How to know *real* GPU utilization?

I know nvidia-smi can show GPU utilization, but it does not seem to reflect real utilization. For example, even if it says 100% utilization, it seems only to mean all cores are activated, but may only use a small portion of the theoretical FLOPs.

It would be great to have something to provide a number representing how full the cores are actually utilized.

I guess nsight may be useful, but not sure whether people use it for this usage.

Thanks in advance for any suggestions!