Best Hardware for Frigate NVR (2026): Coral vs Hailo-8L vs Jetson vs Intel

What Changed in 2026

For years the standard Frigate answer was "buy a Coral." That changed. Current Frigate hardware guidance now states the Coral is no longer recommended for new installations, except on very low-power or otherwise-incapable hardware, and points new builders toward Hailo-8/8L, the Intel iGPU/NPU via OpenVINO, or a GPU. The reason isn't that Coral broke — it's that Coral has stood still while everything around it moved. It runs a single model at a time, its labelmap is a subset of just 17 COCO classes on the default model, and Google has effectively walked away from the product.

Frigate 0.17 also reshaped the detector landscape: the base image moved to Debian Bookworm, the JetPack 4 and 5 Jetson images were removed in favor of a community-contributed JetPack 6 build, the Hailo driver and firmware were bundled into the standard image (4.21.0 series), a quantized YOLOv9 path was added for Coral to claw back some accuracy, and an Apple Silicon NPU detector plus community MemryX and DeGirum detectors landed. The practical upshot: in 2026 you have real choices, and the cheapest correct one is usually already in your host.

How Frigate Sizes a Detector

Frigate is built around offloading detection to a dedicated accelerator so it can run inference an order of magnitude faster than CPU. The sizing math is simple and worth internalizing before you spend money:

max detections/sec ≈ 1000 ÷ inference_ms   and   cameras ≈ (1000 ÷ inference_ms) ÷ detect_fps

At a 10 ms inference time, one detector tops out near 100 detections per second. At a typical 5 fps detect rate on a low-resolution substream, that's headroom for roughly 20 camera-streams before you start queuing. Push detect fps higher, or run a heavier model with a longer inference time, and the camera count drops proportionally.

Two caveats that trip people up. First, detection is not decode. The accelerator only runs the object-detection model; decoding the H.264/H.265 streams is a separate load on the host CPU, integrated GPU, or QuickSync engine. A fast detector cannot rescue a host that can't decode your streams. Second, published inference times move with model, resolution, and Frigate version — treat the numbers below as planning approximations and re-check the live docs for your exact model.

Detector Comparison Table

Approximate published inference times for common Frigate detector + model pairings as of mid-2026. Re-verify against Frigate's hardware and object-detector docs before buying — these shift with model and release.

Inference figures are from Frigate docs and community measurements on representative hosts (e.g. Intel 13th-gen and N100 for the Intel rows). Numbers vary with model width (320 vs 640), Frigate version, and host. Detect throughput, not raw TOPS, is what sizes your build.

Coral EdgeTPU

The Coral is still fully supported in current Frigate, in USB, M.2, and Mini-PCIe formats, and its ~8–10 ms inference on the default SSD-MobileNet model remains genuinely fast for a 4 TOPS part. Frigate 0.17 even added a quantized YOLOv9 path that improves accuracy over the old mobiledet default. So why the demotion?

Three structural limits. It runs one model at a time, which rules it out for the enrichment features (face recognition, semantic search) that newer Frigate builds lean on. Its default labelmap is a subset of COCO — 17 classes — so it sees fewer object types than a YOLO model on a more capable detector. And the platform is, by every external sign, end-of-life: the product site has gone stale, the Hailo-8 is the only comparably fast alternative, and you're buying a 2019-era accelerator in 2026 from a company that has effectively exited. If you already own one, keep using it. If you're spending fresh money, the value case has moved on.

Hailo-8 and Hailo-8L

Hailo is the detector Frigate now points new builders toward, and the integration is mature: the standard 0.17 image ships the HailoRT driver, the detector auto-detects whether you have an 8 or an 8L and picks an appropriate default YOLO model, and the kernel module is DKMS-packaged so kernel upgrades don't silently break it. You can drop a pre-compiled .hef from Hailo's model zoo into your config and skip compilation entirely.

