Comment by ashdnazg

Pretty sure it's optimisations.

Even small things like converting multiplications followed by additions to FMA will reduce the operation count.

Add to that constant folding etc. and a speedup factor of ~3 is not so hard to imagine.