When Vectorized Arrays Aren't Enough
Summary
The post analyzes the hidden costs of NumPy vectorized operations, showing how each binary operation allocates output arrays and how per-element Python overhead undermines performance for large workloads. It surveys alternatives (NumExpr, Numba/JAX, and hand-written Rust/C code) and demonstrates how fusing computations can dramatically improve speed, with a McFACTS case study. The takeaway: benchmark on target hardware and consider fusion or low-level implementations for heavy numeric workloads.