[Detector Support]: Frigate+ Openvino Suddenly Stops Working Despite Working For Weeks #20631
Replies: 1 comment 18 replies
-
|
Did something else in your Proxmox setup change? Virtualization always tends to compound issues like this, which is why we always recommend installing in Docker on bare metal. |
Beta Was this translation helpful? Give feedback.
-
I just couldn't figure it out anymore. The only promising is when I use xe module. It does load 20 cameras with openvino, but eventually, detector gets stuck, then restarts. And oh by the way, it seems I have improved the ffmpeg restart floods with my dahua cameras, I dialed down the fps from the dvr to 5 fps. Im going to observe if this solves the ffmpeg crashes permanently. This does not happen with preset-nvidia though, even without touching dvr settings. |
Beta Was this translation helpful? Give feedback.
-
|
Update: Alright, that was a long journey. I went ahead installing Ubuntu 24.04 because that's where I started. And retraced back on how I started with frigate. Installing cameras, 1 by 1 without adding other configs first such as recordings. I started with the default openvino model which is indicated in the docs: Now, 1 by 1 cameras I added, detection enabled. Seems to run pretty well until 20 cameras. Now as soon as I change to YOLONAS or YOLOV9S from the Frigate+ models, it crashes. Now, to diagnose properly, from cold boot for good measure. I went back adding from start again, but this time, using YOLONAS (also YOLOV9S tested, also the generic YOLONAS and YOLOV9s, all in 320x320). 1st camera, runs fine, but as soon as I add 2nd camera, it crashes. 2 cameras with 704x576 resolution are not that much for YOLONAS or YOLOV9S I guess? I mean, Dell Optiplex 3070 Micro could run same 20 cameras with either model without crashing. I remember now when did all this happen. When 2025.3 model came out from the Frigate+, I trained new model, then tried the new model, and thats where I notice crashes in the logs. Now testing older models yields the same result. Note that above test are very minimal, only detection is enabled, no recordings or what not, no go2rtc, direct streaming from substreams of the cameras, vanilla as they say it. |
Beta Was this translation helpful? Give feedback.
-
|
If you are having issues with 1-2 cameras that does indicate some type of significant problem |
Beta Was this translation helpful? Give feedback.
-
Right? ssdlite works fine with 20 cameras, while yolonas, yolov9s or t, crashes the heck out with just 1 or 2 cameras. It must be software issue. It couldn't be hardware, since 20 cameras with default model, could at least match the resource requirement with just 1 camera of yolo models. If only I could install latest Intel Compute Runtime in the container for testing, however, Im stuck at this version: Anything later than that causes dependency issues. I tried building the source with Debian 13 as base, causes a lot of issues being python 3.11 is not available. Tried python 3.13, but lots of dependencies are not compatible with it. I couldn't just give up. That's why we homelab. |
Beta Was this translation helpful? Give feedback.
-
|
It is very unlikely to be any issue within Frigate, we are aware of many users running these models on OpenVINO with many cameras. |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Describe the problem you are having
Hi,
So I tried all the best I could to resolve this issue on my own, but I already spent 2 sleepless nights scratching my head on what happened. Why does, in all of a sudden, my frigate stopped working with Openvino despite it working very well before.
My setup is as follows:
Frigate 0.16.2 (I started with 0.16.1 but I think this is irrelevant since I retried with that version)
Cameras: 20, mix of Tapo and Dahua DVR, majority is Dahua
Hardware Setup:
Proxmox Ubuntu 24.04 VM
Minisforum MS01 12900H - 96GB RAM - 2TB SSD
Note: I used https://github.com/strongtz/i915-sriov-dkms, 2025.07.22 version to be specific
So iGPU is passed through via SRIOV using that DKMS.
It was working perfectly before. I woke up suddenly my Frigate+ instance is crashing.
I dont understand.
I tried nuking the VM, nuking the host's DKMS driver for iGPU SRIOV to no avail.
What I observed is that when using intel_gpu_top -d sriov on host, a sudden spike of Render/3D to 100%, then follows a crash in Frigate+.
So it was 20 cameras before, perfectly working. Now Im diagnosing with just 1 camera, and it is still happening. So I gave up, and I need your help on figuring out whats happening. Though sometimes, 1 camera does hold a little bit longer, but when I'm slowly readding cameras, it starts getting worse.
Version
0.16.2
Frigate config file
docker-compose file or Docker CLI command
Relevant Frigate log output
Install method
Proxmox via Docker
Object Detector
OpenVino
Screenshots of the Frigate UI's System metrics pages
Metrics UI does not load after crashing, only config, live, etc.
Any other information that may be helpful
No response
Beta Was this translation helpful? Give feedback.
All reactions