While it is true that the GTX 680's GK104 is a data-starved GPU, one should not compare different GPU architectures solely on their memory bandwidth, let alone from different manufacturers. Apart from Turing, NVIDIA's recent architectures (Maxwell and Pascal) are much less bandwidth-dependant than Kepler and Fermi, mainly thanks to the Tiled-Cached rasterization process (and the huge L2 cache, which is a consequence of TCR). A good example for this statement is the GTX 1060 (~192 GB/s, 1.5MB L2 cache), which does not scale that much (if not at all?) when paired with higher frequency GDDR5 and even GDDR5X DRAM.
This is the only gddr5x review I could find:
https://www.techpowerup.com/reviews/KFA2/GTX_1060_6_GB_GDDR5X/34.html
The GDDR5X 1060 had the worst gpu overclock but still managed a 15% boost in performance due to the massive memory overclock. Others did not see any gains with the 9 GB/s gddr5, but it looks like some of the newer titles are memory starved with the 192 bit bus.