[nkurz]

I agree, it's certainly higher than it would be in most cases. I have the luxury of being able to design the algorithms to match the strengths of the processor. You rarely see this much improvement unless you are optimizing specifically for Haswell with intrinsics or assembly, although occasionally Intel's compiler manages some magic vectorization I never would have anticipated. And you almost certainly won't see this amount of gain if you are running an older precompiled binary.