Utilization

Has anyone come across the term “Core Utilization” ? I got the definition of this term as - how much each core is utilized in a program run. I checked all metrics / events provided by nvvp, nvprof - I could not find such metric.
I also checked CUPTI - as expected, it has all metric / events same as nvvp. And I believe Vampire also uses CUPTI for performance, so I assume Vampire will not provide such metric if I am not wrong.

So where did you come across this term? Context would be nice.

In a casual technical discussion…

I see nothing wrong with the concept as defined in the original question, in particular not in a casual technical discussion. However, since “CUDA core” is really just a marketing term, one might want to replace that with “ALU” or another suitable technical term. I assume the discussion centered around idling computational resources due to thread masking, control transfers, or memory hierarchy activity.

I’ve heard lots of people casually throw around the term “utilization” over the years without really knowing what that means or how complex a concept it can actually be…

In your case you could look at at the number of floating point / integer operations being performed over a given period of time, both for addressing and compute ops, and compare that to their theoretical maximum number of OPS over that period of time.

Due note that there are both “CUDA cores” / “Streaming processors” / “FPU:s” aswell as SFU (special function units) that can perform computations.

Anyhow there are plenty of nvprof profiling option flags for this.

The event called “active_cycles” may be related to “core utilization”. In CUPTI, the definition of “active_cycles” is: number of cycles a multiprocessor has at least one active warp.