How to measure the number of HBM bytes read/written for a kernel?

Developer Tools Nsight Systems Profiling Linux Targets
1

Is there a way to measure the number of HBM bytes read/written for a kernel? I am running my code on an H100.

Nsight systems has metrics for dram throughput

DRAM Read Bandwidth - dramc__read_throughput.avg.pct_of_peak_sustained_elapsed, dram__read_throughput.avg.pct_of_peak_sustained_elapsed

However, I was hoping to get a counter for the approximate number of bytes read/written from HBM rather than the percentage of the throughput. Perhaps the percentage of the throughput can be integrated to come up with a number, but I am not sure if this is the right thing to do given I assume that any throughput metric itself is averaged based on the sampling frequency. Is there perhaps a better counter (like total_bytes_read or something) or any other technique that you can suggest to measure the number of bytes read/written for a given kernel?

2

@pkovalenko can you respond to this?

3

Looks like nsight-compute has a dram__bytes_read.sum metric. That looks like the right thing to track.

4

As of now, we can’t display these values in bytes in Nsys and NCU in periodic sampling mode. NCU will only show this metric in kernel/range profiling mode.