High-Speed Networks

Catalog description

This course offers a comprehensive overview of the principles and technologies driving today’s high-speed networks, with a particular focus on wide-area networks (WANs) and the massive-scale networking infrastructures that power modern AI data centers. Students will explore the evolution of networking from traditional WANs to state-of-the-art AI data centers, where hundreds of thousands of GPUs are interconnected to support large-language model (LLM) training and inference. The curriculum emphasizes techniques and technologies for Distributed Machine Learning (DML), aimed at accelerating both training and inference. Topics include strategies to increase communication throughput and reduce tail latency for collective operations such as All-Gather and All-Reduce, which are essential for enabling data, tensor, pipeline, and expert parallelism across GPUs, as well as All-to-All communication in Mixture-of-Experts (MoE) transformer architectures. Key areas of study for AI data centers include: ● High-performance interconnect architectures ● Multipath load balancing and adaptive routing ● Job scheduling for training and inference ● Switch-assisted congestion control using packet trimming and congestion signalling ● Resilient networking for long-duration AI training Coursework includes hands-on assignments using network simulation tools to evaluate algorithms and techniques, as well as quizzes, presentations, and a term project. | Co-requisites: ECE-GY 6353 or another computer networking course