Numba vs. Cython: Take 2

Abstraction: Python performance benchmark comparing Numba JIT and Cython for pairwise distance computation

Key points:

• Benchmark task: pairwise distance on 1000 points in 3D — Python loop=13.4s, numpy broadcast=111ms, Cython=9.87ms, Numba=9.12ms
• Numba uses LLVM JIT; wrapping plain Python with autojit decorator achieves ~1400x speedup with no type annotations needed
• Cython requires explicit type declarations and is a Python/C hybrid; despite full optimization, was ~8% slower than Numba here
• scipy cdist (C): 12.9ms; sklearn euclidean_distances (Cython): 35.6ms; Fortran/f2py: 16.7ms — all slower than Numba
• numpy broadcasting allocates large temporary arrays (MMN elements), causing memory overhead vs. loop-based approach
• PyPy excluded: does not support NumPy, making it unsuitable for scientific Python workflows

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

Source: http://nbviewer.ipython.org/url/jakevdp.github.io/downloads/notebooks/NumbaCython.ipynb