1. compute unit in CGN 1.3(RX 480) is more efficient the same in CGN 1.1 (R9 390) on 15% mostly because of the prefetch and cashing.
2. RX 480 has direct acces to ISA, but you should write you code on lower levels to take advantage from that
3. in games RX 480 is able to compensate brute force of the CGN1.1 by optimizations in geometry units like excluding the poligons of the zero sizes, because of the new index cash end etc.
So the question is can we get same +15% in math tasks if we use high level languages? I thinks its hard to achive if its possible at all. Cahes and prefecth are not able to give big advantage, and maybe only low level direct access can give something, but then you have to write special version just for RX470/480 cards.
But again, RX480 are very good from power point of view. I can take 190-195H/s with 0.95-0.975V voltages if i dont care about power costs, and 180H/s with 0.92V!,temps below 60C if I need power efficiency. Also you can use cheaper PSU for the rigs with RX480.
And again, I wouldn expect that RX480 will perfom close to 390X, till someone will start to develop special miner for RX480