Sorry, I was already irritated by someone else, I might have been a bit of an ass.
Reasoning being, 28nm doesn't have the performance scaling to double performance. Leakage is getting worse, this is why Intel switched to 3-d gates at 22nm, performance characteristics are getting worse and worse for each shrink, for traditionial processes. For example, from 4870->5870 doubled and in some cases more than doubled resources, more than doubled transistor count, from 953 million to 2.15 billion, and increased clocks. Yet performance scaling was only generally in the 60-70% range. If even if we expect kelper to be over 6 billion transistors, as massive or more massive than fermi was(which I don't) and thus double resources, its not going to get 100% scaling. Also, at best, they can only get ~50% more bandwidth unless they go back to 512 bit bus, which makes it even bigger, maybe GTX 280 size. Even that 50% is going cost them a big chunk fo die size for a better memory controller. Reportedly the memory controller from 4870 -5870 had to double in size to handle the faster clocks, also reportedly this is the reason for fermi memory controller not being able to scale up very well beyond 4ghz, because its not beefy enough.
I don't see any way they can make kepler beat a 590 unless its in some benchmark that previously had some odd bottleneck that they've fixed, ala Heaven bench. The hard info we have is that DP is "over 1 Tflop" on kelper, which while its not specific, it kind of gives us a ballpark. Thats only a 25-30% increase. If it was closer to 1.5, or so which would be close to double performance, they probably would have bragged up "over 1.5 Tflop" right? Seems to me that they aren't going for such a huge chip this time, and are trying more for efficency per die area.
Edit: On the "two different architectures" thing, reports are kelper is mostly a die shrink of fermi without too much new. /shrug