[fred123]

SIMD, multithreading, better handling of NaN, maybe other factors.

[ddejohn]

Numpy uses SIMD too.

[sireat]

Exactly why I asked, Numpy is already pretty optimized on ..ahem.. numerical operations.

Maybe some optimization on multithreading

Still I am reasonably sure Numpy utilizes multithreading already. https://in.nau.edu/arc/parallelism-in-python

And here we assume there are no NaNs.

So something else.