AI Systems Performance Engineering: Optimizing Model Training and Inference Workloads with GPUs, CUDA, and PyTorch by Chris Fregly

AI Systems Performance Engineering: Optimizing Model Training and Inference Workloads with GPUs, CUDA, and PyTorch by Chris Fregly

The core thesis of AI Systems Performance Engineering is that throwing more hardware at an AI workload is an expensive, fundamentally flawed strategy; true scalability requires hardware, software, and algorithmic co-design. Fregly targets a widespread, costly failure pattern across modern tech companies: software teams regularly deploy massive generative or neural network models using standard, high-level code configurations without considering memory bandwidth limits, compute bound states, or data communication delays. This results in massive server bills, low hardware utility, and major inference delays.

Instead of presenting high-level conceptual summaries or basic cloud service guides, Fregly digs deeply into low-level infrastructure tuning. He teaches engineers to look past standard processing speeds to analyze "goodput"—the actual volume of useful data a system processes per unit of time. The book provides intensive coverage on diagnosing bottlenecks using advanced diagnostic tools like NVIDIA Nsight Systems and PyTorch Profiler. It shows readers how to bypass restrictive C++ boilerplate using modern compiler stacks like OpenAI Triton to write high-impact custom GPU kernels, handle memory layouts effectively, and remove processing delays across large-scale distributed computing systems.