Hi so i'm looking at rebuilding my nas setup and i'm really tempted to buy something like an Arc A380, with all it's accelerators and surprisingly good support!
I'm running a Jellyfin server and a couple other dockers that can take advantage of either the accelerators or genereal compute of a GPU.
My question is, now that SR-IOV is supports the gpu to be split between multiple VMs, would that also allow the GPU to be split between VMs and Dokers?
I would love to use the card for Jellyfin, but not have it inaccesible from being bound to vfio, once another application or VM is utilising it.

Has anyone had good experience with sharing the hardware in this manner?

Edited April 24, 20251 yr by jztreso
Typo

Personally, I don't see the use of the Intel Arc GPU yet. They are getting better and closer, but a Nvidia 2080 still beats some of them and has more versatile uses.

I virtualize Unraid in a Proxmox system that vGPU splits the 2080 into two GPUs — one for Unraid, the other for a Windows gaming machine.

So, let's review some things first: in VM space, the splitting of a GPU for use is called vGPU or MxGPU. What you have described is MxGPU.
At the moment (ATM), I'm not aware of Intel cards supporting that feature. AMD had some EOL GPUs that supported MxGPU, but Nvidia beat them to it and has better online support through open-source Linux communities.

Also, with other Linux distros, people are using "unlock scripts" to use consumer Nvidia cards for vGPU.
Unraid is based on Slackware Linux. Slackware can do vGPU, but Unraid doesn’t have a package manager available, and doing so may require kernel edits to Slackware to get it working.

 

However, the way Docker works is that it pulls resources from the host to run applications.
Docker, since it uses host devices and shared resources, can use the GPU for tasks (this is similar to opening a movie on Windows and then opening Chrome and playing a YouTube video — they both play and use the GPU).

It’s the same with an LXC.
I would recommend using Dockers and an LXC version rather than VFIO-ing the Intel GPU — use it instead for transcoding in a Jellyfin Docker container.

SR-IOV is important, but it doesn’t mean nor guarantee GPU partitioning.

SR-IOV is needed for IRQ memory mapping and forwarding between host and virtual devices.

 

With Unraid, Jellyfin requires a Docker mod option to use the Intel Arc GPU, as seen here:

 

 

In Unraid, you may (or will) need to install plugin drivers as well.

 

if you want to run some transcode / video intensive application and want a OS. I recommend looking into LXCs on Unraid.

For example, As I do this with a nvidia GPU. I run a plex docker and Local AI. I can use both at the same time as the are both dockers and the docker will use the nvidia GPU when resources demand...

FOr VM, If you vfio bind the intel gpu to pass it to a VM for lets say windows. the the g-card is removed from the host and jellyfin / xyz docker can no longer use the intel card...

Sadly, Unraid doesn't have a vgpu / mxgpu setup plugin or installing to partition like windows gpup or proxmox vgpu setups...
The devs don't like unraid virtualized due to slackware requirements...

But is it possible on unraid yes...

once could use hook scripts but would need to run one or the other not both at the same time...

Edited April 25, 20251 yr by bmartino1
Data - typo - Forum weridness

I'm happy with the transcoding performance but I will say that it's more limited in total use cases.

 

I wanted a dedicated GPU for it on my Ryzen based system and still have a GPU for Gaming VM for my wife.

 

It does what I want for the most part but one limitation that I have run into is that Tunarr doesnt accept Intel ARC and it has to transcode to work properly. So at best I now have to split things up. I do have Steam-Headless docker installed now for my wife's gaming VM instead of running a VM which works but not my favorite way to do it.

 

bmartino gave me alot of help getting ARC working in Jellyfin so it works there but he did have to help me with a modification.

 

 

I had an A770 for a couple years, sold it on ebay for what I paid for it and bought a 4070 super a few months ago before the prices went nuts again. The A770 was an absolute unit for transcoding and power efficiency, but it was a one trick pony in that regard.