But we do know what gpu-util measures, right? If gpu-util is 50% then that means that 50% of the time GPU does something and the other 50% of the time it does nothing. The later is guaranteed complete waist.

If I see that there is 50% gpu-util, I would rather spend time figuring out why half of the time gpu does nothing, rather than trying to optimize the other half where it does something.

Once we reach something close to 100% gpu-util, then it would make total sense to look at a better metric, such as tensor core utilization.