AMD MxGPU aims to give GRID a run for its money

AMD's approach to GPU virtualization is substantially different from NVIDIA's, and could make for an interesting competition down the road.

Earlier this year, AMD released their answer to NVIDIA’s GRID vGPU after what seems like a long wait. Their platform, dubbed MxGPU (for Multiuser GPU), is a hardware-based offering that does all the virtualization on the card itself and has no licenses to buy.

Since AMD does the virtualization in the hardware via SR-IOV (as opposed to NVIDIA, which relies on the driver to do the slicing of GPU resources), users get direct access to the GPU along with the benefit of being able to use the native driver to support the new hardware. OpenGL, OpenCL, and DirectX are all supported. AMD claims that since they’re doing more work in the hardware, MxGPU could perform better, at least in theory, than GRID, but I’d need to see data before I made that call. (It’s one of the things I’m hoping to find out at BriForum during one of our graphics sessions)

The cards themselves are available through server OEMs right now, namely HP, Dell, and SuperMico. There are three models:

  • S7150 ($2,399) – Maximum of 16 users per card, 2048 Stream Processors, 8GB GDDR5 VRAM, 150W power consumption.
  • S7150 x2 ($3,999) – Basically twice the resources of the S7150. 4096 Stream Processors, 16GB GDDR5 VRAM, to support 32 users per card. Power draw is a little less than double the S7150 at 265W.
  • S7100X ($1,925) – An MXM card that is essentially the S7150 in terms of specs. It was just released this week (and I can’t find a link for it!)

The maximum number of users per card is a physical cap related to the amount of VRAM dedicated to each user. AMD’s guidelines for the S7150 card are that it can be split between 2-6 users with Design or Engineering use cases, 6-10 users with advanced use cases that consume high-end graphics but aren’t creating them, or up to 16 users that are typical knowledge workers. Of course your mileage might vary, but those numbers are a good place to start.

The cards currently only support VDI environments built upon vSphere/ESXi 6.x, though XenServer support is in the works. Linux interest is at an all-time high these days, so AMD is also considering adding support for the KVM hypervisor and client OSes.

I’m not about to jump out and say one product is better than another, especially without seeing the comparison data, but I can say that I’m thrilled that there are now options in virtual graphics for desktop virtualization. This competition between AMD and NVIDIA will surely result in more capabilities over the next year, and we’ll likely see costs driven down to where there’s no question about whether or not it’s worth it to add this to your environment.

Join the conversation


Send me notifications when other members comment.

Please create a username to comment.

Since Grid uses timeslicing for compute, it means that the entire power of the gpu is available to all users, where as I understand the AMD has a fixed amount of GPU power. Since most CAD apps need high compute capacity in short peaks, it means that for that kind of workload Grid will enable more power to the task. Running benchmarks does not give you the right picture as they does not reflect how real users work (high peaks, long breaks, vs constant high load). Another thing is that most CAD apps depend on Quadro functionality and will not perform as well on a non Quadro GPU. Grid licensing enables full Quadro functionality for vGPU. OpenCL and Cuda works with the largest vGPU profiles, if you need this functionality you will most likely run on a large profile. So you cannot compare these two techologies as they were the same. For other use cases with taskworkers and knowledge workers needing basic GPU, it's more comparable.
Hi Gabe,
I think the description of the time-slicing is a bit misleading. We have a hardware scheduler built into the GPU. The job of the hardware scheduler is to pull work from the active channels, and deliver it to the GPU.
There’s no direct NVIDIA driver involvement there.
The only difference between SRIOV and GRID is how the channels are set up. Once running both are full hardware acceleration.
Best wishes,
Rachel (from NVIDIA GRID)
Richard Hoffman did a good overview on how NVIDIA GRID is hardware not scheduled in software:
They also need to work on Windows 10 support as the cards only handle Windows 7 and 8.1.

I've been using the AMD FirePro S9050 with XenServer 6.5 and XenApp 7.8 on Dell PowerEdge r730's and they work pretty well. Managed to recently add AutoCAD 2016. I didn't know I had to disable the virtual GPU installed by XenServer and reboot the VM to get Direct3D and OpenGL to work properly.
Hi All,

Yes, I'd like to echo Rachel's comment that the virtualization/scheduling of the NVIDIA GPU occurs within the hardware. See this old post where Jason Southern confirms.

Here's my understanding of how the AMD card carves up the GPU resources. Instead of time-slicing all of the GPU cores as NVIDIA does, AMD allocates a portion of the GPU cores to each machine. So if there are 1000 GPU cores and you have 10 virtual machines, each machine gets 100 cores. Alternatively, NVIDIA would give each machine access to all 1000 cores for a time-slice.

I'll also echo what Magnar said. The idea behind NVIDIA's method is that real-world users look at a CAD model (for instance) for a bit, think about what they are going to do and then rotate/zoom, and repeat. They aren't constantly rotating the model like a benchmark does. These very spikey (real-world) GPU and CPU usages benefit NVIDIA's design to time slice all the cores.

I do look forward to seeing some performance comparisons between the products though.

Thanks for the article, Gabe.

-Richard Hoffman