Learn Kernels, a field guide to GPU performance

Field notes

The book arranges GPU performance as a hierarchy, beginning with threads, blocks, and warps before moving through kernel optimization, inference engines, and distributed serving. Its scope includes matrix multiplication, tensor cores, attention, quantization, continuous batching, mixture-of-experts systems, and current hardware such as NVIDIA Blackwell, AMD, TPU, and Trainium. The homepage also exposes a glossary, reading list, developer area, API, MCP, and llms.txt resources.