I don't know what you mean by "64 bit ISA outweighs the cache penalty for 64 bits." The usual reason for going with a 32-bit OS on a 64-bit capable machine would be that you don't want to buy 2 more memory sticks. I can't see that the code size in a 64-bit MKL FFT would be enough bigger than the 32-bit code size to incur measurable instruction cache misses, but you can easily compare performance of the 32-and 64-bit MKL if you choose a 64-bit OS. I don't know if anyone tested the MKL this way, but there is seldom more than 1% penalty for running a 32-bit application under a 64-bit OS rather than a 32-bit one. If such small differences are important, install both linux versions dual boot and compare them.
