Some interesting stuff today then

We know that Fermi is a killer when it comes to tessellation, and that nvidia still haven't decided upon final clockspeeds. Not as much performance info as I would have liked, but to be honest, about as much as I was expecting.
I think the most interesting thing was the description of the new "polymorph" engine, which takes geometry setup inside the SMs. Aside from increasing geometry throughput by 8x over GT200, it also takes more of the logic inside the modular parts of the chip. This will make it slightly easier to scale the chip down to smaller mid-range and low-end parts (although on the flip side the presence of a large global cache will make things more difficult). Anyway, given the massive geometry performance and the way that Fermi is architected for general compute type stuff, it comes as no surprise that it performs tessellation so well (as demonstrated by the Unigine benchmark that nvidia chose to release).
As for real performance, they can quote whatever stats they like, but at the moment the only real benchmark information we have is from this video:
http://www.youtube.com/watch?v=xCE9kG-ForQ
Here we see, in Far Cry 2, the GF100 getting 84FPS average compared to 50FPS on the GTX285. It's hard to draw too much from this, especially since you have to expect Nvidia to have chosen the best game to show off their new toy, but at least we know that Fermi can be up to 68% faster than a GTX285. Looking at other FC2 benchmarks, this puts the card not too far behind the 5970. I wouldn't expect it to be so close in most games though. AMD will keep the fastest single card crown.
Anyway, it's clear that there are some very interesting new features in Fermi, and that it's going to be a great platform for Nvidia the next generation around. That said, I still think that this entire line is going to end up in the toilet because of TSMC yeilds and general instability on the design. I guess that's what happens when your competitor releases first and you have to rush to catch up. The next generaation should be a beast though, since the new and highly modular architecture will scale very efficiently to a bigger design with more SMs.