Frigate with a Coral TPU in Docker

Updated for Frigate 0.18 (Nov 2026): Frigate's hardware guidance has changed — the Coral is no longer the recommended detector for new installs (the EdgeTPU platform is a dead end; OpenVINO on a modern Intel iGPU or a Hailo accelerator is the current advice). If you already own a Coral it keeps working and this guide still documents it — but do not buy one on old advice. See our Frigate 0.18 guide for the full picture.

Frigate can detect people and cars on a CPU alone, but it will make a small NAS work hard for every camera you add. A Coral TPU — a ~$60 USB accelerator — does that detection almost for free, at a few watts. Getting it working inside Docker is a ten-minute job with exactly two traps. This guide walks the setup and, just as importantly, how to verify the Coral is really doing the work.

Published Nov 17, 2026

Why a Coral, in numbers

Object detection is the expensive part of an NVR. On CPU, each camera stream costs a meaningful slice of your processor, around the clock. A Coral Edge TPU runs the same detection model at roughly 10 ms per inference while sipping power, which is why a four-camera Frigate setup on a modest NAS goes from 'fans audible' to 'basically idle' with one plugged in. If your NAS has a recent Intel CPU, Quick Sync-accelerated detection is the alternative — but for most small boxes, the Coral is the cheapest big win in the whole camera build.

Camera sub-stream→Frigate detect→Coral TPU→Events + clips→Home Assistant

What you need

Step 1 — pass the Coral into the container

The Coral appears on the host as a Google USB device (it changes ID after its first inference — normal). Frigate's official approach is to pass through the whole USB bus, which survives that ID change:

services:
  frigate:
    image: ghcr.io/blakeblackshear/frigate:stable
    container_name: frigate
    restart: unless-stopped
    devices:
      - /dev/bus/usb:/dev/bus/usb   # Coral USB passthrough
    volumes:
      - ./frigate/config:/config
      - /path/to/media:/media/frigate
    ports:
      - "5000:5000"
      - "8554:8554"

The USB 3 trap: Coral USB sticks are famously flaky on some USB 3 ports and hubs — the device vanishes mid-detection and Frigate logs fill with errors. The community-standard fixes, in order: plug it into a USB 2.0 port, or put a short USB 2 extension cable between a USB 3 port and the Coral. It sounds like superstition; it is the single most common Coral fix there is.

Step 2 — tell Frigate to use it

In config.yml, declare the detector. One Coral handles several cameras comfortably; two detectors (a second Coral) is a real configuration if you scale past ~6 busy cameras:

detectors:
  coral:
    type: edgetpu
    device: usb

mqtt:
  host: 192.168.1.10   # your broker, if you use one

cameras:
  front_door:
    ffmpeg:
      inputs:
        - path: rtsp://user:[email protected]:554/stream1
          roles: [record]
        - path: rtsp://user:[email protected]:554/stream2
          roles: [detect]
    detect:
      width: 640
      height: 480
      fps: 5

Note the pattern that keeps CPU low regardless of detector: a low-resolution sub-stream does detection, the full-resolution stream only records. The Coral accelerates the detect role — it cannot rescue a config that feeds it 4K frames.

Step 3 — prove it is working

  1. Open Frigate's web UI → Settings (or the system stats page) and find the detector stats: you want inference speeds around 8–15 ms. Hundreds of milliseconds means you are on CPU.
  2. Check the container log for the line confirming the Edge TPU was found — and for the absence of repeated 'Was the Coral inserted?' style USB errors.
  3. Walk in front of a camera. Events should appear within a couple of seconds, tagged person, with the detector doing the work while host CPU stays near idle.

Troubleshooting

Coral alternatives worth knowing

The Coral is the default recommendation, not the only one. If your NAS has a recent Intel processor, Frigate can also run detection on the integrated GPU through OpenVINO — no extra hardware, and on an N100-class chip it is genuinely competitive with a Coral for a handful of cameras. The trade: it shares the chip with everything else (including Jellyfin transcoding, if you follow that guide too), where the Coral is dedicated silicon that never competes for resources. Hailo accelerators are the newer option in Frigate's world, faster still, but pricier and less battle-tested in home setups. And plain CPU detection remains fine for one or two quiet cameras — the Coral earns its money when camera count, scene busyness, or your electricity bill starts climbing. Start where you are; the config makes swapping detectors a two-line change later.

Planning or upgrading a Frigate setup? The Frigate config validator & storage planner checks a pasted config against the 0.18 rules, computes your shm_size and tiered storage, and generates a valid 0.18 skeleton.

Official documentation and references

Frequently asked questions

How many cameras can one Coral handle?

Comfortably 4–6 typical cameras detecting at 5 fps on sub-streams. Very busy scenes (a street-facing camera at higher fps) eat more inferences; Frigate queues gracefully, and a second Coral is the easy scale-up.

USB Coral or M.2 Coral — which should I buy?

Whichever your hardware accepts. USB is universal and movable between machines; M.2 (A+E key, common in mini PCs) is tidier inside a case and avoids the USB quirks entirely. Performance is the same chip either way.

Does the Coral replace a GPU for Frigate?

For detection, yes — that is exactly its job, and it does it more efficiently than most GPUs. It does not help with video transcoding or Frigate's other AI features (face/plate recognition use different models), which still run on CPU or GPU.

Related guides