Many people pay for NumPy's creation of temporaries and such. If you have a lot of Numpy code and use a lot of long expressions, it's worth looking at NumExpr, which can give greatly improved performance because it eliminates temporaries and does multicore parallelization.