Intel iGPU Platform, QSV higher GPU usage than VAAPI?

[buedi]

Hi everyone,

I run a 10400T with an UHD 630 on a Docker host which runs the Jellyfin Docker image. Works fine since many months, but today when checking GPU usage when transcoding a Video and checking with intel_gpu_top I have seen differences between using VAAPI and QSV. According to the documentation QSV seems to be preferred, because is it more performant than VAAPI, so I usually use QSV.

But when transcoding a single 1080p video, intel_gpu_top shows me 60-100% GPU usage when using QSV and 30-35% usage when using VAAPI.

Now performance is not equal usage. It could be of course that QSV is just much faster because it uses more GPU. But since the framerate of the Video is constant, I suspect that in my case both QSV and VAAPI have to transcode the same FPS and since QSV generates 3 times the GPU usage, it looks less performant to me.

On first sight, I would switch to VAAPI in that case, because lower GPU usage = less power consumption. Also with QSV my iGPU seems to be maxed out and with VAAPI I could transcode up to 3 streams. But maybe I draw the wrong conclusion here.

I wonder if this behavior is normal and other iGPU useres observe the same?

[Efficient_Good_5784]

Overall GPU utilization % isn't 1:1 as much as it seems. It can handle more than one transcoding session using QSV despite a single transcode taking at least 60%. You can test this out yourself by opening up multiple browser tabs and on each playing a different video that's transcoding.

The important part of a transcode is that the transcode fps remains at or higher of the video's fps (to prevent buffering).

You also have to keep in mind that energy used to produce work is measured as a unit over time. So if one method uses more energy at once but finishes the job faster, it's possible for it to use less watt-hours than the slower (and less energy-consuming) method.

You have to measure how much watts your energy usage increases with QSV and see how long it takes for it to complete a transcode. Then the same for VAAPI. You finally compare how much watt-hours each used to see which is more efficient in your system.