Because the architecture is RADICALLY different, they are moving from scalar to non scalar, from VLIW 4 where EVERY single instruction has to be coded perfectly to get every single shader full, and it gets no where near that, to an architecture where there is NO VLIW, 2900xt through 6970 is VLIW architecture, it infact requires the scheduler and the code to be pretty much perfectly in sync, it never, in any game averages 100% ultilisation, the best games approach 3.4(ish)/4, the worst games somewhere closer to 2.5/4.
GCN again has NO Vliw at all, its basically per shader, like Nvidia, 1d shaders not 5d or 4d, its the biggest change in architecture in GPU's in since unified shaders architecture with 8800/2900.
Theoretical fill rate previously wasn't close to average, Nvidia's theoretical fill rate is very very close to their average, AMD's new theoretical fill rate should also be VERy close to the average.
This is an improvement from the per shader efficiency alone, you're talking about realistically in the majority of games a 33%+ increase in efficiency.
Think about it like this, if 2 cards one with each architecture had 1600 shaders, one card would average using 1500-1600 shaders each clock, and the VLIW card would average anywhere between 1000 and 1400, and the majority of the time that will fall in the middle of that range, around 1200.
This is before you think about things like 6870, 1120 shaders performing above a 1440 shader card and close to a 1600 shader card on many occasions due to a lot of tweaking. The actual shaders themselves being pretty different and capable of new things, tweaked front end including many of the tricks they learnt in the 6870 which wouldn't have made it into the 6970 and lastly, VLIW 4 relied on driver based scheduling, GCN does this in hardware likely far more efficiently. The shader count on GCN/Cayman is as comparable as the 7900/8800gtx comparison, or the x1900xt/2900xt comparison...... as in not at all.
