Learn Kernels

Learn Kernels is an interactive book on GPU performance engineering, working from a single inference request down to one kernel and back out to fleets of machines. Its chapters cover threads and warps, matrix multiplication, attention, Triton and CUTLASS, KV caches, quantisation, speculative decoding and distributed serving, and every factual claim links to a primary source. Its sourced, interactive format is a model for technical explainers, and it publishes a JSON API and MCP server for agents.