Cameras

Coral vs Hailo vs Intel NPU for Frigate

Choose a Frigate object detector in 2026 — why Coral is no longer recommended for new installs, and how Hailo and the Intel iGPU/NPU compare on inference time and camera count.

Problem summary

A Frigate detector is sized by inference time, not brand hype: cameras ≈ (1000 ÷ inference_ms) ÷ detect_fps. Frigate now says the Coral is no longer recommended for new installs and points to Hailo-8/8L or an Intel iGPU/NPU (OpenVINO) — and detection load is separate from video decode, which the host still has to handle.

Operator snapshotEvidence first

First proof

Count your camera-streams and target detect fps (usually ~5).

Screen to open

Frigate UI > System metrics (inference speed) and the recommended-hardware docs

Expected signal

You know how many detect streams run and at what fps.

Stop boundary

For a brand-new build, choose a currently-recommended detector instead.

Layer path

1A Frigate detector is chosen by inference time, not brand: Frigate's own capacity model is cameras ≈ (1000 ÷ inference_ms) ÷ detect_fps, so a faster detector simply covers more camera-streams.

2The 2026 shift: Frigate's recommended-hardware docs now state the Coral is 'no longer recommended for new Frigate installations' (except very low-power or otherwise-incapable hardware) and point to Hailo-8/8L, the Intel iGPU/NPU via OpenVINO, or a GPU.

3Object detection (the accelerator's job) is separate from video decode (the host CPU/iGPU/GPU's job) — a fast detector on a host that can't decode your streams still drops frames.

4Approximate published inference: Coral ssd-mobilenet ~10 ms; Hailo-8L ssd-mobilenet ~10 ms / yolov6n ~11 ms; Hailo-8 ~6–7 ms; Intel N100 iGPU (OpenVINO) MobileNetV2 ~15 ms; Intel NPU MobileNetV2 ~6 ms — re-check the live docs, these move with model/version.

Runbook

Step-by-step runbook

Start here. Do each check in order, compare it to the expected result, and stop when the evidence explains the failure or the safe stop point applies.

Inventory cameras and host

Check: List camera-streams, target detect fps, and the host's iGPU/NPU/GPU.

Expected result: You know the detection load and what decode capability exists.

If not: Without this you can't size either detection or decode.

Try the built-in detector first

Check: On a modern Intel host, enable the OpenVINO detector (iGPU/NPU) before buying hardware.

Expected result: Detection runs within budget for your camera count.

If not: If inference is too slow for the count, plan an add-in detector.

Add Hailo if you outgrow it

Check: Fit a Hailo-8L (budget) or Hailo-8 (faster) M.2 module and select it as the detector.

Expected result: Inference time drops and the camera count rises.

If not: Don't reflexively buy a new Coral — Frigate no longer recommends it for new installs.

Provision decode separately

Check: Ensure the host iGPU/GPU decodes all streams via the right hwaccel preset.

Expected result: Frames are stable; CPU isn't pegged on decode.

If not: A fast detector can't fix a decode bottleneck.

Verify and right-size fps

Check: Confirm inference and decode metrics, then keep detect on a low-res substream at ~5 fps.

Expected result: Detection and decode both sit comfortably under capacity.

If not: Raise fps only if you have headroom; otherwise add a detector.

Decision tree

If: New build, a handful of cameras, modern Intel host.

Then: OpenVINO on the Intel iGPU/NPU likely suffices.

Action: Start with the built-in OpenVINO detector; no add-in accelerator needed yet.

If: Many cameras / want the most efficient dedicated detector.

Then: Hailo-8L (budget) or Hailo-8 (faster) is the current recommendation.

Action: Fit a Hailo M.2 module; size camera count from its inference time.

If: You already own a Coral and it's working.

Then: Keep it — it still works and is power-efficient.

Action: No need to replace a working Coral; just don't buy one new as the default.

Safe stop: For a brand-new build, choose a currently-recommended detector instead.

If: Frames drop even with a fast detector.

Then: Decode (not detection) is the bottleneck.

Action: Add/enable hardware decode (iGPU); confirm detect runs on the substream.

Safe stop: If the host can't decode the streams at all, no detector choice fixes it.

Evidence

Evidence table

Symptom	Evidence to collect	Likely layer	Next action
Deciding what to buy for a new Frigate build.	Camera count, detect fps, host iGPU/NPU capability.	Detector sizing	OpenVINO first; Hailo if you outgrow it; not a new Coral by default.
Detection inference high / skipped frames.	Reported inference time vs the model you're running.	Detector throughput	Use the rated model or a faster detector; derate the camera count ~25%.
CPU pegged but inference looks fine.	Whether hardware decode (QuickSync/iGPU) is enabled.	Video decode (host, not detector)	Enable iGPU decode; keep detect on the low-res substream.
Core Ultra host underperforming.	Whether NPU and iGPU are both being used.	Engine assignment	NPU for detect, iGPU for decode/enrichments.

Reference

Commands and settings paths

Read Frigate's detector inference

Frigate UI > System metrics (inference speed) and the recommended-hardware docs

Where: On the running Frigate instance and docs.frigate.video/frigate/hardware.

Expected: You see actual inference ms for your detector and the per-model reference figures.

Failure means: An inference time far above the rated figure means the wrong model or an overloaded detector.

