High density GPU options?

[Najtsob]

What is the best option if you want high density GPU support ? 

I was thinking about vSGA which based on this article (https://blogs.vmware.com/performance/2020/01/vmware-vsga-for-content-rich-vdi.html) seems to be good enough for our use case.

But when I was looking at nvidia documentation it seems that vsga is no loger supported on new nvidia cards. What is then the alternative, AMD cards ? 

P.S. Is VSGA even good option to begin with ? 

[5teelman]

As always - "it depends" - but I can share our setup and experience:

Started off with Nvidia K1 in 2016, worked fine with Win7 VM's - and no licensing nightmare either! We're only using vDGA, easy to setup and upgrade whenever needed. Drivers now support vMotion with GPU as well, works impressively smooth.

We have had a number of Dell 2U servers with 3x Nvidia M10 running our density load for years, Windows 10 clients on Horizon 7.13 with a mix of thin clients and laptops. Normal VDI with 2x full HD monitors for office workers. Needs licenses for each VM.

One M10 card has 4x GPU and 32GB memory, hence we could in theory run 96 VM's on each server with a 1GB GPU profile. Our experience is that rather using 2GB profiles and max out at half the amount of VM's is much better. You'll soon run out of CPU and/or memory. A GPU enabled VM locks up all memory it is provided, and it still uses CPU as not all applications offload graphics to GPU.

We also have Dell 2U servers with 6x Nvidia T4 cards (1GPU, 16GB each), used with 4GB profiles for e.g. video editing. This setup yields a max of 24 VM's with a 4GB CUDA enabled workstation profile. Though at a rather high cost, as the WS license from Nvidia is not cheap. Some "challenges":

• High cost, Nvidia licensing and cards. Looking forward to testing Intel's new initiative (https://www.intel.com/content/www/us/en/newsroom/news/introducing-intel-data-center-gpu-flex-series.... - apparently without license.
• Nvidia licensing server hasn't been that stable up until this year, and handling the licenses is messy. The least "lease-time" before freeing the license is 10 minutes.
• The general user uses much more web based video than before - kills performance sometimes, even with GPU in the VM.

The current Nvidia A16 is the natural choice for density in my opinion, but take a look at Intel and AMD solutions as well. Licensing and licensing servers is Nvidia's biggest disadvantage in my opinion.

I also have to give Nvidia support kudos, they are brilliant and responsive like no-one else I've worked with!

[smut5203]

This is very much helpful. Thanks for sharing

[Najtsob]

Thank you for the first hand answer. 

I know that nvidia licensing is pain, but AMD or Intel GPUs seems to be totally unknown option and I would rather stick with something that is more widely deployed.