[HW Accel Support]: How to ensure VAAPI is actually used for hardware acceleration (AMD APU + LXC + Docker)

[darlezior]

Describe the problem you are having

Hi everyone 👋

I'm running Frigate inside a Docker container, hosted in a Proxmox LXC, on a system using an AMD Ryzen 7 5800U APU. I've passed through the GPU using the following devices to the LXC:

/dev/dri/card0 /dev/dri/renderD128

Inside the container:

• vainfo works and reports supported VAAPI profiles for H.264, HEVC, VP9, etc.
• /dev/dri is visible inside the Docker container.
• Frigate is configured with:

ffmpeg:
  hwaccel_args: preset-vaapi

However, I'm not 100% sure that Frigate is actually using VAAPI for hardware decoding.

I still notice:

High CPU usage (FFmpeg processes often 20–30% per camera).

All CPU cores going into boost mode.

radeontop inside the container shows 0% video decode usage, even during active detections and recording.

The Coral TPU stays active even though motion detection is required first (already configured).


Here are some relevant details:

Cameras are 4K, but detection is only enabled on substreams (lower resolution).

VAAPI is available via Mesa driver: radeonsi.

LXC is configured with full access to /dev/dri, and the Docker container sees it too.

Host is using kernel 6.8.x, Proxmox 8.x.
------
My questions:

1. Is there a reliable way to confirm that VAAPI is actually being used by Frigate's FFmpeg processes?
2. Do AMD APU setups like mine require any manual configuration beyond preset-vaapi?
3. Should I override hwaccel_args to manually set something like:

-hwaccel vaapi -hwaccel_device /dev/dri/renderD128 -vaapi_device /dev/dri/renderD128

...instead of the preset?

4. Could the LXC/Docker layering prevent full GPU access even if /dev/dri is passed through?

5. Are there any debugging flags or FFmpeg log levels I can increase to confirm HW decoding?

Actually I deactivated the HD streams but the issue Is Always there.

