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

[qarium]

I was kind of under the illusion Intel is actually ahead of AMD on sriov in their GPUs.

yes intel is ahead in put SR-IOV into the GPU on consumer cards but it is disinformation that you need this hardware feature to share 1 GPU with many virtual machines.

[qarium]

GPU SR-IOV is the holy grail everyone's waiting for
.. Sadly this news smells awfully like another feature for wealthy enterprise users

years ago i had a talk with bridgman and other people from AMD... about SR-IOV and they pointed out that this hardware feature does NOT do what people think it would do ...

even if your gaming GPU would have it it is not the feature you need to share 1 GPU between multible Virtual machines means you can not run windows 11 in a virtual machine on linux to play a game ,,,

years ago someone or even multible people did spread the disinformation that SR-IOV is the hardware feature who would allow you to do share 1 hardware GPU between multible virtual machines to run whatever game you want...

believe it or not to do so you do not need a hardware feature like SR-IOV at all you can all do it in software Vulkan/SPIR-V just like QEMU VirtIO-GPU means "VirGL is an OpenGL driver for VirtIO-GPU" and "Venus is an Vulkan driver for VirtIO-GPU,"

intel also spread the disinformation that you need SR-IOV and they put SR-IOV into the gpus to make you able to share 1 gpu with many virtual machines but if you look closer most of the parts that make this run is software not a hardware feature.

that you need SR-IOV to share 1GPU with many virtual machines is clearly disinformation.

[DanaG]

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!

[ezst036]

among people who don't want to dual-boot to game on Windows and don't want to have two separate graphics cards with one dedicated exclusively to a VM

This is the thing that has me wondering.

AMD and Nvidia are charging quite a lot for cards capable of SR-IOV, it is likely cheaper just to have a weaker card for regular uses and the beefy GPU for use in the gaming VM and when the feature goes mainstream worry about it at that point. Especially for someone who already has a card they upgraded up from a while back.

Just use two cards and why worry about it.

[ezst036]

I was kind of under the illusion Intel is actually ahead of AMD on sriov in their GPUs.

Probably their onboard, I am not aware that its in the Arc Alchemist PCIe cards.

[flischo]

It is strange that nobody mentions it and it seems to be something that few people know about.

AMDVLK supports SR-IOV virtualisation and works with VMware Workstation.

I've had it running on a virtualised Windows for a while now using a single 7900XT (the same as the host).

[CommunityMember]

their igpus had it. not sure if they still do. I think flex gpus have them too. IIRC there was a mod announced to get it working on the consumer dgpus but I dont think anything has materialized thus far.

My vague recollection is that for the dgpu's it required flashing enterprise (flex?) gpu firmware on the consumer cards by bypassing some version checks. I don't think the details as to how to do it were ever publicly released.

[ssokolow]

Exactly. Or they could simply forbid commercial use (on consumer hardware) in their licensing. That would easily make every big (cloud) company comply. But i doubt nvidia will ever learn to give consumers enterprise features for free.

Compared to AMD? What do you think reliable access to CUDA for things like Stable Diffusioning your own porn (the only actually-marketable use of A.I.) is?

[bernstein]

Yes there is a demand. They could probably limit the number of virtual gpus with the firmware on consumer cards. Most of the consumer demand people are likely to just want to run one vm at a time after all. Obviously that doesn't cover everyone, but most of consumers who'd be interested.

Exactly. Or they could simply forbid commercial use (on consumer hardware) in their licensing. That would easily make every big (cloud) company comply. But i doubt nvidia will ever learn to give consumers enterprise features for free.

[Quackdoc]

I mostly mean using it to run games, applications or Windows itself on desktop linux in a window with full gpu acceleration. I have a Windows vm that works, but is mostly useless with regards to anything that uses the gpu that I use from time to time.

Edit:

Looked over the patches as I hadn't had time earlier and it has a link to a youtube video and it has more info in the patch. Linux and windows guests will work and they have a test linux guest driver that works.

yeah that's the kind of stuff this is designed for. Cloud gaming uses nvidia + vgpu a lot.