I don't think it'll happen ever. The problem is not only in memory per se but in memory bandwidth. Fast memory = expensive memory. And that is what makes a good ETH miner, not the speed of SHA rounds. So I don't expect any breakthrough in ETH ASIC mining. Somebody can try though, I might be totally wrong.

Think "stripped down GPU with SHA3 hardware units instead of general-purpose compute units". Won't be cheap though, nothing near Ant Miners.