vSGA / vDGA Hybrid Configuration

[hammond]

Hi there,

I'm dealing with a customer who wants to provide hardware acceleration to both Windows and Linux desktops from a single server with a single graphics card. My immediate suggestion was that we use the latest Tesla M60 to run everything in a vGPU configuration which, assuming a compatible Linux distro, will be able to provide acceleration to both flavours of operating system.

However, a more cost effective solution might be to install an older GRID K1 card in to the unit, and using the vDGA for Linux desktops features (present since 6.1 I believe) I could pass through a GPU directly in to each Linux desktop. However, lets say for example I have 4 Windows desktops and 2 Linux desktops on a server. The K1 card has 4 GPUs, so could I pass through 2 of the GPUs in to the 2 Linux desktops, and then share the other 2 GPUs between the 4 Windows GPUs in a vSGA configuration?

I guess in summary my question is - if I install the vSGA software VIB driver in to the ESXi host, will I still be able to do PCI pass through on the card, or will the hypervisor remove those devices from the list of available ones?

Many thanks in advance.

[showard1 · Accepted solution]

Hi

Couple of points:

1.     Be really careful with the K1s - those GPUs are very low performance.  If you're wanting acceleration for something like Aero desktops / transparent windows and stuff, sure, they're fine.  If you're wanting to do games or serious 3d stuff, you will likely be disappointed.  In that case, I'd definitely recommend the M60

2.     You can do VDGA and VGPU or VSGA on the same host.  Meaning VGPU and VSGA are mutually exclusive.  Other than that, you can split them up however you want.  Pass through the first 2 gpus directly and have the other 2 do VSGA or VGPU.. 

3.     I have seen almost no difference between VDGA and VGPU when running practical 3d tests.  I don't see any real reason to bother with VDGA over VGPU.  That said, VSGA is a definite performance hit, but depending on what you're doing, may be tolerable.  At least you can vmotion with VSGA.

[hammond]

Thanks for the response Sean. The M60 and vGPU for all desktops is my preferred approach but the cost is possibly prohibitive, as it will mean licencing the desktops to adhere to the GRID 2.0 software model, and Linux support only comes with the middle tier licence which is very expensive. For my use the K1 should be adequate (I've used them before, along with K2s in the past) and the desktops won't need licencing individually, but as vDGA is the only option on those cards for Linux I needed to be sure I could split the architecture to avoid purchasing multiple GPUs/hosts and degrading the cost benefit..

What you've told me means I should be able to go ahead with my proposed design and save considerable cost. It's a prototype trial anyway, if it moves ahead there would be better argument and funding to go with vGPU on M60s and futureproof it all.

Many thanks Sean.