Does pytorch’s profiler support memory bandwidth profiling? I only found memory “allocation” profiling in the docs. I want to determine whether or not my program is bounded by memory bandwidth. Are there any tools to do that?
I’m commenting on this in the hopes that someone more informed than I am might respond, but I had the same question. I was wondering if it would be valid to determine the memory bandwidth of a particular operation by simply dividing the memory allocated by the average time reported by the profiler?