I have a cool algorithm that I've designed while using my laptops M4000M(maxwell) graphics card. I'm looking to get a beefier desktop card to add more oomph to my calculations but find myself in a kind of a pickle as apparently, estimating a graphics card performance for purely computational purposes isn't as simple as counting up the TFLOPs anymore.

My algorithm is HEAVY on bitwise operations. And, or, xor, bitshifts are my bread and butter, making up around 90% of the instructions, 10% is integer math. Nvidia's 3000 series cards looked really nice up until I realized that its peak INT32 performance is only half of its FP32 potential, apparently because half the "cores" are dedicated FP32 while the other half can switch between FP32 and INT32.

Meanwhile the previous generation, turing can apparently do FP32 and INT32 simultaneously so their listed ops/s is actually half of their potential peak if both were to be used somehow. Going back further, afaik all units switched between FP32 and INT32 so the listed ops/s would be accurate no matter which numeric format was used.

My question is this: which of those units(FP32 or INT32) is in charge of bitwise operations in modern GPU architectures? My intuition would suggest INT32 but then again bitwise operations don't really have all that much to do with integer math so maybe both? It would really help me in figuring out what hardware to get.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.