Hm, if people are that interested in the 660TI performance, could someone who has a 670 not do a theoretical 3dmark or heaven benchmark?
Downclock the memory to 192-bit rate. That is, assuming the 660TI runs GDDR5 at 1500Mhz (6000Mhz), running the memory at 1125MHz is 75% should perform similar to 192-bit. Same for the core. Might yield some interesting results. GK104 seems to be rather bandwidth starved so running at 3/4 should make a big difference. Also, while the expected 660TI performance must be aiming to beat pitcairn (Nvidia have had the luxury of knowing the 7850/7870 performance for months now), an other interesting thing would be someone clocking the memory at 1125 and their core as high as possible to see how such a card might compete with an overclocked pitcairn.
Seems Nvidia will be confusing unsuspecting customers again with those two 660s: the 660TI should perform as expected but the GK106 based 660 (no-TI) may be far slower. Nvidia marketing sure know to confuse things (although the GT 640s based on GK107 and GF116 must be a new world record).
EDIT: haven't paid that much attention to the way Kepler has to be overclocked. Is it even possible to dial down the memory like that while keeping the core at a different speed?