Does anyone have experience/thoughts using multiple GPUs Vs. a single as relating to UX/interactivity improvements while in ‘raytraced’ viewport mode?
Does having a dedicated CUDA/OPTIX card and another for regular GPU tasks help with interactivity/workflow, or is it better to have one faster card? I’d like to prioritize interactivity/iteration speed in raytraced mode over absolute render speed for a single render.
I’ll be upgrading from an old 980Ti (so probably modern card will be an improvement), and my options are two 5060Ti 16GB or one 5070Ti 16GB (power limited in workstation).
Also, after searching this forum/google I see that multi gpu support was dropped for a while in the Rhino8 cycle…. would this still be a threat moving forward?
I don’t believe that ever happened, short of some bug. No one serious about their rendering wants to be limited to one device.
People have said that. I never saw any evidence of it. (Experience: once had FOUR 1080ti’s in one machine, so I could actually test this, it didn’t matter. Maybe with older crappier cards it was more visible.)
There is added overhead moving to multiple GPUs that means it’s better to max out a single GPU first. The 5070ti has basically double the CUDA cores of the 5060ti.
One thing to make sure to consider is not just the multi-GPU setup, but what speed those PCI-e lanes run at on the motherboard. Overpopulating lanes if you have a regular board, may cause some throttling in constant viewport refreshes. If you have a more serious workstation board (EATX/Threadripper), then this should not be a problem for a pair of GPUs. My MSI MPG Edge Z790 runs fine with both of my cards at the same time.
For individual renders, it won’t really ever make any difference, but if viewport speed and interactivity (constant updating) is required, then this may become more of a problem.
Yes, I also recall the multi-GPU support was a problem, but it has been fixed for ages.
You may want to consider (if you have the time) waiting for the Nvidia Super series uplifts of the current generation 50 Series cards.
I waited for the 4070 Super, and I think it was hands down the best value of the entire Ada generation cards for the relative performance. But basically anything compared to your 980Ti will be blisteringly fast if you get to the 5060 and above.
Yes, thanks for fixing it. I guess the question was more along the lines of how volatile multi gpu support may be going forward. Is it a ‘solved problem’ now, or are fixes still required every release of Cuda/Cycles/etc…?
Of course, I realize nobody can predict the future 100%, especially of a complex tech stack like nvidia+cuda+cycles+rhino.