Any help or advice would be hugely appreciated 🙏

### Version

0.15.0-6cb5cfb

### Frigate config file

```yaml
mqtt:
  enabled: true
  host: 192.168.1.102
  port: 1883
  user: 
  password: 

ffmpeg:
  hwaccel_args:
    - -hwaccel
    - vaapi
    - -hwaccel_device
    - /dev/dri/renderD128
    - -hwaccel_output_format
    - yuv420p
  output_args:
    record: preset-record-generic

detectors:
  coral:
    type: edgetpu
    device: usb

objects:
  track:
    - person
    - dog
    - cat

go2rtc:
  streams:
    piazzale:
      - rtsp://[email protected]/Preview_01_sub
    giardino:
      - rtsp://[email protected]/Preview_01_sub
    salone:
      - rtsp://[email protected]/12
    ingresso:
      - rtsp://[email protected]/12

snapshots:
  enabled: true
  clean_copy: true
  quality: 80

record:
  enabled: true
  retain:
    days: 10
    mode: all
  alerts:
    retain:
      days: 14
  detections:
    retain:
      days: 14

motion:
  threshold: 25
  contour_area: 10
  frame_height: 360
  improve_contrast: true

cameras:
  #################=PIAZZALE=#################
  piazzale:
    enabled: true
    record:
      enabled: true
    detect:
      enabled: true
      fps: 5
      max_disappeared: 25
      stationary:
        interval: 10
        threshold: 50
    ffmpeg:
      inputs:
        - path: rtsp://127.0.0.1:8554/piazzale
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
    motion:
      enabled: true
      threshold: 30
      contour_area: 10
      improve_contrast: true
    objects:
      filters:
        person:
          min_area: 9000
          max_area: 707000
        dog:
          min_area: 7500
          max_area: 707000
        cat:
          min_area: 4000
          max_area: 707000

  piazzale_hd:
    enabled: false
    detect:
      enabled: false
    record:
      enabled: false
    snapshots:
      enabled: false
    ffmpeg:
      inputs:
        - path: rtsp://[email protected]/Preview_01_main
          input_args: preset-rtsp-generic
          roles:
            - record

  #################=GIARDINO=#################
  giardino:
    enabled: true
    record:
      enabled: true
    detect:
      enabled: true
      fps: 5
      max_disappeared: 25
      stationary:
        interval: 10
        threshold: 50
    ffmpeg:
      inputs:
        - path: rtsp://127.0.0.1:8554/giardino
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
    motion:
      enabled: true
      mask:
        - 0.298,0.000,0.737,0.000,0.735,0.089,0.285,0.089
      threshold: 30
      contour_area: 10
      improve_contrast: true
    objects:
      filters:
        person:
          min_area: 4500
          max_area: 367000
        dog:
          min_area: 3500
          max_area: 367000
        cat:
          min_area: 2500
          max_area: 367000

  giardino_hd:
    enabled: false
    detect:
      enabled: false
    record:
      enabled: false
    snapshots:
      enabled: false
    ffmpeg:
      inputs:
        - path: rtsp://[email protected]/Preview_01_main
          input_args: preset-rtsp-generic
          roles:
            - record

  #################=SALONE=#################
  salone:
    enabled: true
    detect:
      enabled: true
      fps: 5
      max_disappeared: 25
      stationary:
        interval: 10
        threshold: 50
    ffmpeg:
      inputs:
        - path: rtsp://127.0.0.1:8554/salone?video&audio
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
    objects:
      filters:
        person:
          min_area: 3500
          max_area: 286000
        dog:
          min_area: 2500
          max_area: 286000
        cat:
          min_area: 1500
          max_area: 286000
    onvif:
      host: 192.168.1.35
      port: 8080
      user: 
      password: 
    motion:
      enabled: true
      mask:
        - 0.759,0.000,0.769,0.061,0.996,0.066,0.987,0.000
      threshold: 30
      contour_area: 10
      improve_contrast: true

  #################=INGRESSO=#################
  ingresso:
    enabled: true
    record:
      enabled: true
    detect:
      enabled: true
      fps: 5
      max_disappeared: 25
      stationary:
        interval: 10
        threshold: 50
    ffmpeg:
      inputs:
        - path: rtsp://127.0.0.1:8554/ingresso
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
    objects:
      filters:
        person:
          min_area: 3500
          max_area: 286000
        dog:
          min_area: 2500
          max_area: 286000
        cat:
          min_area: 1500
          max_area: 286000
    onvif:
      host: 192.168.1.38
      port: 8080
      user: 
      password: 

  ingresso_hd:
    enabled: false
    detect:
      enabled: false
    record:
      enabled: false
    snapshots:
      enabled: false
    ffmpeg:
      inputs:
        - path: rtsp://xxxxxx:[email protected]/11
          input_args: preset-rtsp-generic
          roles:
            - record

camera_groups:
  Esterno:
    order: 1
    icon: LuTrees
    cameras:
      - ingresso
      - piazzale
      - giardino
version: 0.15-1

docker-compose file or Docker CLI command

version: '3.9'
services:
  frigate:
    container_name: frigate
    privileged: true
    restart: unless-stopped
    image: ghcr.io/blakeblackshear/frigate:stable
    shm_size: "1024mb" # update for your cameras based on calculation above
    devices:
      - /dev/bus/usb:/dev/bus/usb
      - /dev/dri/renderD128 # for intel hwaccel, needs to be updated for
your hardware
    volumes:
      - /etc/localtime:/etc/localtime:ro
      - /opt/frigate/config:/config
      - /cctv_clips:/media
      - type: tmpfs # Optional: 1GB of memory, reduces SSD/SD Card wear
        target: /tmp/cache
        tmpfs:
          size: 1000000000
    ports:
      - "5000:5000" # porta non autenticata
      - "443:8971" # porta autenticata
      - "1935:1935" # RTMP feeds
      - "1984:1984" # go2rtc
    environment:
      FRIGATE_RTSP_PASSWORD: ""
      FRIGATE_MQTT_PASSWORD: ""
      FRIGATE_MQTT_USER: ""
      LIBVA_DRIVER_NAME: "radeonsi"
      PLUS_API_KEY: "-"

Relevant Frigate log output

ND

