[HW Accel Support]: VAAPI Hardware Acceleration Failing on Intel J4125 post upgrade to v0.17

[dr3amr]

Describe the problem you are having

On version 0.16.4 acceleration was working properly with low CPU 30-40% usage and 20-30% GPU usage on my intel Gemini Lake (Gen9) J4125. I am running Frigate in a Docker container (as an Application) as part of the TrueNAS OS (Debian Bookworm / 6.12.33 kernel). Post upgrading, same configuration, CPU usage is 100% and GPU is 0%. I was using i965 driver with general ffmpeg hwaccel_args: preset-vaapi. The Vaapi seems to not work also if changing the driver to iHD. Acceleration is somewhat restored if using the preset-intel-qsv-h264 preset but with about 60-70% CPU load. Is intel VAAPI no longer working in 0.17 version ? although it is still recommended in the documentation, perhaps something I'm missing.

Version

0.17

Frigate config file

environment_vars:
  LIBVA_DRIVER_NAME: i965
#  LIBVA_DRIVER_NAME: iHD

ffmpeg:
#  hwaccel_args: preset-intel-qsv-h264
  hwaccel_args: preset-vaapi
  input_args: -avoid_negative_ts make_zero -analyzeduration 2147483647 
    -probesize 2147483647 -fflags +genpts+discardcorrupt -rtsp_transport tcp 
    -timeout 5000000 -buffer_size 10000000 -use_wallclock_as_timestamps 1
  output_args:
    record: -ss 40 -f segment -segment_time 30 -segment_format mp4 
      -reset_timestamps 1 -strftime 1 -threads 4 -c:v copy -c:a aac -profile:a 
      aac_low -b:a 64k

docker-compose file or Docker CLI command

Using the container as TrueNAS App, no docker-compose.

Relevant Frigate log output

2026-03-01 09:41:22.862450177  [2026-03-01 11:41:22] ffmpeg.Curte_Spate.detect      ERROR   : [aost#0:1/aac @ 0x55b95edc14c0] Non-monotonic DTS; previous: 831137, current: 831042; changing to 831138. This may result in incorrect timestamps in the output file.
2026-03-01 09:41:22.862740827  [2026-03-01 11:41:22] ffmpeg.Curte_Spate.detect      ERROR   : [aost#0:1/aac @ 0x55b95edc14c0] Non-monotonic DTS; previous: 831138, current: 831107; changing to 831139. This may result in incorrect timestamps in the output file.
2026-03-01 09:41:22.863024777  [2026-03-01 11:41:22] ffmpeg.Curte_Spate.detect      ERROR   : [aost#0:1/aac @ 0x55b95edc14c0] Non-monotonic DTS; previous: 831201, current: 831122; changing to 831202. This may result in incorrect timestamps in the output file.
2026-03-01 09:41:22.866770619  [2026-03-01 11:41:22] ffmpeg.Curte_Spate.detect      ERROR   : [aac @ 0x55b95edc5580] Queue input is backward in time
2026-03-01 09:41:22.877007087  [2026-03-01 11:41:22] ffmpeg.Curte_Spate.detect      ERROR   : [aost#0:1/aac @ 0x55b95edc14c0] Non-monotonic DTS; previous: 831245, current: 831163; changing to 831246. This may result in incorrect timestamps in the output file.
2026-03-01 09:41:22.877595254  [2026-03-01 11:41:22] ffmpeg.Curte_Spate.detect      ERROR   : [in#0/rtsp @ 0x55b95ed8c540] Error during demuxing: Connection timed out
2026-03-01 09:41:22.879473242  [2026-03-01 11:41:22] ffmpeg.Curte_Spate.detect      ERROR   : [aac @ 0x55b95edc5580] Queue input is backward in time
2026-03-01 09:41:22.880132545  [2026-03-01 11:41:22] ffmpeg.Curte_Spate.detect      ERROR   : [aost#0:1/aac @ 0x55b95edc14c0] Non-monotonic DTS; previous: 831328, current: 831281; changing to 831329. This may result in incorrect timestamps in the output file.
2026-03-01 09:41:22.880913216  [2026-03-01 11:41:22] ffmpeg.Curte_Spate.detect      ERROR   : [aost#0:1/aac @ 0x55b95edc14c0] Non-monotonic DTS; previous: 831424, current: 831329; changing to 831425. This may result in incorrect timestamps in the output file.
2026-03-01 09:41:22.881509166  [2026-03-01 11:41:22] watchdog.Curte_Spate           INFO    : Restarting ffmpeg...
2026-03-01 09:42:00.736878646  [2026-03-01 11:42:00] frigate.video                  ERROR   : Curte_Fata_2: Unable to read frames from ffmpeg process.
2026-03-01 09:42:00.740689671  [2026-03-01 11:42:00] frigate.video                  ERROR   : Curte_Fata_2: ffmpeg process is not running. exiting capture thread...
2026-03-01 09:42:02.185042518  [2026-03-01 11:42:02] watchdog.Curte_Fata_2          ERROR   : Ffmpeg process crashed unexpectedly for Curte_Fata_2.
2026-03-01 09:42:02.186156798  [2026-03-01 11:42:02] watchdog.Curte_Fata_2          ERROR   : The following ffmpeg logs include the last 100 lines prior to exit.
2026-03-01 09:42:02.186918961  [2026-03-01 11:42:02] ffmpeg.Curte_Fata_2.detect     ERROR   : [aac @ 0x558265b406c0] Too many bits 8192.000000 > 6144 per frame requested, clamping to max
2026-03-01 09:42:02.194944785  [2026-03-01 11:42:02] watchdog.Curte_Fata_2          INFO    : Restarting ffmpeg...

