800 SP just doesn't fit with the rest of the speculation.....

We know that there are not enough TMUs already in rv670. 320 SP with 16 TMUs causes a bottleneck by the TMUs. Even if the TMU count is doubled to 32 it will have an even WORSE ratio of SP:TMU than rv670 which would make no sense whatsoever. 800 SP with 32 TMUs would have a bigger TMU bottleneck than 320 SP with 16 TMUs. It also doesn't make sense from a die size/TDP point of view. rv770 is not rumored to have a monster die size or TDP so the transistor count can't be THAT much greater than rv670.

Also 512-bit + GDDR5 = pointless
current memory bandwidth with 256-bit and GDDR4 @ 2200 = 70.4GB/sec
512-bit + GDDR5 @ 3800 = 243.2GB/sec

There is no need for that amount of memory bandwidth. GDDR5 will not be cheap when it is first released, nor is making a card with 512-bit memory interface cheap. Doing both would make the card more expensive to manufacture with an insignificant performance gain. It will be either 512-bit or GDDR5, not both.