I think that one approach that may yield domain-specific improvements would be to add certain numerical routines into the x86 instruction set.

When I was working in finance as a quant, I was shocked by the amount of time code spent executing the exponential function - it is used heavily in discount curves and similar which are the building blocks of much of financial mathematics. An efficient silicon implementation would have yielded a great improvement in speed.