Changing the GPU Virtualization Game

Today NVIDIA announced the NVIDIA GRID August 2017 release (AKA GRID 5.0), it is a major change for the virtualization industry. Over the next few days I will be publishing a few blogs about the NVIDIA GRID vPC (Virtual PC) and Quadro vDWS (Virtual Data Center Workstation). This blog covers some of what’s new with GRID 5.0 and soon I will publish some details on testing I did using an NVIDIA P4 GPU with GRID 5.0. I will also have additional blogs that go live during VMworld, around GRID vPC and Quadro vDWS. These will be part of the material I cover in my vBrownBag Tech Talk on vGPU Enabled Linux VDI (VMTN6636U)

Now let’s get into the details of whats new…

The first thing that’s worth noting is that NVIDIA GRID vPC is really a software solution for datacenters. You have NVIDIA Tesla (datacenter) GPUs and then you have NVIDIA GRID software (August 2017 release as of this writing), those two combined provide a powerful datacenter solution for Virtual Desktops (VDI). This means you have the ability to get more out of your NVIDIA GPUs and can continue to use them as new GRID software is released.

That gets us to the first big advance with GRID 5.0. It is now possible to use both Maxwell and Pascal GPUs in your virtual environments. This means continued support for legacy Maxwell (M60, M6, and M10) GPUs as well as support for Pascal (P40, P6, P4, and P100) GPUs. Thus allowing companies to maintain their previous investments longer.

Now what could I mean by maintaining investments longer? I buy a GPU and what its capable of is based on the hardware, right? Not exactly… When Citrix announced the ability to monitor GPUs from within Citrix Director (https://www.citrix.com/blogs/2017/06/01/monitoring-of-nvidia-gpus-in-citrix-director/) that was a capability delivered by GRID software. Thus as NVIDIA adds more monitoring functionality with newer versions of GRID, virtualization vendors, 3^rd party app developers, and even you can take advantage of them to monitor GPUs. If a new monitoring feature is added and your support contract is current you can upgrade your GRID software to use the new feature. That’s right the second big advancement, in my opinion, with GRID 5.0 is decoupling hardware and software functionality. NVIDIA is providing the ability to have continual new features even on older GPUs.

You might be thinking this is great! I’ll just go in and download GRID 5 before my license expires and I’m all set! Not so fast, NVIDIA did a fair amount of work on GRID licensing. With GRID 5.0 you have to have valid licenses. If you don’t performance is degraded. I’ve even tested it on Linux based desktops and even Linux desktops have their GPU performance capped if not licensed properly.

One enhancement delivered with the power of both GRID 5.0 and Pascal cards is the ability to run CUDA based workloads across the vGPU profiles. In other words you can build VMs with various vGPU configurations that run CUDA based code such as some Adobe applications, ArcGIS, AutoCAD, etc. You might be thinking this was possible before, and you’re right sort of. You had to use a full GPU profile like the M60-8A. With the Pascal architecture NVIDIA added preemption capabilities to the GPU. This is a fancy way of saying some applications with CUDA capabilities can share the same GPU. Possibly even bringing HPC, ML, DL, and AI into the virtual realm, which I’ll dig into in another blog.

The next cool thing that gets unlocked with GRID 5.0 is actually a feature of the Pascal cards and not a GRID feature. For those who have setup previous generations of GPUs (Maxwell) for EUC/VDI environments, you are probably familiar with having to use the gpumodeswitch command to change a Maxwell GRID GPU from compute to graphics mode. For those not familiar with it, you had to install gpumodeswitch, make the change, reboot, uninstall gpumodeswitch, reboot, install the VIB, reboot, and finally everything is in place. Guess what, you don’t have to do that any longer. The Pascal GPUs are able to switch between compute and graphics without gpumodeswitch. This reduces the reboots and steps required to have a GPU up and running. I’ll cover this during my vBrownbag TechTalk at VMworld and its accompanying blog.

It should be noted that the removal of gpumodeswitch is a feature of the Pascal cards and not something that’s part of GRID 5.0. This means if you are using the M10, M6, or M60 GPU you will still need to use gpumodeswtich before you can use GRID 5.0.

You may have noticed above I mentioned an NVIDIA GPU that you may not have heard of. The NVIDIA Pascal P6 GPU. NVIDIA announced the P6 GPU along with GRID 5.0, this is a GPU specifically for blade server deployments. With deployments that leverage blade technologies it’s possible to provide the latest GPU technology and maintain a consistent operational model in the datacenter.  Hence, if you use rack mount server there are GPUs for it, if you prefer blade servers, there are GPUs for it.

At this point I bet you’re wondering what the various vGPU profiles for the Pascal GPUs are. In the tables below I’ve assembled a list of the vGPU profiles for the P40, P6, and P4 GPUs.

As you can see NVIDIA GRID 5.0 and the Pascal GPUs provide a lot of power for your virtualized GPU needs. If you are interested in discovering more about GRID 5.0 be sure to visit the NVIDIA GRID site on NVIDIA.com. In my next blog I’ll cover some of the testing I’ve had a chance to do with the NVIDIA P4 GPU and GRID 5.0.