Does Nvidia Profiler report bandwidth in metric (SI) units?
I could not find any information regarding the units if they are metric (power of 10) or binary (power of 2) in Nvidia Profiler…
However the Memory Statistics report are in MiB https://docs.nvidia.com/cuda/profiler-users-guide/index.html#memory-chart … but I am not sure if the units are binary elsewhere, too…
For Nvidia Nsight Compute it is documented that the default is metric (SI) units https://docs.nvidia.com/nsight-compute/NsightCompute/index.html#profiler-report-metrics
I assume nvprof reports binary units… i.e.
gld_requested_throughput Requested Global Load Throughput 70.411GB/s 70.411GB/s 70.411GB/s
gst_requested_throughput Requested Global Store Throughput 70.411GB/s 70.411GB/s 70.411GB/s
However, can someone confirm that nvprof and nvvp use binary based units?
Thanks in advance.