NVIDIA Publishes Open-Source Linux Driver Code For GPU Virtualization "vGPU" Support

[Radtraveller]

sorry, reading a bunch lf stuff about sharing vgpu among vms.. while not always necessary it seems most sites I have read want sr-iov enabled.. example https://pve.proxmox.com/wiki/NVIDIA_vGPU_on_Proxmox_VE

[dragon321]

there is a large demand for just running windows 11 in a virtual machine on linux and run games in it who are incompatible with running the same game in Proton natively on linux.

Most games that don't run on Proton are incompatible with Linux because of kernel anti cheat and VM is not gonna solve that issue, unless you want to try to hide VM and risk ban.

[Quackdoc]

sorry, reading a bunch lf stuff about sharing vgpu among vms.. while not always necessary it seems most sites I have read want sr-iov enabled.. example https://pve.proxmox.com/wiki/NVIDIA_vGPU_on_Proxmox_VE

SRIOV is needed to "split" the gpu up and pass it to VMs. Stuff like virgl, venus, gfxstream, forward vulkan/gallium to the host. Virtio native context passes through the kernel driver kinda, but requires mesa to be built for it, and is not a reliable option.

[Nille]

because of kernel anti cheat and VM is not gonna solve that issue, unless you want to try to hide VM and risk ban.

That is the Important part. Its great to improve compatibility but you should not to try games with AC in a VM or make it obviously that its a vm that the AC can decide to not let you play it.

But its still great if you has games that are not running on proton due to other reasons. for example many codex cracked games are not running on proton. they simply wont run.

[muncrief]

I don't know if it's really SR-IOV. But with AMDVLK (on top of Mesa), VMware allows you to enable 3D acceleration, with just Mesa it doesn't work.

There is some performance loss but it is more than enough to be able to use some CAD type applications in a 4K screen.
The GPU usage in the benchmark is high, so the VM is using it.
PwxTYyy.png

Can someone please buy me a replacement GPU for my RX 580? I just discovered amdvlk doesn't work with it so I can't compare its performance under Windows to vulkan-radeon. I have an RX 6600 dedicated to my Windows 10 VM but it would be a real hassle to juggle things around and use it for my Linux GPU, and it would be much easier if one of the very rich people on this forum would just buy me another GPU.

Thank you

[Quackdoc]

Can someone please buy me a replacement GPU for my RX 580? I just discovered amdvlk doesn't work with it so I can't compare its performance under Windows to vulkan-radeon. I have an RX 6600 dedicated to my Windows 10 VM but it would be a real hassle to juggle things around and use it for my Linux GPU, and it would be much easier if one of the very rich people on this forum would just buy me another GPU.

Thank you

if this isn't a mood and a half I don't know what is lmao

[muncrief]

if this isn't a mood and a half I don't know what is lmao

Ha!

Sometimes I'm just silly Quackdoc .

And try to break up the day with a bit of it

[Jabberwocky]

No! No! This doesn’t conform to my narrow view that nvidia is bad and isn’t genuine about going open source!

Please shill somewhere else. No company that has the market share that Nvidia does is genuine about going open source.

This will not go into consumer products. It will only go into overpriced enterprise compute.

There have been real open source projects to get this working on consumer Nvidia GPUs but Nvidia noticed it and blocked open source workarounds in newer GPUs.

https://github.com/VGPU-Community-Drivers/vGPU-Unlock-patcher

[Jabberwocky]

sorry, reading a bunch lf stuff about sharing vgpu among vms.. while not always necessary it seems most sites I have read want sr-iov enabled.. example https://pve.proxmox.com/wiki/NVIDIA_vGPU_on_Proxmox_VE

You're right. The Proxmox guide you linked requires SR-IOV and it's not something you typically see on consumer cards. There's a workaround that I liked in my previous comment for cards like Geforce 1080 or Geforce 2080, some older Tesla cards too. If you get the workaround applied and you're running the correct kernel (because Nvidia is so genuine about open source) then you still need a license from Nvidia in order to run your guest VMs without restrictions. The restrictions here: https://docs.nvidia.com/vgpu/latest/...grid-licensing

Someone I know used vgpu and I tested it briefly on his system, back then he had a 1070 TI. It seem to work quite well, the advantage of going this route is that you can use a Linux host with Windows or Linux guests while still having access to your single GPU in your host (perhaps for transcoding in Proxmox container etc.) The disadvantage is that if you have a Geforce 3000 or 4000 you are out of luck.

The other options are VFIO. I've been doing that since ~2013, I still have my OG Intel i7 3770 (non-K). This method works well if you have a lot of experience and more than one GPU. You can get it working with a single GPU but it's very cumbersome depending on the specific hardware that you have. It's advised to buy parts specifically for this setup. See https://gitlab.com/Luxuride/VFIO-Win...ll-walkthrough

Another option becoming more popular is GPU-P which is Microsoft's GPU partitioning via paravirtualization, yes many P's. I've tried this with Windows 10 Hyper-V and a Windows 10 guest VM. It works flawlessly and seems to support a wide range of hardware. For this test I used a Radeon 7800 XT. I ran Pytorch, Frostpunk 2 and did AV1 hardware encoding inside the VM. It worked better than anything I tested before, with the only drawback being that you need to use Windows... For host and guest oef. See https://forum.level1techs.com/t/2-ga...-hyperv/172234

Sometime in the future we might have Venus working flawlessly in Linux and Windows guests with a Linux host. I'm looking forward to that day.

[Martyn]

Another option becoming more popular is GPU-P which is Microsoft's GPU partitioning via paravirtualization, yes many P's.

WSL2 supports that now and they forced NVIDIA to make GPU-P work for Linux under pain of not having their drivers WDDM 3.0 certified (necessary to leverage new graphics features under Windows 11). Linux kernel devs also require FOSS userland tools to enable testing as a pre-requisite for mainlining kernel modules (Microsoft got stung by this on LKML multiple times, as did EMC for VMWare) so something tells me NVIDIA will be forced to at least support a single VM form of vGPU support without those issues or face rejection.