As part of Microsoft Ignite, Microsoft announced the public preview of the NVv4 Azure Virtual Machines. Which is featuring AMD EPYC 7002 processors and virtualized Radeon MI25 GPU. Which will come in the following instance types:
Now the cool thing is that this is the first set of virtual machines using GPU partitioning on Hyper-V. Which is essentially Microsoft’s own vGPU which NVIDIA has available today on XenServer and on VMware. Which means that we can split up a GPU into virtual cards which we can then allocate to a set of virtual machines. Like with NVIDIA vGPU We have full support for DirectX 9 through 12, OpenGL 4.6, and Vulkan 1.1. But this is using together with AMD something they call AMD MxGPU which uses SR-IOV to do the mapping of the GPU inside the virtual machine.
What is the impact on this? from a cost perspective? I’ve been previosly working a lot on the Nv2/3 series. But since those models only do PCI-passtrough it means that I need to have a whole graphic card or only one core. But still might be a bit to much and combine that with the large instance sizes it is still some pretty large virtual machines.
Just some quick comparisons from a cost perspective if we compare with a regular NV3 series and D2v3 machines (that doesn’t have SSD but just to give some context)
Now from this simple check, we can see that the AMD based virtual machines are close to half the price of a D4v3 instance with almost the same specs but with a GPU attached to it. Also a LOT difference from NV6 (Promo) which is a model which is going out be replaced with a new versoin. The price is in preview so it might change, but who knows. (This is based out of west europe) but clearly in AMD advantage.
And just thinking about the increased CPU performance they will provide as well, this is just a sample benchmark compared with Intel so the AMD EPYC delivers quite a punch and also at a higher clock frequency in GHZ in comparison.
And note that 2nd Gen Intel Xeon Scaleable was released not so long ago, so not the previous version which the Azure virtual machine instances are running on.
When it comes to GPU partitioning is something Microsoft has been working on for some time, if we look back at a slide from the server conference last year
Now with this now available in Azure I also presumed that it was in the roadmap for Hyper-V and close to a release, but it seems that is s still part away.
So if you want to be part of the preview? sign up using this form https://aka.ms/NVv4Signup