I doubt they are short on inspiration for further improvements. A big part of the speed advantage of ifort is the aggressive loop unrolling, pipelining, splitting and multiversioning.

Coming up with these transformations is not the difficult part. That would be actually implementing these transformations correctly and keep the whole compiler optimization framework maintainable.

And of course there is certainly a cost for the improved runtime. Primarily in compile time and code size. For the Intel compiler it generally makes sense to sacrifice these in favor of runtime performance, because it is used on these kind of scientific codes a lot, but for gfortran the balance might different.