Frigate NVR: What the Coral and the GPU Each Do
Purpose
Frigate NVR is a local, private video surveillance system with real-time AI object detection. It ingests camera streams, runs a model to detect people, vehicles, and animals, and records only what matters. Everything stays local with no cloud dependency or external account required. There’s no subscription fee behind it either.
Frigate moved in November 2026. It used to run on a dedicated NVIDIA GPU Docker host, which has been retired. It now runs on a Docker host with an Intel Iris Xe iGPU. Detection did not change: the Coral USB Edge TPU does it. An earlier version of this post said detection ran with TensorRT on the NVIDIA GPU. That was wrong. The old GPU only ever decoded video, and the iGPU now does the decoding for one of the two cameras.
Deployment
Frigate runs as a single Docker container on the iGPU Docker host, an LXC on the second Proxmox node.
| Item | Value |
|---|---|
| Host | iGPU Docker host (Intel Iris Xe) |
| Image | ghcr.io/blakeblackshear/frigate:stable |
| GPU | /dev/dri passed into the container, preset-vaapi for decode |
| Coral | USB Edge TPU, passed into the LXC and the container |
| Network | the proxy network |
| Deployment | Docker Compose, managed locally |
| Logs | gelf log driver, shipped to Graylog |
The Coral sits in a USB 3 port on the Proxmox node that hosts the Docker host, and the container sees it through the passthrough. The move took three evenings for three apps, in the order Jellyfin, Immich, Frigate (brief).
flowchart LR
cam["PoE Cameras"] -->|"RTSP"| go2rtc["go2rtc restreamer"]
go2rtc --> living["Living_room<br/>VA-API decode on the iGPU"]
go2rtc --> garage["Garage_door substream 640x352<br/>CPU decode"]
living --> coral["Coral USB Edge TPU<br/>object detection"]
garage --> coral
coral -->|"detected objects"| recording["Recording Engine"]
recording --> events["Events"]
events -->|"MQTT"| mqtt["Mosquitto"]
mqtt --> ha["Home Assistant"]
ha --> notify["Notifications"]
recording --> clips["Recordings and clips"]
Camera streams arrive via RTSP through the built-in go2rtc restreamer, which handles protocol conversion and reconnections. Living_room is decoded on the iGPU with VA-API, and Garage_door is decoded on the CPU. The frames from both go to the Coral for detection.
Detection
The detector is a Coral USB Edge TPU (type: edgetpu). No custom model is configured, so Frigate’s default Coral model runs. The iGPU plays no part in detection.
Frigate detects people, vehicles, animals, and other categories depending on the model and configuration. Detected objects trigger events that feed into recording retention and downstream automations.
Decode: iGPU for One Camera, CPU for the Other
Living_room uses VA-API on the iGPU. Garage_door decodes on the CPU. VA-API decode of that camera kept crashing, so the CPU path replaced it. It is cheap there: Garage_door detects on its 640×352 substream and records the 1080p main stream by copy. (brief).
Face Recognition and Semantic Search
Face recognition is set to the small model size. On the CPU, large took about 1015 ms per face and small about 751 ms, and small still recognised the person it was tested on. The iGPU’s compute stayed at 0% during the test, so face recognition runs on the CPU here. Semantic search is set to large, and there are no timings for it yet (brief).
Recording
Recording is event-only: the system does not write continuous video streams. When the detector identifies something worth keeping, Frigate saves a clip with configurable pre- and post-event padding. This keeps storage manageable while preserving useful footage: you get the thing that happened without hours of empty hallway.
Home Assistant Integration
Frigate is integrated with Home Assistant via MQTT events published to Mosquitto. When Frigate detects an object, such as a person entering a frame or a vehicle in the driveway, it publishes the event as an MQTT message. Home Assistant’s built-in Frigate integration subscribes to those events and exposes them as triggers for automations.
A typical flow: the Coral sees a person, Frigate publishes an MQTT event, a Home Assistant automation fires, and a notification with a snapshot arrives on a phone. The same pipeline drives state changes and conditional logic without exposing cameras or credentials beyond the homelab network.
See Home Assistant and Mosquitto MQTT for details on those pieces.
Access
The web login port is no longer published. The API returns 401 without a login, while Home Assistant stays authenticated.
The garage camera sits on an isolated VLAN with no internet. Only Frigate’s host and internal networks can reach it (brief).
Notes
The image is now the plain stable tag. The old stable-tensorrt variant is gone, since detection never used TensorRT.
If you are weighing an iGPU against a discrete GPU for Frigate: the Coral does the detection either way, and a small iGPU decodes one camera comfortably. The same host also runs Immich and Jellyfin now.
The Docker Compose configuration mounts a dedicated directory for config and a separate one for recordings and snapshots. The recordings path should live on fast storage, since write speed during a detection event determines whether you keep the clip or lose it to a full buffer.