Well, the early impression was 20% more shaders + 20% more performance per shader, since then we've also seen a potential 30% increase per shader purely by feeding the shaders better. IE 5870 - 30% shaders and the 6870 does beat it on occasion, though behind it in most situations, so you're looking at another 20% of so efficiency there.
I wouldn't be surprised to see it circa 60% faster, the question/problem is, how many of the throughput fixes from the 6870 made it into the 6970. If they were essentially being designed in paralel, its very possible they ripped all the changes out of a 6970 front end and stuck it with evergreen shaders to save time, or power, for the 6870, or its possible those efficiency increases came from the 6870 and went into the 6970, or that they were worked on by different teams and they didn't have time to include them in the 6970.
Its more likely they were in the 6970 originally, as Cayman will essentially be a shrunk(in shader/rop/tmu number and maybe bus aswell) 32nm design, so the new front end was likely a long time coming and just fitted into a 6870 for time to market for the midrange.
So 60%, throw in 5-10% clocks from the 5870 aswell, throw in a potentially huge driver boost from simplication and some other minor tweaks and another 10% wouldn't be surprising either at this point.
In games where, well in Crysis the 5870 already is fantastically competitive with the 580gtx and I can see it being hugely in the lead, in other games, like the rubbish and massive Nvidia backing game like Lost Planet, not so much
The issue for me though, is as per usual I was hoping for a 10% shader, 10-15% downclocked 6950 to be the card I was going for. If its 20% shader drop, and potentially overclocks less well(both companies seem into limiting power to specs and its seemingly a dual 6 pin pci-e power card) then the 6950 won't be anywhere near as close as the 5850 was to the top end. I REALLY hope the 6950 launches with only a 10% or so shader drop, then we'll have a killer card.
