Bojie Li Releases Open-Source AI Infrastructure Book

Bojie Li has published "ai-infra-book," an open-source Chinese textbook titled "Understanding AI Infra: Quantitative Analysis and System Design." Released under the Apache 2.0 license, the project has quickly gained traction, securing more than 3,400 GitHub stars within weeks of its debut. The work serves as a companion to Li's previous book on AI agents, which has amassed over 45,000 GitHub stars.

The new textbook spans 12 chapters, covering critical topics such as model architecture, hardware accelerators, operators, super-nodes, datacenter networks, and distributed training. Drawing inspiration from the quantitative methodology of Hennessy and Patterson's classic computer architecture text, Li's book teaches readers how to design AI systems by analyzing physical hardware constraints like compute power, memory capacity, bandwidth, and communication bottlenecks.

To help practitioners avoid common sizing mistakes, the text structures its analysis around five core questions regarding data movement, focusing on what data is transferred, its volume, its frequency, its path, and which components are left idle. This framework helps engineers avoid errors like neglecting key-value cache sizes or overestimating performance based solely on peak floating-point operations. The repository also includes practical tools, such as a Python-based calculation command-line interface for recomputing per-operator resource costs, alongside roughly 20 gigabytes of Git Large File Storage experiments and a PDF compiled via XeLaTeX.

For AI engineers and system architects, this resource bridges the gap between high-level model design and low-level hardware execution. By learning to calculate order-of-magnitude resource estimates, practitioners can better budget for interconnect traffic across multi-accelerator servers and optimize distributed training workloads across thousands of GPUs.