CUDA Foundations

PMPP is the textbook, and chapters 1–6 are the ones that matter most: the execution model (grids, blocks, warps) and the memory hierarchy are the two mental models every kernel you write will lean on. GPU MODE rides alongside as the closest thing this niche has to a guild — practitioner lectures that recap PMPP with war stories, and a Discord that posts jobs. Don't skip the Stephen Jones talks. Nobody explains better why GPUs are shaped the way they are (latency hiding through massive oversubscription), and that “why” is what survives after the API details fade.

Watch for the occupancy trap early. Occupancy is a means to an end, enough warps in flight to hide latency, not a score to maximize, and chasing it produces slow kernels with beautiful occupancy numbers. And do the humble vector-add task; the load_inline workflow it teaches is how you'll iterate on every kernel through the Kernel Engineering phase.