That's not the whole story though is it? The 5800 series was released on a poor, leaky and relatively unknown process and we don't know how conservative ATi were. Personally I think they were quite conservative given their focus on getting the product out the door.
Given improvements in the 40nm process and more experience and data it's very likely they will have a decent bump in transistor density, though clearly this won't be 25%.
But really my point is that the performance gains of the 6000 series will come from a matrix of factors: increased transitor density, increased die size and new core that is better able to feed the shaders. This makes it difficult to guess how fast it will be, but you can be pretty sure that it will be a significant improvement or they wouldn't bother.
Hmm, if you read my replies, I've said a 25% bump in transistor number is pretty much exactly what I expect, however to double performance in most gens(comparing efficient well made cards, vs equally well made cards) you're roughly looking to double transistor count. The 5870 is exactly twice as fast as the 4890 in Metro 2033, with actually just over twice the transistor count, though some is down to the dx11 move(but amd with dx10.1 was most of the way there to be honest).
2.5-2.7billion would I think be semi comftable for AMD, but would be moving into borderline bad yields/big die and too expensive. There will be a little leeway in increase in process quality, but no where near as big as some people seem to think. Efficiency increases I'd be surprised if they got beyond 10%. the 5870 isn't massively shader limited, but the 5870 in some games is still 10% faster thana 5850 at the same clocks.
If you make the rop/tmu part of the core more efficient and gain 10% speed, then the card will want more shaders to be fed by them, and if you bump up Rop numbers, you'll want more tmu/shaders, basically its a pretty well balanced core in general.
THe biggest issue really is that different games show different limits, thats where the difficulty comes in. A 5870 is almost identical to a doubling of a 4890, and is exactly twice as fast in Metro 2033, and in other less stressful games, its only 60% ahead, it probably averages about 70-80% ahead.
In some games the 5850 is 3-4% behind at the same clocks as a 5870, in other games the 5870 gets as much as 12-13% ahead.
There is no right answer, not least because if they alter the ratio's very slightly to get better performance in more average games but slightly less performance in something like Metro 2033, but then, what if every game out next year uses more and more dx11 features and runs more like Metro 2033, in which case you'd be reducing performance for newer games, to bump performance in older games.
THeres quite literally nothing known about it right now, its all guesswork, though the most obvious place to start guessing is quite simple, what size core can they make, without dropping profits and without effecting yields significantly.
If the process has matured enough that the doubling of via's and extra space to leave room for various sized transistors(one of the key problems with TSMC's process) then they could save anything up to 15% on the die size doing that, which AMD hinted at in interviews was around the cost of adjusting for the issues with the process. Nvidia dropped 33% off their die size, and still end up significantly worse than AMD in transistor density, just to make something they can sell for not a loss, not for a great profit, and don't forget the GF104 is far far newer than the 5870 die, my guess would be they've built in even more redundancy than AMD largely because leakage is still effecting them(not far off a year later now) than AMD because of the much higher clock speeds their design uses.
This all suggest the maturing of the process people seem to think has to happen, well, hasn't, at all. If Nvidia are still making designs 10% bigger with 10% less transistors(that add's up to what, 20% lower transistor density than AMD) then I'd say, the process hasn't improved much in 8 months and Nvidia is seemingly slashing prices to maintain a 15-20% price difference as AMD prices come back down at retailers, so Nvidia clearly don't think the 460gtx is within 15% of 5850 performance, no idea why anyone else thinks so. None of that spells great process and maturing quickly.
.
.
