Deep Learning Systems: Algorithms, Compilers, and Processors for Large-Scale Production

Deep learning in production is not just a matter of choosing a model. Training demands compute and memory; deployment brings its own limits on latency, model size, and power. Andres Rodriguez takes a systems-wide view of these challenges, tracing how algorithms, compilers, and processor design work together to make deep-learning workloads practical at scale.

For data scientists, hardware designers, and performance or compiler engineers, the value of this perspective is in seeing beyond one layer of the stack. Choices made during model development affect the demands placed on software and hardware—and platform constraints can shape what is sensible to build in the first place.

Follow the work from model to deployment

The book builds from deep-learning foundations and model building blocks to applications, training, and distributed training. It also addresses techniques for reducing model size, helping readers understand how the demands of a model can be considered alongside the realities of running it.

Where software meets hardware

Later chapters turn to hardware architecture, compiler optimizations, and frameworks and compilers. Together, these topics illuminate how deep-learning code is mapped to hardware targets and how system-level choices influence execution across different platforms.

A shared view of the deep-learning stack

Rodriguez’s central emphasis is collaboration across disciplines. Data scientists need to account for deployment platforms; hardware designers benefit from understanding the models they may need to support; and compiler and performance engineers must work across diverse models, libraries, and targets. A connected view of these concerns can make technical trade-offs easier to recognize and discuss.

Who may find it useful

This book is especially relevant to engineers and technically minded readers working with deep-learning algorithms, processors, or performance software who want to understand how those pieces fit together. It offers a system-level perspective on the practical work of training and deploying models—not just the model in isolation.