How to Check GDDR (DRAM) Access Time

I am curious about the method to measure the GDDR (DRAM) access time of a GPU in Pytorch-implemented NeRF related code. Specifically, the runtime breakdown of a certain task is quite lengthy, and I want to determine whether this is due to the GDDR (DRAM) access time or the computation time.

To resolve this, I used Nvidia Nsight System and marked the method I wanted to check with nvtx to view the profiling results. However, I am unsure how to extract results related to DRAM access.

Additionally, I have used Pytorch Profiling to check things like Self CUDA Time and CUDA Time, but I am not certain if this was the correct approach.

The environment I am using includes an RTX3090 and Linux.

I am curious about this.

If you take a look at , I think you will be able to use that methodology to determine where you are compute or memory transfer bound.