There are a few threads in the last few months where people much smarter than I am discussed the theoretical maximums on some of the algorithms based on memory speeds, memory bandwidth, etc.  We are much closer to the maximum with public miners than you might believe.

The fact that you see GPU core utilization low just means that the particular algo isn't GPU bound, but is likely bound by memory bandwidth and access times.  Each algo is different and may stress different portions of the GPU Ethereum for example is not very GPU core intensive, but very memory intensive, which is one of the reasons the cards with HBM can achieve faster hash rates at lower memory clocks and why manipulating memory straps have such a large impact on performance.

On the flip side, ZCash is very gpu core intensive, but not very memory intensive, so you can actually underclock your memory and have little impact on your hashing speed.

Monero seems to be a blend of both getting best results with a high gpu core speed and a high memory clock with modified memory straps.