Safe next step: Switch to the rated model / a faster detector and re-check; size with the planner.

Confirm hardware decode is active

Frigate config: ffmpeg hwaccel (e.g. preset-vaapi / preset-intel-qsv); check GPU usage

Where: In the Frigate config and host GPU stats (intel_gpu_top, nvidia-smi).

Expected: Decode is offloaded to the iGPU/GPU, not the CPU.

Failure means: CPU-bound decode shows as high CPU and dropped frames regardless of detector.

Safe next step: Enable the correct hwaccel preset for your hardware.

Size camera capacity

cameras ≈ (1000 ÷ inference_ms) ÷ detect_fps, then derate ~25%

Where: On paper / in the Frigate hardware planner tool.

Expected: The detector covers your camera count with headroom after derating.

Failure means: If the number is below your camera count, you're over-subscribed.

Safe next step: Pick a faster detector, lower detect fps, or add a second detector.

Hardware boundary

Hardware and platform boundary

Change only when

Buy a dedicated detector (Hailo) only after the host's iGPU/NPU (OpenVINO) can't keep up with your camera count — not by default.

Evidence that matters

Inference time for the model you'll run, host decode capability (iGPU/QuickSync), the right M.2/PCIe slot, and Frigate driver support.

Evidence that does not matter

Marketing TOPS numbers and a 'new Coral' reflex — size on Frigate's real per-model inference figures and the capacity formula.

Avoid

Ignoring decode (a fast detector on a host that can't decode 4K/H.265 still drops frames) or quoting inference for the wrong model.

Related tool/calculator

Use the linked calculator or tool to turn this runbook into numbers for your exact setup.

Frigate hardware planner

Last reviewed

2026-06-03 · Reviewed by HomeTechOps. Built from June-2026 research re-verified against Frigate's recommended-hardware and object-detectors docs (the current 'Coral no longer recommended for new installs' guidance, Hailo-8/8L and OpenVINO iGPU/NPU figures, and the cameras ≈ (1000 ÷ inference_ms) ÷ detect_fps model); separates detection from decode and frames the choice as sizing, not brand loyalty. Inference numbers move with model/version — cite the live docs.

Sources/assumptions

Assumes a self-hosted Frigate install doing continuous object detection on several cameras.
Inference times are the figures Frigate publishes for specific models — they change with model and version, so re-check the current hardware docs.
Detector choice is sized on detection load only; video decoding (especially 4K/H.265) is a separate demand on the host iGPU/GPU.

Source-backed checks

HomeTechOps turns official docs and conservative safety rules into a shorter runbook. These links are the source trail for the page direction.

Frigate docs: Recommended hardwareUsed for detector inference times, the cameras-capacity model, and the current 'Coral no longer recommended for new installs' guidance.Frigate docs: Camera setup (detect vs record streams)Used for the detect-on-substream / record-on-main pattern that drives bandwidth and detector load.FS.com: Understanding PoE standards and wattageUsed for the 802.3af/at/bt source-vs-delivered power figures and the switch total-PoE-budget constraint.Frigate Docs: Object DetectorsUsed for Frigate detector choices (Coral, Hailo, OpenVINO, TensorRT) and the rule that NVR object detection needs a dedicated accelerator rather than CPU inference.Hailo: Hailo-8L M.2 AI acceleration moduleUsed for the Hailo-8L as the current budget M.2 NVR-inference accelerator after the Coral USB line was deprecated.Google Coral — USB / M.2 Edge TPUUsed for the Coral Edge TPU (ssd-mobilenet ~10 ms inference) referenced in the detector decision; pair with Frigate's 'no longer recommended for new installs' guidance.Intel OpenVINO toolkitUsed for OpenVINO running object detection on Intel CPU/iGPU/NPU (the Frigate OpenVINO detector path); on Core Ultra, NPU for detection and iGPU for enrichments.

Diagnostic rule

1. Check the simplest signal first.
2. Change one thing at a time.
3. Prove the stream/reachability before deep config.
4. Stop when the evidence says the risk moved.

FAQ

Should I still buy a Google Coral for Frigate in 2026?

Probably not for a new build. Frigate's recommended-hardware docs now state the Coral is 'no longer recommended for new Frigate installations' except in very low-power or otherwise-incapable hardware, and point to the alternatives (Hailo-8/8L, Intel iGPU/NPU via OpenVINO, or a GPU). The Coral still works and is power-efficient, but it's no longer the default recommendation. Size the real need with the Frigate hardware planner.

How many cameras can one detector handle?

Use Frigate's model: cameras ≈ (1000 ÷ inference_ms) ÷ detect_fps, then derate ~25% for real-world overhead. At 5 detect fps, a ~10 ms detector (Coral ssd-mobilenet, Hailo-8L) gives roughly (1000/10)/5 = 20 camera-streams of headroom before derating; a faster ~6–7 ms detector (Hailo-8, Intel NPU on MobileNet) gives more. Detect runs on the low-res substream, so this is about detection, not recording.

Does the detector also handle video decoding?

No — and this trips people up. The accelerator does object inference; decoding the camera streams (turning H.264/H.265 into frames) is a separate job for the host CPU or iGPU/GPU. A fast detector on a host that can't decode your streams will still drop frames. On Intel Core Ultra systems Frigate suggests using the NPU for detection and the iGPU for decode/enrichments.