GPU passthrough

[nywst]

Hi there,

Say, I have 2 VMs under Proxmox: win11 and Jellyfin. Can I passthrough 1 GPU to both VMs? If not, what's the solution please?
Thanks.

 

[leesteken]

Say, I have 2 VMs under Proxmox: win11 and Jellyfin. Can I passthrough 1 GPU to both VMs? If not, what's the solution please?

PCI(e) passthrough is always only to one VM at a time. If you have a GPU that supports SR-IOV (expensive enterprise feature) or mediated device passthrough (integrated Intel or expensive enterprise GPU), then you can share it over several VMs (but you typically cannot have output to a physical display). There is also VirGL that let's OpenGL inside a VM use a GPU on the Proxmox host (but also no output and only OpenGL).
It's not clear to me what you want to achieve, so I think you'll need two GPUs (and a motherboard that supports enough passthrough).

 

[nywst]

PCI(e) passthrough is always only to one VM at a time. If you have a GPU that supports SR-IOV (expensive enterprise feature) or mediated device passthrough (integrated Intel or expensive enterprise GPU), then you can share it over several VMs (but you typically cannot have output to a physical display). There is also VirGL that let's OpenGL inside a VM use a GPU on the Proxmox host (but also no output and only OpenGL).
It's not clear to me what you want to achieve, so I think you'll need two GPUs (and a motherboard that supports enough passthrough).

I need hardware decoding in Jellyfin vm and remote video editing in win11 vm, but seems unlikely based on current setup. My GPU is just 1070 or 5700xt.
Thanks.

 

[leesteken]

I need hardware decoding in Jellyfin vm and remote video editing in win11 vm, but seems unlikely based on current setup. My GPU is just 1070 or 5700xt.

At least you don't need output to a physical device, so Intel integrated graphics might be good enough. If you have only one GPU (which does not support SR-IOV or partitioning or mediated sub-devices) then just don't run both VMs at the same time.

 

[PBoerst]

This is possible with the vgpu unlock script. I'm working on it right now and I actually have a support post for it right now, not officially supported but Craft computing has a guide. I have it mostly working up until launching the VM.
You can find more info here https://youtu.be/jTXPMcBqoi8?si=ulnBFTRbZJsdiyaS

 

[Dunuin]

This is possible with the vgpu unlock script. I'm working on it right now and I actually have a support post for it right now, not officially supported but Craft computing has a guide. I have it mostly working up until launching the VM.
You can find more info here https://youtu.be/jTXPMcBqoi8?si=ulnBFTRbZJsdiyaS

From what I've read so far:
1.) kernel upgrades tend to break it so not great if you rely on your VMs
2.) a very hacky approach
3.) it's at least a legal greyzone to use vgpu without a license which you won't get as a consumer nor for a GTX/RTX that isn`t officially able to use vgpu

 

[PBoerst]

From what I've read so far:
1.) kernel upgrades tend to break it so not great if you rely on your VMs
2.) a very hacky approach
3.) it's at least a legal greyzone to use vgpu without a license which you won't get as a consumer nor for a GTX/RTX that isn`t officially able to use vgpu

1. They never said their vm's were production use so I didn't assume they were 100% uptime required
2. not that hacky imo.
3. Nvidia offers 90 day trials. No grey area there to use it.

 

[Dunuin]

3. Nvidia offers 90 day trials. No grey area there to use it.

Probably for supported enterprise GPUs, not for GTX/RTX that need hacked drivers to enable that feature, as Nvidia don't wanted you to use it with consumer GPUs. And I guess everyone wants to use those VMs for more than 90 days