H200 NVL vGPU on Proxmox

[shbaek]

I have a Proxmox server equipped with four NVIDIA H200 NVL GPUs.
Is it possible to allocate GPU resources to individual VMs in a vGPU-like manner?
If so, are there any documented methods or successful real-world cases of this configuration?

 

[dcsapak]

if you want to use 'vgpu' style gpu with a H200 you have to use nvidia ai enterprise https://www.nvidia.com/en-us/data-center/products/ai-enterprise/
which incurs an extra licensing fee to nvidia (idk the exact details currently, but in the past it was a certain amount of money per year per physical card)

sadly this is currently not officially supported on proxmox, though technically it probably works since it's underlying technology should be pretty similar to regular 'vGPU' usage for other datacenter cards for which proxmox ve is officially supported.

 

[shbaek]

if you want to use 'vgpu' style gpu with a H200 you have to use nvidia ai enterprise https://www.nvidia.com/en-us/data-center/products/ai-enterprise/
which incurs an extra licensing fee to nvidia (idk the exact details currently, but in the past it was a certain amount of money per year per physical card)

sadly this is currently not officially supported on proxmox, though technically it probably works since it's underlying technology should be pretty similar to regular 'vGPU' usage for other datacenter cards for which proxmox ve is officially supported.

Thank you. I’ll test it, and if it works successfully, I’ll share the information through the thread.

 

[dcsapak]

if you test it and it works, and you want it to be officially supported, it wouldn't hurt to tell nvidia that you want proxmox as supported platform

 

[spirit]

if you test it and it works, and you want it to be officially supported, it wouldn't hurt to tell nvidia that you want proxmox as supported platform

Hi dominik, I'm going to test with a H200 soon.

have asked to nvidia support for ai enterprise support (and also about MIG mode not working with windows vms)

 

[dcsapak]

have asked to nvidia support for ai enterprise support (and also about MIG mode not working with windows vms)

well 'proper' vgpus with qemu will only ever be time-sliced and not directly the mig devices (because they're just 'virtual' hardware partitioning)

i currently have access to an rtx pro 6000 blackwell card, which supports timesliced vgpus on top of mig instances, and that should simply work (but it's only available on that card currently)