If we only look at 16x10 and 19x10 resolutions, a GT640 is 24% behind a 7750 in perf/W. (Assuming all tests run at the indicated power.) That's doesn't seem to be an insurmountable gap by switching to GDDR5, considering the large amount of ROP and shaders and pathetic bandwidth.
It's not as if one has to beat the other into submission. Competitive is sufficient. Difference in brand reputation, perceived driver quality, and pricing (smaller die, as you pointed out) can do the rest.
Edit: FWIW, the only straight comparison I could find between DDR3 and GDDR5 for otherwise identical cards, is this
one. If you average the performance of gain for all 16x10 results, you get a speedup of 24.7%. (For this particular case, power consumption was surprisingly identical, but you need many samples to come to a firm conclusion on that one. That's true also for those THG results, of course.) It's a different architecture with different unit ratios, but it shows that a 25% improvement is not some number from lala land and it's better than pretending there is no difference.