Nvidia :

https://docs.nvidia.com/grid/gpus-supported-by-vgpu.html
https://www.nvidia.com/pt-br/data-center/graphics-cards-for-virtualization/
https://docs.nvidia.com/grid/13.0/product-support-matrix/index.html

AMD :
Introducing the AMD FirePro(TM) S7100X, the Industry's First and Only Hardware-Virtualized GPU for Blade Servers

https://www.semiaccurate.com/2018/08/26/amd-v-series-virtualization-gpus-launch-with-the-v340/

https://techreport.com/news/29666/firepro-s7100-graphics-cards-bring-hardware-gpu-virtualization-to-life

https://web.archive.org/web/20160507042500/http://www.tomsitpro.com/articles/amd-firepro-s1750-firepro-s7150x2-hardware-virtualization,1-3129.html

https://www.amd.com/pt/graphics/workstation-virtual-graphics

https://www.amd.com/system/files/documents/gpu-consistency-security-whitepaper.pdf

https://www.techtarget.com/searchvirtualdesktop/opinion/Comparing-AMD-vs-Nvidia-for-virtual-desktop-GPU-cards

https://www.reddit.com/r/VFIO/comments/nqf749/gpu_splitting_on_consumer_amd_gpus_vgpu_mxgpu/

https://github.com/DualCoder/vgpu_unlock - here there is a hack to make consumer GPUs able to provide vGPU

So, coming from all those links, we would have to do :

1. Add entries to the docs or do a separate wiki regarding this subject
2. Design a "test bed" with some guidance so people can test their setups and GPUs
3. Design some methods (and which tools could be used) to perform benchmarks and find possible bottlenecks regarding driver, synchronization, pass-through issues etc.

We also would have to dig into specs of each vendor's GPU families and series, to find anything that could drag us to a more "precise" judgment on which ones does provide support for vGPU.

@olivierlambert Could we create a GPU dedicated section on docs, and perhaps provide some GPU HCL - Hardware Compatibility List - regarding vGPU and who knows what else stuff is relevant ?

Am I missing something here ?

Use case: jellyfin/plex transcoding, Simple Windows VM workstation, etc.

Other cards with the capabilities from AMD are designed for workstations primarily. The workstations based Radeon Pro W6000 Series has these technologies. It would enable the capability for vGPU as it can be used for VDI.

Edit: If you can get a hold of a card from NVIDIA or a card using a chip from this list (https://www.nvidia.com/en-us/data-center/graphics-cards-for-virtualization/) then you can also use it as a vGPU.

So there would likely be possibly dropped frames and/or other issues both graphical and technical.

It would also limit what can be done in virtual machines, on hosts with it. Though it would have its uses, just bear in mind the performance cost.

If your going to go low cost for VDI then the use of Radeon Pro cards can get hardware video acceleration. Additionally there's open source drivers, for this acceleration so missing parts can be developed by community and driver can be forked if necessary. This would occur through the pass through. Though Radeon Pros are for workstations, I don't know how well the hardware acceleration works on the equivalent data centre graphics cards.

Iam running 1xE1230v5 on SuperMicro Server (X11SSL-F). Only got one PCI-E 3.0 x8 free for use. Would mind its not worth to get it a try on this machine.

vGPU is very interesting at all and would be a huge part to take over other virtualization solutions.

Has anyone managed to get XCP-ng 8.1 working with AMD S7150 MxGPU?

Yes. For me It works fine

Has anyone managed to get XCP-ng 8.1 working with AMD S7150 MxGPU?
I got it working on 7.6 with the 1.05 AMD driver but 8.1 with the new 2.0 AMD driver doesn't give any virtual functions.
Does someone have a working setup with XCP-ng 8.1 + AMD S7150(x2) ?

Maybe the real question is more on AMD side of things.

In other-words, if we spend $$$ to continue building on FirePro MxGPU platform, are we eventually going to get stuck without ability to upgrade?

Huge thanks for all your work regarding XCP-ng. I recently started using it and deploying it in production. Open Source technology really helps SMBs in my part of the world (Transylvania, Romania). XCP-ng has been great so far.

Regarding the nVidia Grid K1/K2 cards, I believe there is a huge business opportunity here. I don't have all the information but here is my idea:

nVidia Gamestream technology (video codec, remote protocol) is a dream. It enables 4K 60 fps streaming at very low latency. Unfortunately it relies on nVidia GTX cards and is only doable with GPU passthrough. If the Grid cards would be able to emulate this kind of behavior, it would be a game changer for the VDI industry in my opinion. For this to work, I believe the VM and the OS on it would need to be able to use the GeForce experience software and "think" it's running a proper GeForce GTX card.

Maybe I'm greaming big here, but think of this:
There is a great opensource software called MoonLight. It is an opensource port of the proprietary nVidia GameStream protocol. It enables small single board computers like the Raspberry Pi to act as a high-end thin client. This is working today, but with single hardware passed-through GPUs of the GeForce GTX kind. If this could be done with the Grid cards, enabling multiple VMs to utilize GameStream as if they would have a GeForce card, this would mean a total disruption of the current VDI market. Grid cards are as cheap as $170-$200.

Unfortunately there are some closed source software in the mix, the biggest question being the ability of the GameStream technology to work with Grid cards. nVidia has it's own game streaming cloud solution and I am almost certain they are using Grid cards with their GameStream tech. The only question is how can it be done.

Citrix HDX is pretty expensive, vmWare Horizon with it's own high-end graphics protocol is also expensive. If it would work with XCP-ng, the HighEnd workstation VM market would have a new player with a disruptive solution.