Question regarding memory profiling

Developer Tools Nsight Compute
1

I am trying to profile the global memory access in some code.
While the collected metrics are great I do want to have a more in-depth data and wondered if it is possible

for instance I saw that a function took X amount of time and had a certain read and write bandwidth.
Is it possible to know how much of the run time was spent on the reads or writes to the memory? Can we know which happened first, or did they happen at the same time?
Or in other words, can I tell if the bandwidth access is always at peak value for some of the time or is it sustain for all the function’s runtime?

Also I saw that I can see the memory coalescence, are there other metrics that can indicate the addresses/strides accessed by the code

Thanks in advance

2

You can use PM sampling to see the values of certain metrics over time. For most chips, this will include DRAM read and write bandwidth. Collect the PmSampling section to enable this.

3

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.