I’ve previously written about the new upcoming NVv4 series virtual machine instances which are the new series of GPU based instances in Microsoft Azure which are now coming with AMD Radeon based GPU’s (https://msandbu.org/introducing-the-nvv4-azure-virtual-machines/) these new instances are the first coming with support for true GPU partitioning, which is a of somewhat similiar technology to what NVIDIA is doing with their vGPU solution. So essentially the hypervisor is slicing the GPU into different profiles which are then attached to the virtual machines.
This capability in Azure is provided using a technology called MxGPU, which is built on top of SR-IOV (Single-Root I/O Virtualization) and something that AMD has provided for XenServer and VMware ESXi for some time already, this is however the first time GPU sharing has been enabled for a Hyper-V based platform with Azure. Unlike vGPU from NVIDIA this does not require any custom drivers in the hypervisor layer, but only on the guest VM level and is done using true hardware based GPU partitions. This essentially means that the drivers on the guest VM handles access to the physical GPU using SR-IOV.
The AMD MxGPU feature exposes all graphics functionality of the GPU to the Virtual function at its PCI-E partition allowing for full support for graphics APIs like DirectX and OpenGL but also GPU compute APIs like OpenCL. This means that Code written in these standards for the physical device need not be adapted or altered to function in the virtual environment as it operates as a regular physical GPU.
Now I finally got access to the NVv4 Preivew, which of now is only available in US South Central datacenter region, but will be expanded upon to other regions later (West Europe coming soon)
As it is with most solutions that requires a driver to be installed is typically available as an extension, but since this is in preview the current way is to download the customized drivers and install which can be found here –> https://download.microsoft.com/download/7/e/5/7e558ac0-3fff-413d-af62-800285a2fc53/Radeon-Pro-Software-for-Enterprise-19.Q4.1-Technical-Preview.exe
And as mentioned before, it provides full capabilities for DirectX and OpenGL.
And not so shabby performance on the lowest instance type which provides 2 GB of video memory on the MI25 card, which provides a total of 16 GB of memory of the physical card.
Running the AMD EPYC CPU running at 2,5 GHZ Core speed, which is a bit higher compared to the Dv3 series that Microsoft has in Azure running on the Intel Platform.
Unreal engine benchmark.
Full size here –> https://i.imgur.com/7soVUsX.png
Now looking at the pricing perspective of NVv4 compared to the most common instance types for Dv3 series. At the current price setup the Nvv4 is a lot cheaper compared to the D4v3 series running at almost the same specs, but a bit lower memory. It is as of now, 56% of the cost of the D4v3 series, might not be the case when the final price is out, but seems like it Microsoft can sqeeze out more compute for a lower cost. The way I see it, this is going to be the recommended instance type for most VDI projects moving forward, regardless if it is WVD, Citrix or VMware.
But it should be noted that the Nv4 series is currently in preview, so it might take some time before the final price details as published.
It should also be noted that the Nvv4 compared to the Dv3 are equipped with Intel® Xeon® 8171M 2.1GHz (Skylake), Intel® Xeon® E5-2673 v4 2.3 GHz (Broadwell), or the Intel® Xeon® E5-2673 v3 2.4 GHz (Haswell) processors in a hyper-threaded configuration.