[Detector Support]: iGPU or ARC A380 GPU

[MomosX]

Describe the problem you are having

Hello,
I'm using Frigate in my Truenas Scale EE (24.10) with a Coral USB and HW acceleration form the Intel CPU with good results.
However i recently got an Intel ARC A380 GPU and wanted to configure Frigate to use that.

In the Frigate app from Truenas Scale EE there is only the option to pass GPU to the app.

This works to pass both the iGPU and the A380 to the Frigate App (confirmed looking at the transcode settings in Plex where both are available to choose from.)

However i'm not sure how to only choose the A380 as detector and not the iGPU in Frigate. Using the standard config seems to pick the iGPU, but i'm not sure about that either.

Hardware info from Frigate:
Vainfo Output
Return Code: 0
Process Output:
vainfo: VA-API version: 1.17 (libva 2.10.0)
vainfo: Driver version: Intel iHD driver for Intel(R) Gen Graphics - 23.1.1 ()

This would seem is the iGPU.

Any ideas please ?

Thank you

Version

0.14.1

Frigate config file

detectors:
  ov:
    type: openvino
    device: GPU
model:
  width: 300
  height: 300
  input_tensor: nhwc
  input_pixel_format: bgr
  path: /openvino-model/ssdlite_mobilenet_v2.xml
  labelmap_path: /openvino-model/coco_91cl_bkgr.txt

docker-compose file or Docker CLI command

-

Relevant Frigate log output

-

Relevant go2rtc log output

-

FFprobe output from your camera

-

Install method

Docker Compose

Object Detector

OpenVino

Network connection

Wired

Camera make and model

Wyze V3

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

No response

[NickM-27]

the main way to do this is via docker, so it is not great that truenas does not let you select a particular GPU

[MomosX]

There is not much flexibility in the Truenas Electric Eel apps. They are good for novice users like myself but then you get into this sort of issues. There is however terminal access to the Frigate container.
What would be the secondary way please ?
Cheers

[MomosX]

Turns out there is a way to see what the ARC card doing in Truenas 24.10.

Seems like Frigate is using the card also. But the performance is not very good.

Using the Coral USB , Detector CPU usage never looked so busy.

[NickM-27]

right, it is using a different model which requires different steps to prepare the image

[MomosX]

Cheers ! will keep an eye on the performance using the ARC card and switch back to Coral if needed.

[MomosX]

Hi, i reopened the topic because i simply don't understand how it works :(

Please let me explain in detail with pics worth a 1000 words:

Intel iGPU and Intel ARC A380 , both available to Frigate

This is what is reported from gpu top, both outside the Frigate contaier and inside it.

I gave up for now to use the A380 as detector, switched back to Coral USB.

I would like to make the HW acceleration work on the A380.

TEST1:

Setting it up like this:
detectors: coral_usb: type: edgetpu device: usb ffmpeg: hwaccel_args: -hwaccel vaapi -hwaccel_device /dev/dri/renderD129

works, but the result is weird: huge CPU usage, very bad Coral Inference.

gpu-top shows that A380 is used:

TEST2:

Setting up like this:
detectors:
coral_usb:
type: edgetpu
device: usb
ffmpeg:
hwaccel_args: preset-vaapi

works and the result is good:

but gpu-top shows that the A380 is not used, instead is used the iGPU

I did all sorts of tests and settings. Can you @NickM-27 please help an old timer figure this out ?

thank you !!!

[NickM-27]

There is no way to specify a different GPU be used when both are intel and both inside of the container. I am not familiar with truenas enough to suggest how you would limit it to only one GPU

[MomosX]

I am trying to limit it to one GPU, but to the A380 m not the iGPU.
renderD129 specified in ffmpeg will pick the A380, renderD128 will pick the iGPU.

But the results with renderD129 are very bad.

How do i reopen the topic maybe someone has the same scenario please ? Can't seem to find the button...

[MomosX]

Ha ! i managed to solve it !
As a result CPU usage dropped from 18-23% to 3-6%, memory usage droped with about 50% also ! And inference speed on the Coral is better also. Power for the entire sistem went down from 110-115 W to 85-95 W - probably because the CPU was drawing more power then the A380 is.

I would say it's a good result :)

For others who might be struggling with something similar the config that worked is:

detectors:
coral_usb:
type: edgetpu
device: usb

ffmpeg:
input_args:
- -rtsp_transport
- tcp
- -r
- '25'
hwaccel_args:
- -hwaccel
- vaapi
- -hwaccel_device
- /dev/dri/renderD129
- -hwaccel_output_format
- yuv420p

Will start playing next with openvino detector to compare with the coral.

[NickM-27]

Well in general both of the sets of args being manual and not a preset is problematic. But if you are set on using the A380 then the hwaccel args will have to stay if there is no way to specify this in the truenas config.

and to be clear, we already have had a few users including myself testing with Arc cards in Frigate

[MomosX]

Using a preset is the best of course.
I was merely trying multiple options , trial and error to see what works.

Is it possible to have something like ?
ffmpeg:
hwaccel_args:
- preset-vaapi
- hwaccel_device /dev/dri/renderD129

So to use a preset but specify the device ? It doesnt work in the form i wrote it.
I think it all boils down to be able to specify the renderD129 and it should work in theory.

[NickM-27]

Like I said currently it is not possible, as generally the preset will try all GPUs and pick the first one that works which in this case is the iGPU. We could add an explicit option to the config but this is not currently supported

[MomosX]

ok. thank you very much for the patience:)
would be great if such an explicit option would be allowed/added , from all my searches there are quite a few folks that struggle with the same issue.
Cheers !

[MomosX]

Hi,
An update as i wrote in the Truenas forum and devs replied that "Dockers GPU support for selecting is limited to nvidia only."
So nothing they can do about selecting Intel GPU's for now.
https://forums.truenas.com/t/intel-gpu-passthrough-to-apps-scale-ee-24-10/27956/3

Plex app in the same Truenas system has this option to select between GPU and it's working, so hopefully when you get a chance and time maybe this could somehow be implemented also.

Thank you !!!