CPU Usage Still Relatively High with HWA. More filters to GPU? (/HWA config help req)

[Ale Lav]

Hello all, I'm very sorry if what I'm asking could've been found in a Jellyfin doc somewhere, but I don't think I saw this.

I finally got hardware acceleration to work out on my installation, it's an Intel® Core™ i5-3210M CPU running Debian Buster/Bookworm, set to VAAPI on its renderD128. I'm very sure that it's working because when I turn off HWA on Jellyfin settings, it's using 99.9% of all cores. It's a bit of an old repurposed laptop, not very powerful.

The thing is, when I have HWA turned on, it's actively still using 60%+ ish of every core at once, which is an improvement, sure, but I wonder if I can do better. Are there things I should change about my HWA configuration, or ways to offload more Jellyfin-FFMPEG filters to the GPU? What things is the CPU still having to do that the GPU can't be doing instead?

Right now I have every HWA setting as the default apart from VAAPI set and pointing to the renderD128 device in the Docker container.

Cheers, and thank you for your time

[TheDreadPirate]

It is likely still transcoding audio. If the subtitles are being burned in, that also happens on the CPU. Even if the GPU does the final encode.

[Ale Lav]

I am indeed burning in subtitles because they would, without fail, lose sync out of nowhere any time I watched anything when transcoding. Interesting to know that's the CPU doing that. I guess my question is how much control do I realistically have over the pipeline.

[TheDreadPirate]

None, unless your are a C# mega wizard and know Jellyfin's code base to make code changes. Regardless, I believe this is how ffmpeg works anyway.

[tdjf]

not complaining, just curious, what prevents that github intro skipper from making it into jellyfin as a built in feature?

[TheDreadPirate]

A first party plugin, or straight up integration into the server, is being worked on for a video analysis based intro/outro skipper function. But no ETA as there are other priorities at the moment.