Streaming multiprocessors and threads on GPU Understanding about Streaming-Multiprocessors, Threads,

[tera]

“Keeping dimBlock.x a multiple of 16 often gives you good coalesced memory access patterns.”

I believe this should be 32 on Fermi devices since they are not accessed by half-warps but by whole warps.

That’s true. I tried to keep it simple and thought it’s not worth mentioning Fermi here as the cache will help mitigate bad memory access patterns there. However I’ll edit the post accordingly.