We will see, the graphic card market is very complex one.
2900xt which is a great great grandad of 6970, and has a same aritechture. When it came out, it was hot, inefficient, relatively slow and expensive as well as delayed.
It took 4-5 generations to get the architecture perfected.
Nvidia is at that stage now with it's fermi. Femi as architecture has much more life than current Ati version. Expect, Kepler to move things on massively forward.
"To understand the future u must know the past"
Fermi is NOT massively new, to say so is utterly silly, yes the arrangement isn't the same as the 8800, but the base shaders mostly are. Thats the point of a mostly fairly simple 1 for 1 shader architecture.
Gf100 is way closer to 8800 than the 6970 is to 2900xt. The 6970 has a different front end, different memory bus, different internal communication(radically different) different shaders, and seemingly different tesselation unit.
A cayman doesn't come close to performing like a x number of shaders 2900xt card, at all, in any way. 8800 128 shaders, 280gtx, 240shaders(I keep hearing it was supposed to be 256shaders but was basically too big, I assume it used to be more 16 shader clusters rather) and 256 would have been the natural progression. The 480gtx is 512shaders, again a natural progression.
The whole time memory bus has been fairly standard, internal communication has been fairly standard, rops/tmu's have been fairly standard, ratio's have been kinda standard and performance scaling has been exactly where you expect. Double shaders(and several other things) get circa 80% more performance. This is EXACTLY what you expect for a very similar architecture doubling in size every generation.
The 2900xt was big because it was designed for 65nm, and was pushed up to 80nm as 65nm was VERY late, simple as that, it was a huge core, because it should have been on a MUCH smaller node, thats all, it would likely have had significantly more shaders aswell to make more use of that ringbus architecture.
The 5870 bears very little relation to the 2900xt, the 6970 bears no relation to it now.
Fermi is NOT the new and inefficient design that will get better with new process's. Architecture + process don't "suit" one another, its a simple case of too big = problems making it, this "too big" size has been pretty much the same for generations but also shrinking generation by generation.
Keep in mind that 2900xt was HUGE, on the wrong process and ate power, and yet yields still beat Gf100 quite easily, 2900xt was also almost top to bottom a completely new architecture to previously, on top of having to be pushed back a process, on top of having many dx10.1 features(and the original dx10 spec features) included that Nvidia/MS teamed together to have removed. Fermi has literally none of that going for it in terms of excuses, 32nm was never going to be out last year, Fermi is the exact architecture you expected from a double shader count 280gtx with generational improvements, it doesn't show any significant architectural efficiency over 280gtx(except in DP).
Fermi is as "new" an architecture as 280gtx was to the 8800, some, but expected with no massive jump in efficiency.
The 5870 was exactly the same, its "new" its got dx11, its got stuff the 4870 didn't have, but it performed EXACTLY where you thought a double sized 4870 was, there was little to no efficiency increase, it had similar shaders and front end. 4870/5870 are two cards you would call the same architecture, the 2900xt, the 3870, the 4870/5870, the 6870 add the 6970 are all "different" architectures with pretty fundamental changes.
