According to my notes, MMX is short for Matrix Mathematics eXtension. MMX uses SIMD (Single Instruction Multiple Data) operations, which allegedly gives it an eight-times performance speedup for those things it is ideally suited for. For those who care about such things, the reason MMX is incompatible with FPU operations is because Intel, not wanting to create new registers, decided to map the eight MMX registers to the 64-bit mantissa part of the 80-bit floating point registers.