The core clocks are not more important than the memory bandwidth on the Kepler cards because they have huge memory bandwidth bottle-necks. The 4GB models would be memory bandwidth bottle-necked to an extreme long before you could make use of 4GB of memory versus 2GB. If you do something that needs more memory, then you are better off getting one or more Radeon 7950s or 7970s instead because those 4GB GTX 670s and 680s are no better than the 2GB versions.
Have any of you wondered why the 670, despite having a significantly crippled GPU compared to the 680, still has about the same performance (an average of an approximately 3% performance drop is not worth the extra $100 for a 680 IMO)? This is because they are so memory bandwidth bottlenecked and have the same memory bandwidth that the 680's faster GPU doesn't matter much at all in gaming. In fact, the GK104 GPU in the GTX 670 is still a faster GPU than the Tahiti GPU in the Radeon 7970 for gaming purposes, yet it's huge memory bandwidth bottle-neck holds it back. That is why the 7950 and 7970 pull ahead as the resolution and AA are increased.