Beating Kepler in Cycles and GPGPU isn't going to be hard. In fact, Kepler saw a pretty big regression in performance compared to fermi in GPGPU. It isn't always a regression, but if I remember correctly, it's a massive regression in things that use dual precision, like LuxRender:
Agree with that, but I just want to point out that neither Luxrender or any other path/light tracing software is using double precision as it wouldn't make any sense. It's a Monte Carlo based simulation where the important is doing a lot of calculations, not the their precision.
Nvidia claimed nothing. I did the math there and may be wrong but I'm pretty confident about it. In fact I was aiming pretty low. The GTX 670 (Which isn't a full GK104) at 294mm^2 is 183% faster (2.8x) than the GTX 650 at 118mm^2 (GK107, segment that the GM107 is replacing) and the GTX 750 Ti is x2.4 the performance of a GTX 650 with just a 25% larger die.
If the GM104 jump in performance per square mm is similar to the Kepler one going from the 107 to the 104 chip you may see a huge gain if it is a ~300mm^2 again. Enough to beat the crap out of the GTX 580 and get close enough to the GTX 780 Ti which is 2.5x the performance of a GTX 750 Ti.
Remember that I'm always talking about Cycles performance, nothing else.