Bank conflicts in shared memory still a thing?

Is it true that compute capability 3.5 GPUs, (in my case the k40), do not experience bank conflicts like in the older generation GPUs? I.e threads can access the same bank at the same time without having to wait in a queue?



Yes, it’s still a thing. :)

Shared memory on sm_3x and sm_50 devices has 32 banks.

Bank conflicts are still a thing on all architectures. They are a fundamental attribute of highly banked memories that are accessed concurrently at arbitrary indices.

Access scheduling in the presence of bank conflicts should have improved from sm_35->sm_50.

Thanks allan and Greg