Let me criticize a little the correctness and consistency of the conclusions in this article as to relation of GTX 660 TI's memory interface and its behavior here in the "Testing For Memory Interface Limitations".
The authors of this article state that the only reason of the problem of GTX 660 TI (when its FPS drops at AA 8X) is its memory bus 192 bit. And then the say that its 192 bit bus is sufficient reason to not buy it. But ...
As we all know, the memory bus itself is not important. What is important is the product of the memory bus and the frequency of memory, which is called "memory bandwidth".
Remember, the old card HD 4830 has 256 bit bus and DDR3 and its bandwidth is 57.6 GB/sec? But a little newer HD 4770 which has only 128 bit bus but the faster DDR5 has almost the same bandwidth 51.2 GB/sec, HD 4770 being often faster than HD 4830.
Here we have:
HD 7950 bandwidth - 240 GB/sec, bus 384 bit;
GTX 670 bandwidth - 192.3 GB/sec, bus 256 bit;
HD 7870 bandwidth - 153.6 GB/sec, bus 256 bit;
GTX660TI bandwidth - 144.2 GB/sec, bus 192 bit - note, the bandwidth is only 6.5% less than for HD7870.
OK, let's assume that 6.5 % is still enough for such a big FPS drop at AA 8X.
Now, look at the graph "Absolute Framerate 2560x1440 (FPS)". What do we see?
1) Let's compare the red (HD7950) and the blue (HD7870) curves. They behave pretty much the same, the red one just being shifted up accordingly to higher performance of HD7950. The distance between them at AA 8x is even less than at noAA. But the bandwidth of HD7950 is 56.25 % larger than that for HD7870. And if you prefer to compare the buses - 50% larger. So, why such a big difference in memory bandwidth and bus does not lead to any changes in the curve behavior? It is if the bandwidth is allegedly so important.
2) Analogously, let's compare the green (GTX670) and the yellow (GTX660TI). Again, the differences for the bandwidth and the bus are considerable (both 33% ) and still no change in behavior of the curves - only the shift due to different performance.
3) The most interesting - let's compare the green (GTX670) and the blue (HD7870) curves. GTX670 has the same 256 bit bus and 25.2% larger bandwidth than HD7870. But GTX670 still drops below HD7870 at AA8X. What, again because of more narrow bus or less bandwidth?
In short, the conclusions about the main role of "bad memory interface" of GTX 660 TI in this article are, to put it mildly, exaggerated. There should be other explanations.
To my opinion, the memory bus, and the more so the bandwidth, have nothing to do with these "results" at AA 8X.
P.S. Why did they choose the game Batman Arkham City? It looks this game doesn't love the nVidia cards.
P.P.S Where is the comparison with Zotac AMP! Edition model of GTX 660 TI that has 158.6 GB/sec memory bandwidth which is 3.2% higher than for HD 7870?
P.P.S I didn't find any comments on power consumption comparison. Why does GTX 660 TI (TDP 150 Watts) consumes more watts than HD 7870 (TDP 175 Watts) while it has less TDP?