N00b CUDA specific qestions.


I had some CUDA specific questions.

I’ve been looking at a lot of benchmarks and I’m confused. For instance, are the amount of CUDA cores the most important factor on GPU specific number crunching (BOINC, Password cracking, etc) tasks? Or is it the clock speed of the card in general?

Would a lower (older) spec. card with more CUDA cores out perform a newer “faster” card with less CUDA cores?


It depends on a number of factors. If your kernels are compute bound, speed will be roughly proportional to the number of cores times the shader frequency. If they are memory bound, speed will usually be about proportional to the memory bandwith (which is proportional to the bus width times memory clock), although in some cases the memory latency will also be important (I don’t know though if there are significant latency differences between the cards, I guess not).
In general, you will likely want newer cards for their new features like double precision, larger register set, improved atomics. Fermi will bring many more like caching etc.