GTX 280 - 576mm^2 die size, 65nm process
GTX 480 - 529mm^2 die size, 40nm process
Basically all the additional performance came from shrinking the process. This doesn't happen very often.
It happens whenever there is a process change, the architecture seemingly has naff-all to do with it.
The process plays the largest part for overall performance gains, but it is architecture that makes the crucial difference. All being equal with NVidia and AMD running on the same process, it is the architecture than seperates them (die size also helps).
Looking at refreshes within the same process such as 5870 vs 6970 indicates what benefits architecture can have. AMD made some basic changes to shader layouts and gained 20-30% performance in some areas for little additional transistor count.
Now, this is where the 7900's are a little disappointing. Looking at the basics, the 7970 has 33% more shaders, 50% more memory bandwidth, and almost twice as many shaders as the 6970. All of this extra muscle only provides ~35-40% real world clock for clock gains. I would agree that in the case of the 7900, all of the extra performance comes from the die shrink, and that very few architectual gains have been made over the 6900's (atleast very few that make a difference within games).
It is almost as if AMD have implemented Intel's tick-tock approach whereby every other generation focusues on either die-shrink, or architecrure, with just a tiny overhang of the two.
If AMD simply die-shrunk and factored-up the specs of it's greatest ever card- the 9700Pro to a 4bn transistor monster, it would be utterly crap by todays standards. Architecture is vital and makes a world of difference. Unfortunately it does not evolve as quickly as die-shrinks which are the easy way to obtain gains. Just take AMD vs Intel in the CPU world, where Intel has AMD whipped for architecture. On the same fab process Intel are literally years ahead of AMD. AMD latest processors can only just about beat Intel's 3 year old i7 socket 1366 parts which were made on a much larger process.
Below is a simple graph showing the accumulated affects of 0%, 5% and 10% architectual gains per generation. These gains may be small compared to an average 40% gain aquired through shrinkage, but they certainly add up. This is where AMD have lost out to Intel since the days of the Athlon XP's.
