Yeah, but it is pretty consistent when you use data just from guru3d as well.
Their
7970 Overclocking Review 1125Core 21% Overclock
Battlefield 3 : 17% improvement with 21% OC
Crysis 2 : 16% improvement with 21% OC
AvP : 16% improvement with 21% OC
Their
680 Overclocking Review 1300Core 29% Overclock
Battlefield 3 : 15% improvement with 29% OC
Crysis 2 : 16% improvement with 29% OC
AvP : 17% improvement with 29% OC
You see the same playing out in OC results elsewhere for the 680.
If you contrast Kepler scaling to Fermi you see that Kepler does not scale nearly as well as Fermi does with overclocking. Another
review from guru3d with a GTX 580 overclocked by 28% from 772core to 987core.
There you see Fermi scaling like Tahiti/7970 does with near perfect scaling. 29% improvement in BFBC2 with 28% OC. I'd assume the mem overclock helped here as well with the better than 28% OC to performance result.
Taking Kepler OC results contrasted to the results you'd gain from a good clocking 7970 Lightning would make the 7970 Lightning more appealing to me at the same pricepoint of current 680s, $499. Tahiti scales better with overclocking and the Lightning models are cards built to overclock.
Non-reference 680s may be impressive, but they're going to have to clock really high to give big gains with the less than Fermi-esque scaling Kepler has with core speed increases.