The performance/efficiency tradeoff in ethash is becoming less significant with newer hardware.  With a R9 290 card it was very difficult to maximize performance without increasing power consumption.  With Polaris, where performance is mainly limited by memory and not the core clock, it is much easier to get near peak performance with minimal power.

That said, I think there is room for some tangible improvements in ethash (and probably cryptonight) mining efficiency through optimizing the mining kernel.  For example a Rx 580 doesn't need active 36 CUs to do 30Mh/s; it's closer to 1 CU per Mh.