oh, right. I think that's the answer.
BONUS: for next year, the G6630 gives 50% performance increase + the expected node shrink allows a further ~30% overclock. 1.5*1.3 = 1.95 ~ 2.0.
For the following year, they could go for G6630MP2 (the 6 clusters can be multicored as a whole). Next (three years from now), they could go G6630MP3 plus another node shrink. By that time, there could be 8/10 clusters available G6830, G6A30 etc, and an entirely new architecture in the works, though probably not yet available.