Relevant go2rtc log output

ND

FFprobe output from your camera

ND

Install method

Proxmox via Docker

Object Detector

Coral

Network connection

Wired

Camera make and model

Reolink and Ctronics

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

No response

[inkeep-ai-assistant[bot]]

For AMD/ATI GPU hardware acceleration with VAAPI, you need to ensure the correct driver configuration. Based on your setup, there are several issues to address:

Configuration Issues

Your current configuration uses manual VAAPI arguments, but for AMD GPUs, you should use the preset-vaapi configuration (1). The documentation states that VAAPI supports automatic profile selection and works with both H.264 and H.265 streams (1).

You need to change the driver to radeonsi by adding the environment variable LIBVA_DRIVER_NAME=radeonsi (1). I can see you already have this in your docker-compose file, which is correct.

Recommended Configuration

Replace your current manual hwaccel_args with the preset:

ffmpeg:
  hwaccel_args: preset-vaapi

(1)

Verification Methods

To verify hardware acceleration is working, the documentation provides specific testing commands. For VAAPI, you can test inside the container with:

ffmpeg -hwaccel vaapi -hwaccel_device /dev/dri/renderD128 -hwaccel_output_format yuv420p -re -stream_loop -1 -i https://streams.videolan.org/ffmpeg/incoming/720p60.mp4 -f rawvideo -pix_fmt yuv420p pipe: > /dev/null

(2)

Hardware Detection

The hwaccel presets not only replace longer arguments but also give Frigate hints about available hardware, allowing Frigate to make other optimizations using the GPU such as when encoding the birdseye restream or scaling streams (3).

Troubleshooting

Based on a similar case with AMD GPU in Docker, ensure you have the correct devices mapped and that vainfo shows proper driver initialization (4). The fact that vainfo works and shows supported profiles suggests the driver setup is correct, but using the preset instead of manual arguments should improve compatibility.

The high CPU usage you're experiencing suggests hardware acceleration may not be active. Switch to preset-vaapi and monitor the system metrics to see if CPU usage decreases significantly during video processing.

📚 Sources:

1. Hardware Acceleration
2. Contributing To The Main Code Base
3. FFmpeg presets
4. Trying to use AMD GPU in Docker container on Unraid

---

Have another question? Just tag @inkeep.

[darlezior]

Thank you for the detailed response.

I've removed all manual hwaccel_args and switched to the recommended preset-vaapi in the config.
Also, I confirmed that the LIBVA_DRIVER_NAME=radeonsi environment variable is set correctly.

I tried running the VAAPI test command inside the Frigate container, but it returned:
bash: ffmpeg: command not found

It seems ffmpeg is not available within the container, or not in the default path.
Should I test hardware acceleration using a separate container instead?

Thanks again for the help!

[NickM-27]

Looks like it made a slight mistake. Ffmpeg is not part of $path so you have to run the fully qualified path. But with the changes made you should be able to run frigate and verify via GPU stats that it is running.

[darlezior]

Tests performed:

1. Test with ffmpeg inside the container

/usr/lib/ffmpeg/7.0/bin/ffmpeg -loglevel verbose
-init_hw_device vaapi=va:/dev/dri/renderD128 -filter_hw_device va
-hwaccel vaapi -hwaccel_output_format vaapi
-rtsp_transport tcp -i rtsp://:@/Preview_01_sub
-vf 'scale_vaapi=1280:720,hwdownload,format=yuv420p'
-f null -

Confirms radeonsi_drv_video.so is loaded correctly.

pix_fmt: vaapi is being set.

2. radeontop on the Proxmox host while Frigate is running

Video Decode stays near 0% even with active RTSP streams.

Overall GPU Load is very low.

3. VCN status check

cat /sys/kernel/debug/dri/0/amdgpu_pm_info

Output:

GPU Load: 1 %
VCN: Powered up
Unified Video Decoder Medium Grain Clock Gating: Off
VCN Medium Grain Clock Gating: Off

Indicates that VCN is powered up but apparently not handling video decoding.

---

Even though VAAPI seems configured, the CPU cores assigned to Frigate remain in constant boost, with CPU temperature fixed between 80 °C and 90 °C.
This suggests decoding is still happening in software instead of hardware acceleration.

LIBVA_DRIVER_NAME=radeonsi is already set.

ffmpeg preset changed to preset-vaapi as recommended in the docs.

---

{"return_code":0,"stderr":"","stdout":"vainfo: VA-API version: 1.22 (libva 2.10.0)
vainfo: Driver version: Mesa Gallium driver 22.3.6 for AMD Radeon Graphics (renoir, LLVM 15.0.6, DRM 3.57, 6.8.12-13-pve)
vainfo: Supported profile and entrypoints
VAProfileMPEG2Simple : VAEntrypointVLD
VAProfileMPEG2Main : VAEntrypointVLD
VAProfileVC1Simple : VAEntrypointVLD
VAProfileVC1Main : VAEntrypointVLD
VAProfileVC1Advanced : VAEntrypointVLD
VAProfileH264ConstrainedBaseline: VAEntrypointVLD
VAProfileH264ConstrainedBaseline: VAEntrypointEncSlice
VAProfileH264Main : VAEntrypointVLD
VAProfileH264Main : VAEntrypointEncSlice
VAProfileH264High : VAEntrypointVLD
VAProfileH264High : VAEntrypointEncSlice
VAProfileHEVCMain : VAEntrypointVLD
VAProfileHEVCMain : VAEntrypointEncSlice
VAProfileHEVCMain10 : VAEntrypointVLD
VAProfileHEVCMain10 : VAEntrypointEncSlice
VAProfileJPEGBaseline : VAEntrypointVLD
VAProfileVP9Profile0 : VAEntrypointVLD
VAProfileVP9Profile2 : VAEntrypointVLD
VAProfileNone : VAEntrypointVideoProc"}

---

[abovevirtual]

Following as I am having a similar problem, not on proxmox running bare metal but same issue in my docker compose that my CPU is being hit hard and GPU is not really much at all but have a similar configs and all match the provided documentation

[darlezior]

Hi there, still same issue here. Is there any chance to solve It? What could I try?

[shineangelic]

same here, in my case using an older radeon with ffmpeg: hwaccel_args: preset-vaapi, but supported as viewed in vainfo:

I am also experiencing low GPU loads, apart from used GPU memory:

average radeontop:

I have A very similar configuration, with salotto and cancello :)