It's been clear from all reviews and even nvidia themselves that they sacrificed non-gaming performance to make the 680 a great gaming card. So I wouldn't expect it's folding performance to be nearly as high as a 7970.
That's not a logical conclusion because you have automatically assumed that compute performance = Double precision performance. Folding @ Home benefits from many fast CUDA cores, not gobbles of DP performance. If this was Collatz Conjecture, MilkyWay@Home or PrimeGrid, it would be different. What's the score that HD7970 gets in F@H?
7000 points at stock clocks? That would put it around 9000-9500 @ 1200mhz using that data point. The added compute performance doesn't seem to benefit HD7970 for this program unless there is a specialized OpenCL client that takes advantage of its double-precision capabilities?
GTX460 768 beats an overclocked HD6990 and is >2x faster than a single HD6970.
If Double precision performance mattered, this would have never happened.
GF104 chip has 1/12th the SP performance in double-precision.
Cayman has 1/4th of SP performance in DP.
GTX460 768mb = 113 Gflops of DP performance
HD6970 = 676 Gflops of DP performance (
6x more)
HD6990 = 1,352 Gflops of DP performance (
12x more than a GTX460)
Based on this estimate a 1200mhz HD7970 should roughly get 8000-10000 points, if that.
It looks like Folding @ Home primarily benefits from many fast CUDA cores and memory bandwidth. Therefore, GTX680 should crush a 7970. Considering a GTX460 > HD6990 despite 12x lower DP performance, I can't see how HD7970 can possibly outperform a 1536 SP @ 1058 mhz GTX680 with same memory bandwidth as a GTX580.