Looks like the first reviews of the GTX580 are coming out!
The initial one is pretty exhaustive.
As expected, GF110 used by GTX580 is not based on the GF104 architecture, it’s GF100 based. Compared to the GTX480, shader and memory clock rates are up about 10%, SMs are up to 512 from 448, and power use is down about 10%, and the GTX580 has an even better cooler on it making it quieter than the GTX480. This all matches the rumors and speculation from the past two weeks.
The only uncertainty is whether GF110 will support the Tesla features of ECC and full throughput DP. One clue is that there are slightly fewer transistors in GF110 (3.0B as opposed to GF100 3.2B), but this could be from anything including simply better layout. (My personal guess is that ECC was removed, and we won’t see a GF110 Tesla, but this could easily be wrong.)
For CUDA programmers, GF110 is certainly superior to GF100, but won’t make any change to the way you code or use CUDA… it’s just a nicer card with about 15-20% more horsepower.