I doubt Bitmain has a "smaller" varient on the S7 in the works.
A 2 string per board design would run more like 270 watts per board, which is what many of us EXPECTED the S7 to be - 2 strings 2 boards like the S5 but 18 chips per string instead of 15.
The plan to create and sell an S7+ is obvious, given the S5+ was apparently intended in part as a "prototype" style design using their older gen chips.
Trivial design to do this time around, the design work appears to have already been done.
It seems that Innosilicon's A3 will be in the same ballpark as the BM1385 - probably about .2 watts / GH at the chip level .26 at the system level, based on what I've seen out of them and Lketc - none of those figures are definite before working hardware shows up though and I might be misinterpreting the figures I HAVE seen.
It's hard to understand some of those websites when I don't read Chinese.
I am starting to doubt we'll see an S8 design, Bitmain seems to have gotten tired of dealing with "internal power supply" issues, and the S5+ sold out fast enough despite the widespread assumption (since demonstrated to be almost definitely TRUE based on the S7 announced sale availability date) that the S7 was probably already in production for internal Bitmain/Hashnest usage when the S5+ was announced.
I also suspect there will be a U4 as an attempt to fill the "home miner / quiet miner" niche.
I'm not betting on it being a pod design, but I'm not betting against it either.
I would NOT be shocked at something close to the form factor of that 100ish GH Rockerbox model, Bitmain could probably fit a single-string miner into that size board.
Looking forward there are no huge performance gains
14/16nm full custom should be able to get at least double the efficiency of 28nm full custom.
Possibly more, though it appears that quantum effects are making the gains noticeably smaller than JUST the size of the gates would indicate (in theory, if nothing else caused issues, half the size of the gate features should equal half the voltage and a quarter the power, but quantum effects haven't allowed that for a few generations).
THAT is likely to be the last "big gain" for a few years though, as that will bring Cryptocoin mining chips up to the "state of the art".
I suspect Bitmain will skip anyting other than "full custom" on their announced "working on" next generation chip.
I also suspect that chip won't show up before the halfing next year, and might not arrive before end of the year timeframe in 2016.