Might be worthwhile to run the same code with an appropriate numba decorator. My guess is that you'd get at least as much speed up but without having to change the sieve declaration, but I'm not sure.