Hi, is the warp size (32 theads) going to change in the next couple of years ? Any indications by NVIDIA people here ? I know one shouldn’t rely on the fact that the warp size is 32 but it didn’t change for the last couple of years so maybe it’s just ‘a good compromise’ (e.g. bigger warp size would mean typically higher warp divergence, smaller warp size would mean that SIMD advantages are decreasing).
I think of it like this. If you want to tune your code for a current gpu like fermi, you can get perf benefits from baking in assumptions about the warp size. If you want your code to be portable, make it correct for all warp sizes.