That seems to depend on a number of assumptions -- your timeline, whether you expect a soft or a hard takeoff, the centrality of raw intelligence vs. cultural effects to research quality, possible nonlinearity of network effects on intellectual output. But I'd bet that the big one is time: if you think (unrealistically, but run with it) that you can improve a test population's intelligence by 50%, that could be very significant if you're expecting a 2100 singularity but likely won't be if you're expecting one before they graduate from college.