Decode 8 × 1080p30 H.265 cameras and still run detection? On Orin Nano the decoder sits at 73 % of its published 11 streams; the GPU does none of the decoding.

Answer

Eight 1080p30 H.265 streams are 497.7 MP/s against a published 11 × 1080p30 row (684 MP/s at this profile): 73 % of the decoder, PASS, class A because 1080p30 is a printed datasheet point. Decode runs on NVDEC, a separate block; it takes nothing from the GPU.

Whether there is "power for object detection" is therefore an inference question, not a decode question. With the engine's default detector (yolov8s INT8 640x640 @ 15 fps) the inference row below shows 120 detection frames/s required against roughly 23 available (513 %, FAIL). Lower the detection frame rate, use a smaller variant, or run detection on motion only to open the margin; the live link lets you change those inputs.

The thread was asked about the original Jetson Nano in 2021. This page answers it for the current-generation Jetson Orin Nano, the module the question maps to today; the principle in NVIDIA’s reply (NVDEC is separate silicon, the GPU is untouched by decode) holds for both.

The question

• 2021-11-02 · thread 193819

  “If you decode all 8x1080p @ 30fps h.265 streams (8 ip cameras), will there be power for object detection?”

• 2021-11-03 · reply by NVIDIA staff (DaneLLL)

  “On Jetson platforms, there is individual hardware encoder/decoder so the decoding is not execute on GPU. GPU is free for object detection.”

Every stage as a constraint

Required against available in the same unit; PASS below 80 %, NEAR_LIMIT from 80 %, FAIL at 100 %. Evidence class and confidence per row; the source line is the document the available figure came from.

Headroom ladder for this workload: 1 cameras comfortable (≤ 60 %) · 1 likely viable (≤ 85 %) · fails at 2. Overall confidence LOW (weakest stage).

Assumptions the engine used

• Orin NVDEC clock is identical in every nvpmodel preset; the power mode does not change decode capacity.
• Single-stream external benchmark 30 fps (benchmark_exact, YOLOv8 Variant Benchmarks); multi-stream efficiency 0.78 at 8 streams.
• Hardware (VIC/GPU) preprocessing budget ≈480 detection frames/s per node (engineering heuristic).
• Shared memory budget: weights 11 MB + workspace 1024 MB + decode buffers 128 MB + preprocess 256 MB + OS 900 MB + 10% reserve.
• 8 × ~2.5 Mbit/s average (4 peak) H265 at 1080p@30 from the camera bitrate table; 1 GbE ingress assumed.

Reproduce and cite

Live permalink (carries the method and dataset versions; the page flags if either has changed since):
https://edgeaistack.ai/engines/camera-stream-capacity/?hw=jetson_orin_nano&cams=8&fps=30&mv=1.0&dv=2026-09-07

Static page: https://edgeaistack.ai/decode/answers/orin-nano-8x-1080p30-h265-with-detection/

Sources

• NVIDIA · DS-11105-001 v1.1 §1.9.1 Multi-Standard Video Decoder, Table 5 · verified 2026-09-07 · class A

  Jetson Orin Nano (8GB / 4GB) H265 decode: 1 × 4K60, 2 × 4K30, 5 × 1080p60, 11 × 1080p30 (690 MP/s)

• C · YOLOv8 Variant Benchmarks · class C

  yolov8s int8 tensorrt on jetson_orin_nano: 30 fps single stream

• NVIDIA · forums.developer.nvidia.com/t/380328 2026-08-17 · reply by NVIDIA staff (DaneLLL) · verified 2026-09-07 · class A

  NVDEC/NVENC maximum clock is the same in every nvpmodel power preset on Orin, so decode capacity does not change with the power mode.

• NVIDIA Developer Forums, thread 193819 · reply by NVIDIA staff (DaneLLL), 2021-11-03 · class A

Method: Camera Stream Capacity methodology v1.0. Module decode tables: Jetson Orin Nano (8GB / 4GB) H.265.