[maldev]

No, it uses the hypervisor to implement a scheduler for the GPU between the different vm's and gives them access to it. VFIO on linux requires passing the entire GPU to only a single VM and making it so the host doesn't have access.

[sekh60]

Sorry again, is this like nVidia MIG? That seems to be the more current technology. Not sure if the host can retain a slice though.

[maldev]

Kind of, i'm not to familiar with the enterprise gpu sharing, but I believe that requires hardware support for it on the GPU to split up contexts for the OS. Whereas the GPU-P solution works on any GPU and is vendor agnostic as it's done on the OS/Processor level. It's really cool and I have no idea how they can ship this since it does trample on the enterprise tech pretty substantially. But I belive it's a byproduct of Microsoft wanting to compete on the datacenter level and it gives them a killer feature for Azure, and as a byproduct it entered as a semi undocumented feature into consumer hyper-v.

[sekh60]

Thanks for the knowledge share!

MIG does need hardware suppoet, and from what I understand really expensive licensing. I don't know if AMD has a parallel technology.

That is pretty cool that it is vendor agnostic. I've found a few docs from a few years ago talking about stuff like it on Linux, but development of it seems to have stopped or just not progressed at all.

But she is they would use it in Azure, I imagine there they are using enterprise GPUs. But in Andheri scale datacenters it definitely sounds like an advantage!