How to read it

If you are new to the field, start at Prerequisites. It is the one chapter that assumes nothing, and it exists so that the others can assume a great deal: what a GPU trades away to be fast, how to read a kernel listing, what a transformer computes, and why generating text one token at a time turns that computation into two different jobs. Then read Foundations and do not skip it. From there on the book assumes you can say what a warp is and why a kernel might be memory bound. After that the chapters are reasonably independent: read them in order for the argument, or jump to whichever layer you are currently stuck on.

If you already write kernels, skip Prerequisites entirely; it contains nothing you do not know. The chapters after it will still move faster than you need in places, and the reading list is probably the more valuable artifact for you: it is a curated set of primary sources, and the chapters are best treated as a map of which of them to read and in what order.

One convention is worth knowing before you start. This book keeps vendor claims and independent measurements apart, and says which it is showing you. A figure attributed to a specification is that document’s own number. A figure from a paper or a worklog is that author’s measurement, on their hardware, in their configuration, and the chapter says so. Where two sources disagree, and they do, the disagreement is reported rather than settled by picking whichever number reads better.

Terms with a dotted underline are defined inline; hover or tap them. Most are collected again in the glossary, where each definition is quoted from its primary source. Figures in this book are there to show structure, not decoration, and the caption always says whether the numbers in one are from a specification or merely illustrative.