The Mentor's Handbook
Teaching Kernel Engineering — everything a mentor must learn, from matrix multiplication up, to deliver the workshop. Every chapter teaches the idea from the ground up — plain words, a metaphor, a by-hand example, the real math, where it runs in production today, and a minute-by-minute plan for teaching it. Read it in order; by the end you can deliver the entire workshop. 37 chapters.
Start: how to use this handbook →The mentor's mindset, how to use this handbook, and the single story that connects all fourteen sessions so you always know where you are.
The bedrock. Matrix multiplication from the absolute beginning, why it is the whole game, what 'fast' really means, and the compute-vs-memory idea that governs everything. Feel these in your bones before anything else.
What a GPU actually is, taught with pictures. Threads and warps, the memory hierarchy, tensor cores, and latency hiding — the vocabulary and intuition students need before any code.
The heart of the course, session by session. For each rung of the ladder: the idea in plain words, the metaphor, the board work, the live demo, the number, and where students trip.
Where the GEMM skills meet real LLMs. Fusion, softmax, attention and FlashAttention, and why chatting with a model is a memory problem — all in teachable pictures.
So you sound current and credible. Hopper and Blackwell in plain words, DeepSeek's real kernels, AI that writes kernels, and how to debug — each tied to what shipped this year.
Turning understanding into a great live session. Minute-by-minute plans for all eight lectures and six workshops, the questions students always ask, and how to run the capstone.