Memory bandwidth in terms of SM number

[zhiqi-0]

Hi, really thanks for the nice tool!

Just some quick questions, I'm not familiar with underlying designs of GPUs, and curious about the used memory bandwidth in terms of SM used number.

Can we only use one SM to saturate the memory bandwidth theoretically? Or will there be some contentions if multiple SMs are involved for accessing data in global memory, comparing only using one SM?

I'm not sure if it's correct to test the memory bandwidth achieved by, e.g., only one SM, by changing the following code to const int TOTAL_BLOCKS=1;.

gpumembench/cachebench-cuda/cache_kernels.cu

Line 353 in 838a4fc

const int TOTAL_BLOCKS = BLOCKS_PER_SM * SM_count;