So were looking at 3200 shaders on the highest end part..
Not a bloody chance in hell.
GF100 is all but dead because frankly something that big is just so low yield, with no profits and with limited production capacity its daft to drop the output so much with such a big core.
Assuming it is 40nm (98% chance of that) then, GF100 is 3.05billion transistors, it would seem Nvidia is also, with the 460gtx using around 10% less transistors yet its 10% bigger than the 5870 die. So AMD can probably fit in around 3.2-3.3 billion transistors in something as big as the GF100 die, but they won't want to come close to that size, its simply not possible to make a decent yield high end part. AMD also aren't massively better in power usage, GF100 uses around 60% more power, it also is 60% bigger, so if AMD increased transistors by 60% or so, it would use 60% more power aswell, okay they are better than Nvidia, I would expect it to use less power, but not much.
So realistically AMD can't come close to the 3.3billion transistor mark, nor will they want to bump it beyond, what, 225W at most? At 3200shader part would frankly also require higher bandwidth, more rops, etc, etc, you'd be looking at a 4billion transistor part, its not going to happen at 40nm.
At best, they might go up to around 2000 shaders, if the efficiency increases means the rops/tmu's/bus take up a little less space throughout then they might just get enough room to get up to 2200.
I'm expecting something with a roughly 25% increase in die size/transistor count, something around 2.6billion transistors, and maybe a 30-35% performance increase.
The problem being, thats well without the performance limits 5850/5870/480gtx overclocked, however of course, if you overclock the new cards you get that lead back somewhat.
They might just go for absolutely smegging off Nvidia, if they can save 20% die space with efficiency increases, smaller more efficient rops/tmu's and a more effective balance of shaders/rops/tmu's/bus, then they could release things with similar performance at a lower cost, but I also don't really see that happening.
I think the new 6770 type cards(if there are any) could offer 5830>5850 performance at a lower cost though.
LIke I said, its interesting just because, due to the 40nm limit and complete inability to come close to the usual transistor count doubling, we have no idea where they will go, what they can get out of such a small increase available at 40nm and what cards they'll bring out at all.
The 5770 is a killer card largely because of its die size, while more performance is great, a bigger die is a bigger die, it will cost more. They might well continue to produce both and not EOL the 5770 series. With the 5xxx series becoming essentially the smaller value brothers of the bigger brother higher performance 6xxx series cards.