Thanks for the info fellas, pretty interesting to see how things have shaped up over the last few months.
Ya, the interesting part is with latest drivers and 1050mhz clocks on the 7970 GE, it is slightly ahead of GTX680 on average but GTX680 performs very well in Frostbite 2.0 games (BF3, and upcoming Medal of Honor game).
As you add more AA and increase resolution, HD7970 pulls away even more (but I admit at 2560x1600, a lot of enthusiasts are probably using 2 GPUs).
1920x1080 8AA HD7970 @ 1.05ghz is 15% faster than GTX680
2560x1600 8AA HD7970 @ 1.05ghz is 20% faster than GTX680
http://www.computerbase.de/artikel/grafikkarten/2012/test-amd-radeon-hd-7970-ghz-edition/4/
And HD7970 with 1.1ghz clocks is available for sale for
~$450 with free games too. So it's cheaper and faster than a 680. But because early drivers had many issues with 7900 series and performance was poor, the reputation that 670/680 cards are faster is set in stone. Not many people revisit reviews 5-6 months later.
Makes me wonder what AMD and Nvidia's next move will be on 28nm. We've already seen a potential monster from Nvidia from the compute side of things but the fact that both sides are practically even this generation changes the game somewhat. It'll be interesting to see if either company manages to make their architectures more efficient - I reckon AMD has more room to grow in this regard.
For gaming, NV is half a step ahead I would say. They have an excellent track record of making huge 450-530mm^2 die chips. GK104 is just 294mm^2. They can easily increase performance by 50% for games without too much sweat by just adding more SPs, ROPs, TMUs and widening the memory bus width to help with AA. Their power consumption is also better this round and Kepler architecture seems to handle deferred MSAA better.
AMD has a much more difficult task with HD8900 series. Their Tahiti XT chip is 365mm^2, and their strategy involves heterogeneous computing (so including FP64 and dynamic scheduling into gaming cards). They have already implemented a wide memory bandwidth bus (little room to grow there without GDDR6 on the market).
Die size wise, AMD has very little room to grow if they stay with Tahiti XT chip and expand it. If they ditch FP64, rebalance the ROPs to 48 and maybe add some enhancements to cache, rasterization/geometry engines and start off with Pitcairn, they can probably add a lot more gaming performance. But if AMD keeps FP64 and full compute intact, I have no idea how they'll add 40% more performance over 7970 GE on 28nm without sending power consumption past GTX480 levels. My gut feeling is in 2013 GTX780 will go back to GTX580/6970 (or GTX480/5870) days with NV having 10-15% more performance. NV might not want to go to large die if they don't need to since I bet they are making more $ on Kepler than on Fermi. We'll have to see though.