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "choose OS: Linux, Package: Pip, Language: Python and Compute Platform: CPU.",
      "class": "A",
      "backs": [
        "neousys_nuvo9531 runtime pytorch supported"
      ]
    },
    "ort_x64": {
      "publisher": "Microsoft",
      "document": "ONNX Runtime — Get started with Python",
      "doc_id": "onnxruntime-get-started-python",
      "section": "Builds",
      "url": "https://onnxruntime.ai/docs/get-started/with-python.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "onnxruntime | CPU (Release) | Windows (x64), Linux (x64, ARM64), Mac (X64),",
      "class": "A",
      "backs": [
        "neousys_nuvo9531 runtime onnx supported"
      ]
    },
    "ort_openvino": {
      "publisher": "Microsoft",
      "document": "ONNX Runtime — OpenVINO Execution Provider",
      "doc_id": "onnxruntime-openvino-ep",
      "revision": "EP v5.8",
      "section": "Support coverage",
      "url": "https://onnxruntime.ai/docs/execution-providers/OpenVINO-ExecutionProvider.html",
      "retrieved": "2026-09-29",
      "verified": "2026-09-29",
      "quote": "CPU refers to Intel ® Atom, Core, and Xeon processors. GPU refers to the Intel Integrated Graphics.",
      "class": "A",
      "backs": [
        "neousys_nuvo9531 runtime onnx supported"
      ]
    }
  }
}
