Can you use SR-IOV to provide GPU resources to two separate consumers?
Can you pass through an SR-IOV Virtual Function of the GPU to a VM?
The answer to 2. is “yes”. 1. is trickier.
Conceptually this is possible, and as I understand it Nvidia enables it for their professional GPUs. Intel is currently working on this for Linux, but the work is not complete as I understand it. I think laptop GPUs have this mostly working, but not desktop GPUs.
This is interesting… I just tested today on Unraid server this functionality (brought up by a user… as usual on Unraid ) and for a while it worked… for a while I had the same intel iGPU active on those 2 Win11 VMs at the same time… so I was kinda curious if this scenario is currently supported on truenas (not that it is super important but it would be a nice feature). So my usage would be quite simple… like while watching a youtube video inside the VM I won’t bog down the CPU… nothing related to video editing or stuff like that
But as you said @ericloewe it seems like it isn’t supported THAT well from GPU companies as well… so I guess I’ll stick with VM with cpu as gpu
From my looking into this, it appears that Nvidia actually has it working, but locks it behind licenses, but there are vGPU unlocks.
Meanwhile, Intel has switched from their vt-g extensions to just using SR-IOV, but not sure where that is up to, and integrated graphics pass through requires chipset AND bios support. Which a lot of Supermicro boards do not have.
Since all my servers 1) do not have integrated gpus and 2) don’t support the pass through anyway, I haven’t bothered pursuing that path further.
I’m looking into the Tesla/quadro nvidia thing currently for the same reason, I want to run multiple windows VMs, and without a gpu the rdp perf is not good enough
Yeah newer Intel IGPU use SR-IOV ( I think from 11th gen onwards) but this thing seems so new that I guess Truenas will take a long time before integration (IF it will ever be integrated)