Warp efficiency

In the metrics definition [1], I see

warp_execution_efficiency smsp__thread_inst_executed_per_inst_executed.ratio

warp_nonpred_execution_efficiency smsp__thread_inst_executed_per_inst_executed.pct

Based on my understanding from the CUPTI manual, the metric definition is the same for both name. However, the difference is in the way that this metric is reported. For example, if the maximum number of executed warps is 32 and the theoretical value of the number of supported warp is 32, then

X.pct = 100%
X.ratio = 32

However, as you can see two names are assigned, pred and nonpred. Is that correct? I guess it is a bit confusing.

[1] https://docs.nvidia.com/nsight-compute/NsightComputeCli/index.html#nvprof-metric-comparison

1 Like