Relevant go2rtc log output

No errors

FFprobe output from your camera

Stream 1
Video:
Codec: H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10
Resolution: 2592x1944 (4/3 aspect ratio)
FPS: 15/1
Audio:
Codec: PCM A-law / G.711 A-law

Install method

Docker Compose

Object Detector

Coral

Network connection

Wired

Camera make and model

Misecu x 4 5MP / Asecam x 2 8MP. PoE enabled.

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

No response

[inkeep-ai-assistant[bot]]

Based on the symptoms you're describing—VAAPI hardware acceleration working in v0.16.4 but failing in v0.17 with 100% CPU usage and 0% GPU usage on your Intel J4125 (Gemini Lake/Gen9)—there are several troubleshooting steps to consider(1).

Hardware Acceleration Configuration

Hardware acceleration is critical for reducing CPU usage when decoding video streams(1). The preset-vaapi configuration you're using should offload video decoding to your Intel integrated GPU(1)(2).

Your Intel J4125 has Gen9 graphics, which should support VAAPI. However, the driver selection matters:

• i965 driver: Older Intel driver, typically for Gen 4-9 graphics
• iHD driver: Newer Intel Media Driver, typically for Gen 8+

Since you've tried both LIBVA_DRIVER_NAME: i965 and iHD without success in v0.17, this suggests a potential compatibility issue with the newer Frigate version or underlying FFmpeg changes.

Troubleshooting Steps

1. Verify GPU Access: Ensure the Docker container has proper access to /dev/dri/renderD128(3). In your TrueNAS setup, confirm the device is passed through correctly:

devices:
  - /dev/dri/renderD128:/dev/dri/renderD128

(3)

2. Check FFmpeg Logs: The logs you provided show timestamp and connection issues, but not hardware acceleration failures. Look for specific VAAPI initialization errors in the Frigate logs that would indicate why hardware acceleration isn't being used.

3. Verify VAAPI Support: Inside the Frigate container, run:

vainfo

This will confirm if VAAPI is properly detected and which driver is active.

4. Alternative Preset: Since you mentioned preset-intel-qsv-h264 partially works (60-70% CPU), this indicates the GPU is accessible but VAAPI specifically may have issues. QSV (Quick Sync Video) uses a different code path than VAAPI(2).

Configuration Recommendation

For Intel integrated GPUs, the standard configuration should be(2):

ffmpeg:
  hwaccel_args: preset-vaapi

(2)

With the device mapping:

devices:
  - /dev/dri/renderD128:/dev/dri/renderD128

(3)

Potential v0.17 Changes

While the documentation still recommends VAAPI for Intel hardware(1)(2), there may have been underlying FFmpeg version changes between v0.16.4 and v0.17 that affect VAAPI compatibility with Gen9 graphics.

Next Steps

1. Verify /dev/dri/renderD128 is accessible in the container
2. Check Frigate startup logs for VAAPI initialization messages
3. Run vainfo inside the container to confirm driver detection
4. If VAAPI continues to fail, consider using preset-intel-qsv-h264 as a workaround, though it's less efficient
5. Check if there are any kernel or mesa driver updates available for your TrueNAS system that might improve compatibility

The fact that QSV partially works confirms your hardware is capable—the issue appears specific to the VAAPI code path in v0.17.

📚 Sources:

1. High CPU Usage
2. Getting Started
3. Installation

---

Have another question? Just tag @inkeep.

[dr3amr]

vainfo seems to show that the iHD driver is loaded even though LIBVA_DRIVER_NAME: i965 is added to frigate config environment_vars section.
LE: seems the solution is to add the environment variable "LIBVA_DRIVER_NAME: value" directly in the container definition / docker compose, as it no longer works while defined in Frigate config (bug? @blakeblackshear ). I know this option was intended to work for the Frigate HA addon, but this worked for rest of deployments in 0.16x versions. For J4125 CPU and GPU the combination of i965 + preset-vaapi works best, and given the i965 driver was not getting loaded, and instead only iHD (default driver) was loaded, it was not functioning properly. preset-vaapi does not work at all with the iHD driver, at least for this CPU / GPU. Now time to enjoy / test the new yolov9s model for Coral :).
LE2: seems the new yolov9s has rather 12-15ms than the reference 10ms on the Coral USB version. CPU usage also increased a bit to 50-60% (includes all ffmpeg processing for 6 cameras).