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

[bernstein]

Heck, even just making PCIe reset perfectly reliable would already be a massive improvement over the way things worked (or rather, the way thing failed) the last few times I tried PCIe passthrough with AMD GPUs!

Pcie reset has been working flawlessly on my rx6700xt for years…

[Jabberwocky]

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.

Very interesting.

Just by chance I found this random gist thread: https://gist.github.com/krzys-h/e2de...316dfb794f4d6a

I really want to try this out, but working 16 hours a day at the moment.

[Quackdoc]

did you EVER meet a person in this forum who had such a amd radeon pro v620 card ?

yes...

[qarium]

then AMD should stop advertising V series cards for cloud gaming, but they do advertise it for such, because thats what they are good for

did you EVER meet a person in this forum who had such a amd radeon pro v620 card ?

i clearly have never seen such a person.

https://www.amd.com/de/products/accelerators/radeon-pro/amd-radeon-pro-v620.html

if i search such hardware on price search engines i never find any of such hardware.

geizhals.de nothing google product search nothing..

if you see current product offerin: https://www.amd.com/de/products/grap...kstations.html

there is not a single amd radeon pro v card

this means all current workstation PRO cards have no SR-IOV

its pretty sure only the MI cards have SR-IOV right now... and they do not do graphic at all.

[Quackdoc]

now i have a unapproved post to..

you want to look here: https://www.phoronix.com/forums/foru...-drivers/page5

as you see SR-IOV is not what people think it is...

then AMD should stop advertising V series cards for cloud gaming, but they do advertise it for such, because thats what they are good for

[qarium]

beep boop unapproved

now i have a unapproved post to..

you want to look here: https://www.phoronix.com/forums/foru...-drivers/page5

SR-IOV is only available on V-series (e.g., https://www.amd.com/en/products/serv...adeon-pro-v620) and MI-series (e.g., https://www.amd.com/en/graphics/inst...r-accelerators) boards.
SR-IOV does not work the way most people probably think it does. Something like virtio would make more sense for end users IMHO. See https://www.phoronix.com/forums/foru...40#post1422640

I think there is a misconception that SR-IOV will magically make your single GPU show up as multiple GPUs and everything just works transparently. That is not the case. You need explicit support for SR-IOV in the hypervisor and guest drivers to make this all work. You could make SR-IOV the basis for a consumer-targeted solution, but you could also use a number of other things (virtio, etc.).
MS already has a paravirtualization solution for Windows host + Windows guest and Windows host + Linux guest. For Linux host + Linux guest you can use virtio-gpu and mesa. So the only thing missing is Linux host + Windows guest.

as you see SR-IOV is not what people think it is...

[qarium]

beep boop unapproved

just keep in mind the feature what people want is AMD MxGPU

https://pve.proxmox.com/wiki/MxGPU_with_AMD_S7150_under_Proxmox_VE_5.x

https://docs.omnissa.com/de-DE/bundle/Desktops-and-Applications-in-HorizonV2312/page/PreparingforAMDMxGPUforWindowsFullCloneVMs.html

you find the old forum threat with more infos about that here:

https://www.phoronix.com/forums/forum/linux-graphics-x-org-drivers/open-source-amd-linux/1421838-amd-radeon-pro-w7700-launches-as-999-gpu-with-fully-open-source-upstream-linux-drivers/page5

and there are the AMD people who talk about this... see;

SR-IOV is only available on V-series (e.g., https://www.amd.com/en/products/serv...adeon-pro-v620) and MI-series (e.g., https://www.amd.com/en/graphics/inst...r-accelerators) boards.
SR-IOV does not work the way most people probably think it does. Something like virtio would make more sense for end users IMHO. See https://www.phoronix.com/forums/foru...40#post1422640

I think there is a misconception that SR-IOV will magically make your single GPU show up as multiple GPUs and everything just works transparently. That is not the case. You need explicit support for SR-IOV in the hypervisor and guest drivers to make this all work. You could make SR-IOV the basis for a consumer-targeted solution, but you could also use a number of other things (virtio, etc.).
MS already has a paravirtualization solution for Windows host + Windows guest and Windows host + Linux guest. For Linux host + Linux guest you can use virtio-gpu and mesa. So the only thing missing is Linux host + Windows guest.

as you see SR-IOV is not what people think it is...

[qarium]

beep boop unapproved

by the way i told the amd people to add SR-IOV to every consumer hardware...

they say the transistor count penalty is to high and the overhead of a SPIR-V software solution like colabora venus for only 1-2 VM vGPU is so low that it is not worth it.

the transistor count penalty is only worth it if you are really data centr and you really run 32 instances per gpu...

you can say over and over again that intel and nvidia add SR-IOV to their chips as well but they all disable it for cheap consumer cards via closed source microcode firmware.

also keep in mind as soon as intel has more success in the GPu space they would try the same tactics to get more money from professional workstation customers. the start of the ARC intel gpu series is a abnormality what will not repeat itself.

just see it from the other side means complete different perspective if we have a SPIR-V software solution like colabora venus solution that works well this would cause all the professional workstation users to buy cheap desktop hardware instead of the expensive workstation hardware ... more or less this would be the only way to pressure AMD/Nvidia/intel to add SR-IOV to every consumer gpu to get better performance in VM vGPU benchmarks.

[qarium]

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.

most desktop users would be happy with only 1 single VM... with vGPu support.
but really why is this not possible now and why in hell was it not possible in the past
for me is really a pain in the ass.

you know what is the biggest joke ... Microsoft target children with their wrongdoing.

for example if a children want to play minecraft on linux with a controller means not with keyboard and mouse the java version of minecraft does not support anything else than keyboard and mouse on linux.

on windows 11 the minecraft version in the appstore is the dx12 native windows version with full controller support. linux people could play this version with Proton/DXVK but you can not download and install the dx12 windows version you can only download and use the java version on linux.

its really bitter that children who use linux gets bullied by microsoft.

to run windows 11 in a VM with cGPu support would also make the directX12 native windows version of minecraft run.

[Quackdoc]

beep boop unapproved