ETH is absolutely bandwidth limited, achieving incredible efficiencies on many GPUs
If I am incorrect on that assumption then feel free to correct me. In answer to that all I can say is ETH may be Memory Bound on some cards but that doesn't mean it is globally bandwidth limited. If it is then link me to that proof please.
Ethash is entirely memory bandwidth and latency limited. That is why OC-ing the GPU does nothing for performance and why reducing CAS on the VRAM and increasing the memory clock speed is where all the performance improvements lie. The only way you are going to improve on it is with faster memory. And 4GB of QDR2 SRAM, even if it were possible, would be prohibitively expensive.
This is why the only available ethash "ASIC" yields no power efficiency improvement over a 1070. You _might_ be able to improve on it with a custom ASIC with 4GB of eDRAM, but that kind of silicon would also be obscenely expensive to produce, even for the current big players like Bitmain and Baikal.