Is it possible, practical or even effective to mix a Titan V or RTX and a Tesla V100 in a single system? With the Titan performing graphics duties and the Tesla acceleration and compute tasks? (And if so, how about 2+2) Both being Volta, is it possible to link 2 or more in a NV Link to allow full speed communication? I don’t do much ML/DL yet, though I will be before long. But I do a lot of graphic design and graphics-intensive work from 3D (soon to add VR) on down to intense gaming sessions with my son. If not, could you recommend a setup? Thanks.
{Edited} Please see below.
So, please allow me to clarify, now that I have found some more information in independent study. I am still curious as to the same questions, just a more likely pairing if my case scenario is possible at all. Can I have a Quadro GV100 and a Titan RTX in the same case? In NVLink/SLI? And if not, what about a Quadro RTX?
Yes, but why. You can NVLink 2 GV100s together. There are techniques with PyTorch that will use two different gpus. I think programs like Blender can do it as well.
Generally, one can mix a gfx and a compute GPU. Both need to be enterprise GPUs, If you mix consumer and enterprise GPUs, some enterprise features will not be enabled…
You can also use to gfx GPUs, and have one of them be assigned for compute…
Gfx GPUs do compute just well, but FP64 precision…
NVlink is ALWAYS ONLY possible between identical GPUs!
–>> SO: if using multiple GPUs, always try to use the exact same GPUs, so you have no driver, compatibility, feature differenc problem, and hence affinity issues…
IF you need one GPU to NOT do compute, so its exclusivly used for gfx, to keep a high and stable fps, one can use CudaVisibleDevices and HIDE this gfx GPU for compute. And flipped, a compute GPU that does NOT need to do ANY gfx, can be switched (in Windows) into TCC mode, so some of MSFTs gfx ‘rules’ (like TDR) do not apply…