Hailo-8L (13 TOPS, ~$70 as the Raspberry Pi AI Kit) is the budget pick. YOLOv6n runs around 10–13 ms, which comfortably covers several cameras at typical detect rates while drawing about as much power as a phone charger. One community caveat worth knowing: some users report weaker night-time detection on the default YOLOv6n versus Coral's MobileDet under IR lighting, and that swapping to a newer model revision can shave a few milliseconds — worth testing on your own footage.

Hailo-8 (26 TOPS, M.2, ~$200) roughly halves inference to 6–7 ms on YOLOv8n, handling high stream counts and leaving headroom for concurrent pipelines on an x86 host.

Rule of thumb: on a Raspberry Pi 5, the Hailo-8L AI Kit is the cleanest first-party path. On an x86 mini-PC with a free M.2 slot, the Hailo-8 is the call if your host iGPU/NPU can't already keep up.

Intel iGPU and NPU (OpenVINO)

The most overlooked Frigate detector in 2026 is the one you already paid for. If your host is a modern Intel mini-PC — an N100 box, a NUC-style unit, a Beelink — its integrated GPU (and, on newer Core chips, a dedicated NPU) can run detection through OpenVINO with zero added hardware. MobileNetV2 lands around 15 ms on an N100 iGPU and around 6 ms on an Intel NPU, which covers a typical home deployment without buying an accelerator at all.

Frigate's own guidance is to buy a dedicated detector only after the host's iGPU/NPU can't keep up — not by default. The bonus: an Intel host with QuickSync also handles your video decode, so detection and decode live on one efficient part. For a new build with a handful of cameras and a modern Intel host, OpenVINO on the iGPU or NPU is very often the right and cheapest answer.

Jetson (TensorRT / ONNX)

Jetson is supported via the TensorRT or ONNX detectors on the JetPack 6 image (-tensorrt-jp6). Inference is typically 20–40 ms depending on the YOLO model, the Jetson platform, and the nvpmodel power mode. That's slower per-frame than a Hailo, and the DLA path is more power-efficient but not faster than the GPU. So for pure detection, a Jetson is usually overkill and over-budget next to a Hailo or an Intel NPU.

Where Jetson earns its place is everything around detection. The JP6 image auto-detects the Jetson for GPU-accelerated enrichments — face recognition and semantic search — which a Coral can't do at all. And if you're already running a VLM or LLM on the box (scene description, "is the package still on the porch," natural-language search over clips), the Jetson is doing double duty and the detection cost is marginal. Buy a Jetson for Frigate when the Jetson is also doing other AI work, not as a dedicated detector.

RK3588 (RKNN)

Rockchip RK3588 boards run object detection on their built-in 6 TOPS NPU via the RKNN detector, with YOLO-NAS s landing around 25–30 ms using all three NPU cores, and Frigate supports hardware video decode across Rockchip boards. The appeal is integration: one cheap (~$120–$180) ARM board provides the host, the decode engine, and the detector in a single low-power package. The trade-off is a smaller software ecosystem and more setup friction than the Intel + Hailo path. It's a strong all-in-one for a fixed, modest camera count where simplicity and power draw matter more than raw flexibility.

Decision Framework

Choose Intel iGPU/NPU (OpenVINO) if

• You're building on a modern Intel mini-PC and have a handful of cameras
• You want zero added hardware and a single part doing decode + detect
• Your detect load fits within the host's spare iGPU/NPU headroom

Choose Hailo-8L if

• You're on a Raspberry Pi 5 (the AI Kit is the first-party path) or need a cheap M.2 upgrade
• The host iGPU/NPU has run out of detection headroom
• Budget is ~$70 and a handful of cameras is your scale

Choose Hailo-8 if

• You're on x86 with a free M.2 slot and a high stream count
• You want sub-7 ms inference and room for concurrent pipelines
• You may add GPU-style enrichments later and want detection off the critical path

Choose Jetson if

• You want GPU-accelerated face recognition or semantic search alongside detection
• You're co-locating a VLM/LLM on the same device
• You already run a Jetson and detection is incidental

Keep your Coral if

• You already own one and it covers your camera count
• You're on very low-power or otherwise-incapable hardware where it's the only fit
• You don't need enrichments or a broad labelmap — but don't buy a new one for a new build

Frequently Asked Questions