what is the mean of `gpu__compute_memory_access_throughput`

Developer Tools Nsight Compute
1

in the --section MemoryWorkloadAnalysis,

Metrics {
    Label: "Max Bandwidth"
    Name: "gpu__compute_memory_request_utilization_pct"
    Options {
      Name: "gpu__compute_memory_request_throughput.avg.pct_of_peak_sustained_elapsed"
      Filter {
        MinArch: TU10x
      }
    }
  }

what is the means of gpu__compute_memory_access_throughput
and how to get memory load/store efficiency using nv-nsight-cu-cli for sm_75
thanks in advance

2

In the section file you pasted, there is no metric gpu__compute_memory_access_throughput, did you put the wrong name by any chance? In general, you should be able to get descriptions for almost all metrics using the command line query functionality, e.g. for SM 75, RTX2080Ti

nv-nsight-cu-cli --query-metrics --chip tu102

This will print a (long) list of all metric base names with their descriptions. You can check Nsight Compute CLI :: Nsight Compute Documentation for more details on how to use that option, and how to query the valid suffixes for those metric names on Volta or newer architectures.

If you already know the metric (base) name to query, e.g. gpu__compute_memory_access_throughput, you can also use directly

nv-nsight-cu-cli --query-metrics --chip tu102 --metrics gpu__compute_memory_access_throughput

For the metrics you listed, the descriptions are

gpu__compute_memory_access_throughput: the average Compute Memory Pipeline : throughput of internal activity within caches and DRAM, as a % of peak burst rate over active cycles
gpu__compute_memory_request_throughput: the average Compute Memory Pipeline : throughput of interconnects between SM<->Caches<->DRAM, as a % of peak burst rate over active cycles

Note that their are currently no direct mappings for most nvprof *_efficiency metrics in Nsight Compute (see also Nsight Compute CLI :: Nsight Compute Documentation )

3

thanks for your detailed answer.
by the way, how can i get the descriptions of other metrics? and, can i get *_efficiency by the combination of some metrics?

4

thanks for your detailed answer.
by the way, how can i get the descriptions of other metrics? and, can i get *_efficiency by the combination of some metrics?

5

how can i get the descriptions of other metrics

nv-nsight-cu-cli --query-metrics