SOLVED: Intel A380 AV1 HW Transcoding Issues?

[danger0339]

Let me start out that if this is out of scope for this forum, please feel free to delete this, but I don't feel like I am getting very far in my own investigation, so I am turning to the community for potential guidance. There are clearly a multitude of variables at play here, so I readily admit there a million things that could potentially be wrong, but I am eager for any thoughts on what I may possibly want to explore.

So, first off, I am running Proxmox with a 14th generation Intel i7 CPU along with an Sparkle Intel A380 video card that I intended to use for transcoding Plex and Jellyfin content as an alternative to an old PowerEdge server that I was running previously with VMware and Plex.

In an effort to get off on a good start, I built a new VM with Ubuntu 24.04 LTS (to try to avoid legacy kernel issues that seemed to crop up with 22.04), to the best of my knowledge I have setup hardware pass thru of the A380 to the VM and the VM sees the card and I can do intel_gpu_top and it seems to see the right device and when I play content via Jellyfin, but admittedly it only seems to show activity on some of the content, but not on others.

Jellyfin is running in the linuxserver.io jellyfin container and I am using docker mod "jellyfin-opencl-intel" add-in. The compose file is rather cookie cutter in nature with me passing /dev/dri into the container and all the rest of it is just volumes really.

At any rate, the issue that has driven me here is that whenever I enable AV1 decoding in the transcoding features, the video never plays and in intel_gpu_top it shows the Compute row spike to 100% and I have to restart the container to be able to interact with Jellyfin to disable AV1 again. When AV1 is disabled in transcoding it shows activity in top and based on the banner across the top it is showing the A380 GPU (not the iGPU), so I don't understand by disabling AV1 it theoretically uses the GPU, but if I say "hey use hardware transcoding for AV1" it chokes and dies.

I feel like there must be a good guide or something out there on this and I am just missing it. I worry that I am not actually using my A380 and all of this is falling on the iGPU and I am just missing it somehow.

Any comments or feedback would be greatly appreciated. Once again, Jellyfin seems to work, I can watch all my stuff, but AV1 clearly doesn't work right and one of my family had issues with some content constantly buffering but for me locally it worked fine. That one may be related to his playback device (a Roku), but I have no idea to be honest.

Thoughts?

[toytown]

I would assume this is a bug as intel_gpu_top shouldn't be showing any activity from jellyfin unless its hardware (gpu) transcoding.  However to be 100% sure, turn off AV1 encoding, start a movie that needs transcoding, then grab the FFMPEG log for that playback and paste it here.

Then turn on AV1 encoding, playback the same content and again paste the new log here.

but admittedly it only seems to show activity on some of the content, but not on others.

Not all content needs transcoding, if the client has the capabilities (codec support) and has no bandwidth limitations it will just direct play/remux.

[danger0339]

After reading around some more, in this section:

https://jellyfin.org/docs/general/admini...l-on-linux
https://bugs.launchpad.net/ubuntu/+sourc...ug/2072755
https://github.com/jellyfin/jellyfin/issues/11380

It talks about the fact that in Ubuntu 24.04 LTS there is a known bug with the stock upgraded kernel (6.8.0-3 and the solution is to downgrade to 6.8.0-36 until a solution is found.

Once I did that I switched hardware decoding back on for AV1 it seems to work. Looking at my logs, I think I had always had the -38 kernel, so it explains the behavior at least to a degree that I have been observing. Now standard top shows a sleepy system, but intel_gpu_top shows a good amount of activity suggesting it is clearly using the card and it doesn't just spike the Compute item and force a restart of the docker container.

Scanning around in an AV1 video still doesn't feel instantaneous like it does for standard videos, but it is definitely a marked improvement over what was. Still not sure I have all the bells and whistles tuned correctly, but definitely making progress.

[TheDreadPirate]

Huzzah! You read documentation and found my bug report. You're my new favorite user.

One of the quirks I found with the 6.8.0-38 issue was that transcodes that scaled the resolution were the affected (4K to 1080P or 1080P to 720P). But transcodes that did not scale resolution were not always affected. I didn't test it that much before rolling back.

One of the two patches to fix this issue was already scheduled to be incorporated into -40. But the second patch that fully fixes the issue won't make it until the next kernel release.

[TheDreadPirate]

Ubuntu released 6.8.0-44 with all the patches that fix this issue.