4 ms·

Huh... that's actually faster than I expected. I figured that the normalization processing overhead would be higher than that.