GPU Programming for AI Engineers
An Independent Guide to CUDA, PyTorch and Performance on NVIDIA GPUs
-
- $17.99
-
- $17.99
Publisher Description
GPU Programming for AI Engineers is an independent, practical guide for Python engineers who train and serve models and need to understand what happens on the GPU. It starts from a slow job and asks why it is slow. It explains how a GPU is organized, from warps and thread blocks to tensor cores and caches; why memory capacity, bandwidth and transfers usually decide performance; and how the driver, CUDA Toolkit, libraries and PyTorch fit together, including their license terms. It then teaches correct timing and asynchronous execution in PyTorch, vectorized thinking, CUDA C++ kernels, tile kernels and custom PyTorch operators. Later parts cover profiling with torch.profiler and Nsight tools, input pipelines, mixed precision, torch.compile, out-of-memory diagnosis, DDP and FSDP2, inference optimization, containers, Kubernetes, MIG, reliability and cost per result. One fictional running example, a water utility's pipe-inspection model and field assistant, connects every chapter with code, worked numbers, checklists, a lab and a glossary.
This book is an independent publication and is not authorized, sponsored or endorsed by any company whose products it mentions. All trademarks belong to their respective owners and are used only for identification. This book was drafted with generative AI tools and checked with automated tools. Tools, features and prices change often; confirm them in current documentation.