Cameras

Frigate hardware planner

See how many cameras a detector (Coral, Hailo, Intel, Nvidia) can handle at your detection fps.

Use this to right-size a Frigate detector: it applies Frigate's own cameras ≈ (1000 ÷ inference_ms) ÷ detect_fps capacity model and flags when you're over capacity.

Low priority

Hailo-8L can handle this — up to ~13 cameras at 5 fps

At ~11 ms inference, Hailo-8L sustains about 68 detections/sec (75% derated). 6 cameras × 5 fps = 30 detections/sec, so this is within capacity.

Detector capacity: 68 det/s
Your load: 30 det/s
Max cameras: 13
Verdict: Within capacity

First checks

Detector: Hailo-8L (~11 ms inference).
Load: 6 cameras × 5 fps detection = 30 detections/sec.
Inference numbers are model-specific (ssd-mobilenet / yolov6n class) — measure your own in Frigate's stats for accuracy.

Step-by-step troubleshooting

1
Confirm your detector and inference time
Check Frigate's stats for the actual inference speed (ms) of your detector + model.
Expected: A measured inference time instead of a table estimate.
Next: Enter it in the override field and re-run.
2
Plan decode separately from detection
Ensure the host has hardware video decode (iGPU/GPU) for all camera streams.
Expected: Decode and detection both have headroom.
Next: If detection is over capacity, lower fps or upgrade the detector.

What your answers suggest

Capacity = (1000 ÷ 11 ms) × 0.75 ≈ 68 detections/sec.
Max cameras ≈ capacity ÷ detect fps = 68 ÷ 5 ≈ 13.

Likely cause area

Detection capacity and video DECODE are separate — 4K decode needs an iGPU/GPU (QuickSync/VAAPI/Nvidia), independent of the detector.
Run detection on a ~5 fps substream; detecting on the full main stream wastes capacity.

Safe actions

You have detector headroom — keep detection on the substream at ~5 fps.
Make sure the host can also DECODE all the camera streams (separate from detection).

When to stop

Don't run near the theoretical ceiling — real frame-skipping starts before 100% (this figure is already derated to 75%).

Assumptions

Inference times are representative 2026 figures and are model- and resolution-specific.
Derated to 75% of the theoretical 1000/inference ceiling because real frame-skip starts earlier.
Video decode load is separate and not included in this detection estimate.

What should I check first?

Pick your detector and check its inference time in Frigate's stats (ms) if you already run it.
Set a realistic detection fps per camera (Frigate default is 5; 10 for fast objects).
Confirm the host can also DECODE all camera streams — that's separate from detection.

What is likely wrong?

Choosing a Coral for a new build — Frigate's docs now list it as no longer recommended.
Detecting on the high-res main stream instead of a ~5 fps substream.
Forgetting that 4K video decode needs an iGPU/GPU independent of the detector.

What is safe to try?

Detect on the substream at ~5 fps to stretch detector capacity.
For new builds, prefer a Hailo-8L or an Intel iGPU/Arc over Coral.
Lower detection fps or split cameras across two detectors if you're over capacity.

When should I stop?

Stop before running near the theoretical ceiling — real frame-skipping starts before 100%.
Stop assuming detection capacity covers decode; size 4K decode (QuickSync/VAAPI/Nvidia) separately.

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 ≈ (1000 ÷ inference_ms) ÷ detect_fps 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.

FAQ

Is the Coral still a good choice in 2026?

Frigate's hardware docs now state the Coral is no longer recommended for new installations — prefer a Hailo-8L or an Intel iGPU/Arc, which are widely available and support modern models. The Coral still works and is very power-efficient, but it's limited to older model types and the USB version lacks the throttling protections of the M.2/PCIe versions.

Why derate the capacity to 75%?

The 1000 ÷ inference-ms figure is a theoretical ceiling. In practice, frame-skipping and scheduling overhead start before you hit it, so the planner uses ~75% of theoretical capacity. Measure your detector's real inference time in Frigate's stats and enter it for the most accurate result.