lots of avoidable copies in NumPy, and even if you manage to get rid of that, the python interpretation overhead isn't amortized until your matrix dimensions ar

lots of avoidable copies in NumPy, and even if you manage to get rid of that, the python interpretation overhead isn't amortized until your matrix dimensions are in the several-hundreds at least.