[some-natalie]

That's 100% fair. Good thing it's not too difficult to assign VFIO w/i QEMU for virtual machines despite the manufacturer shenanigans. :)

The Arch wiki has a great guide here - https://wiki.archlinux.org/index.php/PCI_passthrough_via_OVM...

It does get a little tricky if your GPUs are identical, but I've done this for years and maintain a guide for doing this (as well as the ACS-override patched kernel RPMs) for Fedora.

- Writeup - https://some-natalie.dev/blog/fedora-acs-override/

- Code + RPMs - https://github.com/some-natalie/fedora-acs-override

As far as concerns around stability with ACS override, I tend to only enable the override for the specific GPU (or other hardware) that I'm passing through and haven't encountered any stability problems or memory leaks that'd interrupt desktop or light server usage. I also used to run this for a bunch of white-box GPU hardware for a customer at a former job and it worked well for exploratory AI/ML workloads before investing in the big Nvidia DGX boxes. YMMV, of course!

[galangalalgol]

Is there a reason an emulator like qemu couldn't just pass along spirv like wasm does with webgpu? Would that be way slower?

[vGPU]

A better way [for me, anyways] is getting GRID drivers running. However, this only works with the 9xx cards up to the 2080.

https://gitlab.com/polloloco/vgpu-proxmox

[throwaway698585]

It's not difficult but it misses the point. SIOV supports 1k's of VF's because that's what you need if you want a sandboxed app-per-VM security model. When statically compiled VM's are just as performant as containers but more secure.

[pm2222]

Why is this a requirement? I thought Linux takes over after it loads.

  Your guest GPU ROM must support UEFI"