I read somewhere that Nvidia are purposely capping the performance of their gaming cards when used for “scientific” type calculations for this next 6xx generation. I can’t remember the link, but check out this forum discussion: http://www.tomshardware.co.uk/forum/347529-15-560ti-cuda . Apparently the 580 cards are better to use if you can’t afford a Quadro…
The GeForce GTX 680 trails AMD’s Radeon HD 7900 cards in 32-bit math.
And it gets absolutely decimated in 64-bit floating-point operations, as Nvidia purposely protects its profitable professional graphics business by artificially capping perfrmance.
Vlado, do you think the GTX 680 will get better / un-capped performance when used with CUDA instead of OpenCL? And is Phoenix using CUDA too at this point?
On the topic of video cards, is there one that is noticably the best for Vray RT? I know that several other GPU renderers say that game cards work better than Quadros, is that the case with Vray RT? are there benchmarks anywhere? How about GTX 680’s in SLI? or the GTX 690? It is surprisingly hard to find good info.
Also, in vray RT, when you are using CPU and GPU simultaneously, what is better, to have more slower cores or fewer faster cores (assuming they add up to the same speed). I am looking to build a beast machine that is specifically for doing lots of GPU rendering and I am curious what is the best way to go for this. Let’s pretend money is no object.
Gaming cards do work better than Quadros, but there is a risk of overheating if you are not careful, especially if you stuff several cards in the same machine. Good cooling in that case is essential. In terms of performance, the 680 card is roughly similar in performance to 580, but consumes less power. You don’t need SLI to use multiple GPUs for rendering.
If you have good GPUs, I would recommend using the GPU only and not hybrid CPU/GPU - this might end up being slower.
CUDA tends to be somewhat more reliable, and there is no need to go through a compilation phase like with OpenCL. Performance-wise they are very close as the GPU code compilers are very similar after all. However for future versions there might be things that we support on CUDA only.