[Bug]: [0.17.0][HA Add-on][OpenVINO Intel iGPU] face_recognition (model_size: large, device: GPU.0) crashes embeddings_manager with CL_INVALID_VALUE

[NapalmCandy]

Checklist

• I have updated to the latest available Frigate version.
• I have cleared the cache of my browser.
• I have tried a different browser to see if it is related to my browser.
• I have tried reproducing the issue in incognito mode to rule out problems with any third party extensions or plugins I have installed.
• I have asked the AI at https://docs.frigate.video about my issue.

Describe the problem you are having

Frigate starts and object detection works correctly on OpenVINO Intel iGPU, but enabling face recognition with model_size: large and device: GPU.0 consistently crashes frigate.embeddings_manager during OpenVINO model compilation.

The crash appears specific to the face embedding path (ArcFace/OpenVINO GPU). Main camera streams and go2rtc continue running.

Steps to reproduce

1. Run Frigate 0.17.0 (Home Assistant Add-on) with OpenVINO detector on Intel iGPU.
2. Use a working object model (yolo26s_384_FP32_E2EOFF_ONNX.onnx) and confirm normal startup.
3. Enable global face recognition with:
   • enabled: true
   • model_size: large
   • device: GPU.0
4. Restart Frigate.
5. Observe frigate.embeddings_manager crash with OpenVINO GPU error clCreateSubBuffer ... CL_INVALID_VALUE.

Version

0.17.0-f0d69f7

In which browser(s) are you experiencing the issue with?

Google Chrome 145.0.0.0 (Windows 10 x64)

Frigate config file

mqtt:
  host: 192.168.1.180
  port: 1883
  user: REDACTED
  password: REDACTED

detectors:
  ov:
    type: openvino
    device: GPU

model:
  path: /config/model_cache/yolo26s_384_FP32_E2EOFF_ONNX.onnx
  width: 384
  height: 384
  input_tensor: nchw
  input_dtype: float
  input_pixel_format: rgb
  model_type: yolo-generic
  labelmap:
    0: person
    1: bicycle
    2: car
    3: motorcycle
    5: bus
    7: truck
    15: cat
    16: dog

ffmpeg:
  hwaccel_args: preset-vaapi
  input_args: preset-rtsp-generic

face_recognition:
  enabled: true
  model_size: large
  device: GPU.0
  min_area: 500
  detection_threshold: 0.7

cameras:
  interfono_frigate:
    ffmpeg:
      inputs:
        - path: rtsp://127.0.0.1:8554/interfono_frigate
          roles: [record, detect]
    detect:
      width: 1280
      height: 960
      fps: 4
    face_recognition:
      enabled: true
      min_area: 10000

version: 0.17-0

docker-compose file or Docker CLI command

N/A - Home Assistant Add-on installation (Supervisor managed container).
Add-on: Frigate NVR
Install method: Home Assistant Add-on

Relevant Frigate log output

[2026-03-03 21:17:45] frigate.app INFO : Starting Frigate (0.17.0-f0d69f7)
[2026-03-03 21:17:45] frigate.app INFO : Embedding process started: 979

Process frigate.embeddings_manager:
Traceback (most recent call last):
  File "/opt/frigate/frigate/embeddings/__init__.py", line 49, in run
    maintainer = EmbeddingMaintainer(...)
  File "/opt/frigate/frigate/embeddings/maintainer.py", line 163, in __init__
    FaceRealTimeProcessor(...)
  File "/opt/frigate/frigate/data_processing/real_time/face.py", line 94, in __init__
    self.recognizer = ArcFaceRecognizer(self.config)
  File "/opt/frigate/frigate/embeddings/onnx/face_embedding.py", line 155, in _load_model_and_utils
    self.runner = get_optimized_runner(...)
  File "/opt/frigate/frigate/detectors/detection_runners.py", line 316, in __init__
    self.compiled_model = self.ov_core.compile_model(...)
RuntimeError: Exception from src/inference/src/cpp/core.cpp:126:
Exception from src/inference/src/dev/plugin.cpp:58:
Exception from src/plugins/intel_gpu/src/runtime/ocl/ocl_engine.cpp:205:
[GPU] clCreateSubBuffer, error code: -30 CL_INVALID_VALUE

Relevant go2rtc log output

21:17:39.070 INF go2rtc version 1.5.0 linux/amd64
21:17:39.071 INF [api] listen addr=:1984
21:17:39.072 INF [rtsp] listen addr=:8554
21:17:39.075 INF [srtp] listen addr=:8443
21:17:39.076 INF [webrtc] listen addr=:8555/tcp

No go2rtc errors around the crash window; the failure is in frigate.embeddings_manager.

