A long time (10 years) ago I wrote some custom C++ code to perform relatively simple arithmetic operations on very large arrays, which was considerably slower in Numpy for some reason [1]. As I said that was long ago so maybe Numpy has improved by now, but back then there was a significant penalty when doing these operations directly in Python using Numpy, even though they were only a handful of things (e.g. calculate asin(x)+bcos(y) on some large array), so I didn't expect much overhead from going back and forth between Numpy-internal code and Python.

1: https://github.com/adewes/python-qubit-setup/blob/master/lib...