Been saying this for some time, it's not new.
Take a "standard" process drop and take a 200-250W chip and call it 400mm^2.
The goal for a GPU company with this chip is effectively two fold, use the new process to produce the same number of chips at half the size with as close to halving the power usage as possible and you have your new midrange chip. Then to use the pure power saving to produce a chip that has twice the number of transistors and as close to twice the performance at the same die size/power usage.
So you relatively speaking want 50% drop in power per transistor and as close to 2x the transistor density. one issue is not everything now is dictated by transistor size, in the old days the gaps between transistors were so big comparatively that every drop in transistor size and gap between transistors was fine. So to keep clean signals and to keep transistors cool now dropping the distance between transistors doesn't always work so some parts of chips scale very badly now with process drops compared to other bits. Memory controllers are said to be one of the things that scale badly.
Either way, when you go from a 50% power reduction to 25%, and this is on the low power process, not the high power which is less and less where these processes are tuned for. Then a 290x replacement with lets say only 15-20% more power, effectively means the same power chip would at twice the transistor density end up going from a 450mm^2 chip to a 300mm^2 250W TDP but only 15-20% faster chip. A 450mm^2 chip would be effectively 70-80% faster but use most likely 400W + which just isn't a viable product for most companies.
16nm brings with it a 40% power reduction....... I'm still unclear if the 40% power reduction is supposed to be from their own 20nm, which is brilliant, or in comparison to existing 28nm, which is much less good but still would be enough to be the difference between a meaningful new high end and a crap one.
I'm unsure if they'll try a 7970/680gtx replacement at 20nm, and wait for 16nm for the high end or if they'll skip it entirely.
20nm transistor density isn't bad, 1.9 increase is good, but this really has to be used alongside a huge power drop to make it worthwhile. Don't forget that 16nm isn't a fully new process, it is 16nm finfet's on top of the existing 20nm base metal layer, the equipment, fabs, everything is basically the same so there is no reason to believe in huge delays for 16nm. It should be in volume production a year after 20nm and we'll likely see significant benefit for high end gpu's.
I'd prefer, MUCH prefer them to switch to 16nm as early as possible with genuine 70-80% faster cards in a given segment, than have some overly expensive, painfully small improvement 20nm cards which effectively push back 16nm cards. Everyone that works on the 20nm cards won't be working on the 16nm project, with limited resources I hope Nvidia/AMD both focus on 16nm.