Operating system

Home Assistant OS

Install method

Home Assistant Add-on

Network connection

Wired

Camera make and model

Dahua-based RTSP streams via /cam/realmonitor (NVR channels), plus two Dahua IP cameras (exact model numbers unknown).

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

• Regression context: in Frigate 0.16, face recognition worked for this setup.
• In 0.17, object detection works with OpenVINO iGPU and YOLO26s model.
• Crash occurs when face_recognition uses OpenVINO GPU path with model_size: large.
• Same behavior observed with device: GPU and device: GPU.0.

[NickM-27]

We haven't had any other reports of this, it could be related to the Intel drivers in the HA OS builds, have you tried just with GPU as the value?

[NapalmCandy]

Yes, I already wrote on the final lines of the report post

"Same behavior observed with device: GPU and device: GPU.0."

I also tried with several AI (ChatGPT, Claude and Gemini) to solve the problem, but I think there is nothing I can't do.

The problem started when I updated to 0.17, my old yolo11s model doesn't load (so I generated a new one, the XML/bin do not work, so I use ONNX) and also had the problem with the Face Recognition with large model.

[NickM-27]

I see, we may need to wait for other reports then to understand what the difference is.

[tete1030]

I actually have reported this earlier, there is a bug in openvino 2025.3
#22059

you can fix it by either upgrading to openvino 2025.4 or patch this PR: openvinotoolkit/openvino#31180

[Wolfgang6200]

I am running Frigate+ 0.17 with the yolov9s‑relu6 (edgetpu) detection model. Detection works very well, but the embedding stage always runs on the CPU. Even with 7 Coral PCIe TPUs installed, Frigate+ does not offload embeddings to any TPU. This leads to severe performance issues:

Embedding inference time: ~900 ms per object

Embedding throughput: 0 inferences/s

CPU load: ~99%

Re‑ID and tracking become unstable

All 7 TPUs show detection load only, no embedding load

GPU remains unused for embeddings (expected due to TFLite pipeline)

After analyzing the behavior, it seems that Frigate+ currently does not provide a TPU‑compatible ReID/embedding model. YOLOv9s‑relu6 produces bounding boxes but no embeddings, so Frigate+ falls back to CPU‑based embedding.

This behavior is fully reproducible:

Detection model: yolov9s-relu6-edgetpu

No TPU‑compatible embedding model available

Frigate+ attempts to generate embeddings

Fallback to CPU

CPU load spikes to 99%

Embedding latency ~900 ms

Expected behavior:

A TPU‑accelerated ReID/embedding model should be available in Frigate+

Embeddings should run on TPU (5–10 ms instead of ~900 ms)

CPU load should drop significantly

Tracking and Re‑ID should become stable

Multi‑TPU systems (e.g., 7× Coral PCIe) should be fully utilized

Questions / Suggestions:

Is a TPU‑compatible ReID/embedding model planned for Frigate+?

Could embedding results (e.g., license plate text) be exposed in the UI, not only via MQTT?

Is GPU‑accelerated embedding (TensorRT‑based ReID models) something that could be considered in the future?

[hawkeye217]

I think there is some confusion here. Frigate does not employ ReID or generate embeddings for tracked objects, and the object detection model that runs on the Coral is for object detection, it is not an embedding-generating model.

A TPU is very different from a GPU - Frigate does not run any of its embedding models for its other enrichments on a TPU, as documented here.

If you have other questions or need additional assistance, feel free to open your own support discussion as your post is unrelated to the original issue in this thread.

[Legendhidde]

I am having this same issue on 0.17 on an Intel Ultra 235H with Reolink Cameras, I cannot get enrichments to work properly for some reason. LPR never runs and face recognition has worked but only sometimes

[Legendhidde]

I’d love to give more information. I’m running on Bazzite with docker on bare metal with 0.17. I get the same issues. For now I’ve switched to small model to avoid these issues. Large model gave me the same errors but sometimes, when it did work. It would give me completely random images as faces that it was trying to detect.

[tete1030]

Reported this earlier, there is a bug in openvino 2025.3
#22059

you can fix it by either upgrading to openvino 2025.4 or patch this PR: openvinotoolkit/openvino#31180

[noldevin]

I'm having the same issue, even tried a custom docker compose with the updated openvino. Soft crash when i try to upload that first photo. No CPU usage, needs a reboot to come back.
I am new to frigate and started on 0.17, so can't say it ever worked before.
I have frigate running in its own proxmox LXC with 14500T igpu passed through and 8gb of RAM assigned.