How cluster influence GEMM or other application?

Accelerated Computing CUDA CUDA Programming and Performance
1

I tested cutlass’s cluster, well… it has some influence, but not too much… what I can find is TMA’s multicast is within a cluster. Any other factor we want to use cluster? For… GEMM or other application?

2

What do you mean with it has some influence? Influence on what? Compared to what? What did you expect, what was the result? Speed? Which platform?

1 Like
3

Such as you are writing cutlass code with sm90, you set cluster_shape=<1, 2, 1>, and then you change it into <1, 4, 1>. You will have different running time. Why?

Or my question can be, for GEMM, what optimization can cluster do? Given that, cutlass does not implement SM-to-SM communication, and NCU shows the DSMEM usage is 0. Why cluster influence performance?