From 9c2f69adc3a9c097262cfb6f81cccf739bdffa85 Mon Sep 17 00:00:00 2001 From: Erich Blume Date: Tue, 17 Feb 2026 18:26:16 -0800 Subject: [PATCH] Update docs to reflect actual ZMQ detector performance (~50-80ms) CoreML runs 468/662 model nodes with the rest on CPU, yielding ~50-80ms inference rather than the initially estimated ~15ms. Co-Authored-By: Claude Opus 4.6 --- docs/changelog.d/frigate-zmq-detector.infra.md | 2 +- docs/reference/services/frigate.md | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/docs/changelog.d/frigate-zmq-detector.infra.md b/docs/changelog.d/frigate-zmq-detector.infra.md index 91519c9..0035dc1 100644 --- a/docs/changelog.d/frigate-zmq-detector.infra.md +++ b/docs/changelog.d/frigate-zmq-detector.infra.md @@ -1 +1 @@ -Add Apple Silicon ZMQ detector for Frigate — inference moves from ONNX CPU (~117ms) to CoreML/Neural Engine (~15ms) +Add Apple Silicon ZMQ detector for Frigate — inference moves from ONNX CPU (~117ms) to CoreML/Neural Engine (~50-80ms) diff --git a/docs/reference/services/frigate.md b/docs/reference/services/frigate.md index 63c7008..492e0ba 100644 --- a/docs/reference/services/frigate.md +++ b/docs/reference/services/frigate.md @@ -49,7 +49,7 @@ Camera credentials are stored in 1Password and synced via [[external-secrets]] t ## Detection -Object detection uses the [apple-silicon-detector](https://github.com/frigate-nvr/apple-silicon-detector), which runs natively on [[indri]] as a LaunchAgent (`mcquack.eblume.frigate-detector`). It communicates with Frigate via ZMQ over TCP (`tcp://host.minikube.internal:5555`), leveraging CoreML and the M1 Neural Engine for ~15ms inference (down from ~117ms with ONNX CPU). +Object detection uses the [apple-silicon-detector](https://github.com/frigate-nvr/apple-silicon-detector) with a YOLOv9-t model (`yolo-generic`, 320x320), running natively on [[indri]] as a LaunchAgent (`mcquack.eblume.frigate-detector`). It communicates with Frigate via ZMQ over TCP (`tcp://host.minikube.internal:5555`), using CoreML with partial Neural Engine acceleration (~50-80ms inference, down from ~117ms with in-pod ONNX CPU). A `driveway_entrance` zone is configured for alert filtering — only detections in this zone trigger review alerts.