[Sesse__]

Useful, then, that you can start several vectorized floating-point muls each cycle. (E.g., most modern x86 are 3/0.5 cycles for vmulps. No 20 cycles in sight.)