VAAPI vs QSV resource usage on Intel Arc A310, Ubuntu

[d3relict]

i have an Intel A310 dedicated GPU which i use in frigate.

there is a huge difference in resource consumption between VAAPI and QSV, and i'd like to understand what this is, and if this is intentional.

VideoEnhance usage immediately increases tenfolds when i enable QSV, and iowait doubles on the cpu, yet the user space consumption stays about the same, or drops about 3-5% (this is random, i have other test runs, where the user space usage did not change at all). the power usage of the machine raises about 10 watts (according to the UPS) attached to it.

all streams are h.264, and mostly ingested via RTSP, but i have 6 streams with this setup:

living_room_1:
      - ffmpeg:http://xxxx/flv?port=1935&app=bcs&stream=channel0_main.bcs&user=XXX&password=XXX#video=copy#audio=copy#audio=opus

hwaccel args are passed in the config root as such:

ffmpeg:
  hwaccel_args: preset-intel-qsv-h264
  # hwaccel_args: preset-vaapi

docker compose parts:

    devices:
      - /dev/bus/usb
      - /dev/dri

using Frgate 0.16.0-c2f8de9, Ubuntu 24.04, kernel 6.8.0-78-generic

afaik there are nothing else relevant, but please tell me if there is anything else i shoud share.

[inkeep-ai-assistant[bot]]

On Intel Arc GPUs, both VAAPI and QSV are supported for hardware-accelerated video decoding in Frigate, and you can choose either by setting the hwaccel_args in your configuration. For Intel Arc, the documentation states both preset-vaapi and preset-intel-qsv-* are supported options(1).

Resource usage differences between VAAPI and QSV can occur due to how each driver and hardware interface manages decoding workloads. The documentation does not specify that one should always be more efficient than the other, nor does it indicate that a large difference is intentional(1). It is normal to see differences in GPU and CPU utilization depending on which method is used, as QSV and VAAPI use different paths in the Intel graphics stack.

If you notice significantly higher VideoEnhance usage and increased iowait with QSV compared to VAAPI, this aligns with user observations in the community, but there is no official explanation or recommendation that one is always superior for Arc GPUs(1)(2). Either can be used based on your observed performance.

For Intel Arc, both of the following are valid:

ffmpeg:
  hwaccel_args: preset-vaapi

or

ffmpeg:
  hwaccel_args: preset-intel-qsv-h264

(1)

If you want to further investigate the efficiency difference, you may want to compare driver versions, host kernel, and FFmpeg builds, as these can impact hardware acceleration performance. There is no explicit setting in Frigate to optimize further between the two outside of the documented presets(1).

📚 Sources:

1. Video Decoding
2. [HW Accel Support]: Intel-qsv or intel-vaapi is used

---

Have another question? Just tag @inkeep.

[NickM-27]

The AI is not really correct here, in my testing VAAPI does not work well on the Arc cards. The Arc cards run the newer Xe driver which is only designed to be used with QSV and Intels new media SDK

[d3relict]

how should this manifest? i see no functional differences, i only observe the usage values change drastically.

otherwise i'm sure it's not cpu decoded, because user percentages essentially the same for both cases (even if i remove the hwaccel_args it still uses the gpu somehow), and only if i remove the device from the compose file completely, then cpu usage goes up about 5%.

[NickM-27]

When I tested between them GPU usage was substantially higher with VAAPI and there was more latency in the decoding. Switching to QSV and everything was much faster.

[d3relict]

did you measure the latency, or was it simply perceivable? any data or "guesstimate" about the magnitude of the difference?

[d3relict]

i checked both QSV and VAAPI, and found that QSV has around 2-3 seconds of latency, while VAAPI preset is almost instantaneous. i was using the debug view, and gave visual cues on camera, and measured when they were displayed.