The drop from 40nm to 28nm is, proportionally, one of the biggest drops we've ever seen. Transistor density will increase by ~2.05x, and I'm expecting roughly double the transistor count in the top end GPUs from both sides. Since GPUs are inherently parallel devices, performance scales close to linearly with increased transistor count.
Barring major design failures, thermal problems, or issues with the 28nm manufacturing process at TSMC, I'm expecting a significant increase in performance this generation. Both AMD and Nvidia should see a near-doubling of raw rendering power in their top-end GPUs. I expect this will translate to somewhere in the region of 70-80% improvement in GPU limited games.
We won't know anything for a few months by the looks of things, but there is every reason to expect a big jump in performance. For me, AMD will be the more interesting to watch this time around: They're making a major revision to their architecture (as Nvidia did with Fermi last generation), and it will be interesting to see how well it works out. With a brand new design there's always the possibility for a great success (9700pro, 8800GTX), or for teething problems that aren't resolved until later iterations (FX5800, x1800xt, GTX480).
