Without giving any range of numbers, without giving away anything ...
(I am speaking theory here, not about SandyB itself)
Benchmarks will scale based on the Amdahl's law ( http://en.wikipedia.org/wiki/Amdahl's_law ).
If some sub systems were doubled, and you use a benchmark that uses 99% of the time this subsystem, you ll see a scaling of 2X.
now, there are some specs out there for SandyB that already release publically, not very detailled, but it is there ... you know that it has AVX.
For example, if you look at linpack, it is mostly L1 cache sensitive, and Floating based ... you add on the top of this that AVX is 256 bits, and SSE4.1 is 128 ... You get some indication of the improvement ... I see a lot of speculation here, but I don't see a lot of people trying to do the math properly.
Please try to raise the debate a little, I know there are smart people reading here, I just don't see them posting the math they are asking me questions about by email ...

I am not allow to give away any performance information, just understand that for SandyB, doing the math right will help you more than using a early prototype, those are not tuned properly.