Count the bytes, not the FLOPs
Summary
The article analyzes tsbootstrap’s performance improvements by avoiding materializing large intermediate arrays. It outlines two main bottlenecks—memory traffic and Python state overhead—and presents a fused, kernel-based approach with a counter-based RNG to achieve significant speedups (3.1x to 20x) on sustained workloads. It also discusses when materializing the full tensor remains useful and why careful measurement and reproducibility matter in performance engineering.