Optimization via theoretical maximums

Why performant code matters (but gets widely ignored), with Casey Muratori

Topic 5 of 18

Published Aug 26, 2026 · Clip 23:07–27:04 (3:58)

Casey argues that the common profiling-and-measure approach to optimization is wrong; great optimizers instead determine the theoretical peak capability of the hardware, measure the delta between that and current performance, and work to shrink the gap with explainable reasons. This approach finds anomalies like undocumented CPU features (e.g., new Intel rename/rat table behaviors) and doubles as a learning mechanism, analogous to Simon Eriksson's Napkin Math benchmarks at Shopify that caught wrong benchmarks.