[citation][nom]nforce4max[/nom]I have my doubts that it will consistently perform to those expectations. People need to factor in the bandwidth needs of those 16 extra TMU and 8 more ROP units. This card is likely to end up more bandwidth starved than the GT640 and GTX650 which will almost certainly bottleneck the shader. Nvidia should have kept all three memory controllers active and traditionally their designs allowed for decoupled memory controllers which is still true with Kepler. If it had remained as a 192bit card its performance would have been noticeably higher than what will hit the shelves.[/citation]
Having more ROPs doesn't mean that it needs more bandwidth to have the same performance. Even without the extra ROPs, this card should be able to outperform the 650 and the 550 Ti significantly. Adding ROPs and such doesn't hinder performance, although the added parts might not be effectively utilized for the performance gains that they could have given had other components such as the memory controllers been able to keep up with them.
It will probably be much more memory-bandwidth starved than either of the other two cards, but it should be able to outperform them nonetheless. Heck, the GTX 680M has only about 30% more memory bandwidth than this card, yet it can fight with the Radeon 7970M (although I'd take the 7970M over it), a card that is easily almost as fast as Radon 7770 CF. This 650 Ti will undoubtedly have inconsistent performance (even the 650 has inconsistent performance where it can jump around from roughly on-par with the 7770 to weaker than a GTS 450), but at least on average, it should out-perform the 7770.
Maybe I'm overestimating its capabilities, but even if I'm correct, I'm not saying that I'd buy nor even recommend this card. The 7850 is only a little more expensive and is without a doubt a far superior card. and even if it loses at stock, the 7770 should have no trouble keeping up in overclocking performance even if it takes a little 2x MSAA to even out the performance comparison a little.