The simplest suggestion is to call the Intel Math Kernel Library for your matrix multiply (DGEMM or similar). You will automatically get the best parallel performance. The Fortran compiler can recognize some forms of matrix-multiply and call MKL on your behalf, if you compile with auto-parallelization and the highest optimization level, but that will also allow some optimizations that may not be as precise as you want.
The /fast option is shorthand for a set of options that usually improve performance, but note that this implies /QxHost which assumes you are running the program on the same CPU type you compile on.
