Pythonic Perambulations

Abstraction: Benchmark comparison of Numba JIT versus Cython for Python numerical code

Key points:

• Benchmark task: pairwise distance computation on 1000 points in 3D — a common pattern in ML and statistics.
• Pure Python loop is ~100x slower than NumPy broadcasting; NumPy broadcasting uses hidden temporary arrays causing memory overhead.
• Numba autojit decorator (LLVM-based JIT) yields a ~1400x speedup over pure Python — about 50% faster than the previous year's Numba release.
• Highly optimized Cython (with all type annotations and bounds-checking disabled) is a few percent slower than plain Numba despite requiring years of Cython expertise.
• Fortran/f2py is ~2x slower than both Numba and Cython; scipy.spatial.distance.cdist is ~50% slower than Numba; sklearn.metrics.euclidean_distances is several times slower.
• Timings summary: Python loop 13.4s, NumPy 0.111s, sklearn 0.0356s, Fortran 0.0167s, scipy 0.0129s, Cython 0.00987s, Numba 0.00912s.

Connections: Numba · Cython · Numpy · Scikit Learn · Jit Compilation · Scientific Computing · Performance Optimization

Source: http://jakevdp.github.io/blog/2013/06/15/numba-vs-cython-take-2/