Proxmox, Ubuntu, Docker, Jellyfin, Intel Arc A380

[mandopatriot]

Hi everyone,

Big fan of Jellyfin, been using it for a while now on different systems, primarily on Intel NUCs with Ubuntu server direct installed and running in Docker with iGPU usage for hardware transcoding. So I'm semi-familiar with hardware transcoding, but I'm sure I'm doing something wrong here.

I built a new system so I could run more VMs. Ryzen 9700x, SuperMicro H13SAE-MF, 192GB DDR5 ECC, 2x 1TB NVME SSD and 2x 1TB SATA SSD, Intel X710-DA2, and Intel Arc A380. I also have a Synology NAS that I am mounting to NFS, until I get all my new drives to build a new ZFS storage array.

I installed latest version of Proxmox with ZFS mirror on the 1TB SATA SSDs and am using a ZFS mirror for VM data on the 1TB NVME SSDs. I created a new Ubuntu Server 24.04 VM, with UEFI BIOS and machine type q35. I chose the HWE version of Ubuntu Server and after install, I shutdown and added the raw PCI devices for the GPU and audio controller. Both are set to full functions, ROM-BAR and PCI-Express both enabled.

In Ubuntu, I see both devices when running lspci. I followed the steps from Intel (Arc Ubuntu Install) for installing the packages, as well as adding my user to render and video groups. I see the GPU when I run "clinfo | grep "Device Name"" (recommended by Intel), I also see card0 and renderd128 in /dev/dri, and I see the GPU in intel_gpu_top. I also followed the Jellyfin docs to configure low power mode.

I installed Docker and copied over my compose file from my Intel NUC, as well as the Jellyfin data, and connected to my NFS share. I updated the compose file to point to the new data directory, and started it up. I'm using the linuxserver container with OpenCL mod. I'm able to connect, see all my previous data and media.

I updated transcoding page in Jellyfin to the supported codecs from the Intel QSV wiki. I'm using Intel QSV for hardware acceleration, QSV device set to /dev/dri/renderD128, and have all of these chcked - prefer os decoders, enable hardware encoding, and both Intel LP hardware encoders. I also have tone mapping enabled and set to defaults (not VPP).

When transcoding a video regardless of which device, I see higher than normal CPU usage. Intel_gpu_top show the Arc A380 is being used, but nothing in the Render/3d engine, only video and compute. I verified on my other system that Render/3d and video are both used (doesn't have compute), and the CPU usage is low.

Is this a configuration issue? Attached is one of the transcode logs (zipped) and some screenshots of what I see. I can get any other data or screenshots as needed, please let me know and thanks in advance for any help!

[ArneR]

I'm no expert, but as far as I can tell everything looks normal in the log. But do remember that audio is being transcoded by the cpu, not gpu, which could explain the usage.

[mandopatriot]

I'm no expert, but as far as I can tell everything looks normal in the log. But do remember that audio is being transcoded by the cpu, not gpu, which could explain the usage.

Thanks for looking into it. That's what I was thinking, but I barely see any CPU usage on the NUC when transcoding, and the 9700x is much more powerful than the 1250P in the NUC.

I also see render/3d usage when transcoding on the NUC, but I don't see it on the Arc A380. Granted, I don't know if that even matters (no idea what the engines do specifically), but figured it's worthwhile to note.

[ArneR]

Same here on my A310, only "video" and "videoenhance" graphs move when transcoding a tv-show right now. Don't think I've ever seen the "compute" graph before, on mine it just says "[unknown]"
Might be because I'm running Truenas Scale, idk.

   

[mandopatriot]

Same here on my A310, only "video" and "videoenhance" graphs move when transcoding a tv-show right now. Don't think I've ever seen the "compute" graph before, on mine it just says "[unknown]"
Might be because I'm running Truenas Scale, idk.

Awesome, thanks for confirming that it's working correctly. New to dedicated GPUs so I wasn't sure.

Side question - what settings do you use in Jellyfin for transcoding? Anything you recommend based on your experience with the A310, such as HEVC/AV1 encoding?