Machine Learning Systems · Fall 2026

AI Infrastructure

Not every student will become a great AI infrastructure architect, but a great architect can emerge from any student.

Instructors Professor Congliang Chen, Li Shang, and Yuedong Xu Term Term 1, AY 2026–2027 Format Build what you use, and use what you build

Course overview

AI Infrastructure (also known as Machine Learning Systems) examines how algorithms, systems software, and hardware are co-designed to make AI workloads efficient, scalable, and reliable. The course covers the full stack: GPU and NPU architecture, memory hierarchies, communication systems, accelerator programming, compilers and runtimes, large model (LM) distributed training and inference.

The course is hands-on and engineering-driven. Lectures establish architectural and systems foundations; in-class hackathons then ask students to build, profile, debug, and optimize components used in modern training and inference systems. Emphasis is placed on quantitative reasoning, reproducible measurement, and iterative system improvement.

A recurring theme is agentic AI for systems development. Students will explore multi-agent workflows for accelerator profiling, kernel generation and optimization, experiment management, and training and inference infrastructure. The goal is to understand both what these workflows can automate and how their outputs should be evaluated and verified.

The course will also feature industry-defined project opportunities. Candidate topics may be contributed by AI-infrastructure organizations, such as Huawei, Tencent, and ByteDance. Final partners, topics, awards, research opportunities, and internships are subject to confirmation and will be announced through the course website.

We believe the future of AI infrastructure development should not be limited to a small number of highly specialized engineers inside large organizations. By combining systems thinking, engineering practice, and agentic methods, students will learn to build and optimize sophisticated AI systems while developing the judgment needed to validate correctness, performance, and reliability.

Prerequisites

Recommended preparation includes computer architecture, operating systems, machine learning, and compiler fundamentals. Basic programming proficiency is expected; prior GPU or NPU programming experience is helpful but not required.

Learning outcomes

By the end of the course, students will be able to:

  1. Explain the architectural foundations of modern AI infrastructure. Analyze how AI workloads interact with accelerators, memory hierarchies, interconnects, compilers, runtimes, distributed training frameworks, and inference systems.
  2. Develop practical optimization skills. Implement, profile, debug, and optimize accelerator kernels and system components across computation, memory, communication, training, and inference workloads.
  3. Apply systems thinking and hardware–software co-design. Identify bottlenecks across the stack and translate high-level algorithmic requirements into efficient implementations under real engineering constraints.
  4. Build and evaluate agentic systems-development workflows. Use LLM-based agents for profiling, analysis, kernel generation, debugging, experiment management, and optimization while verifying correctness and performance.
  5. Work across emerging AI-computing ecosystems. Adapt tools and methods to GPU, NPU, and domestic accelerator platforms and address practical portability, deployment, and optimization challenges.
  6. Design next-generation intelligent computing systems. Integrate system expertise, empirical evidence, and agentic methods to propose and evaluate new infrastructure designs independently.

Tentative schedule

Week Theme Format Topic
1 Introduction Lecture The Intellectual Map of AI Infrastructure
2 GPU/NPU Architecture Lecture GPU Architecture
Lecture slides: Part 1 · Part 2
Supplemental readings: Buck (2004) · Buck (2015)
3 GPU/NPU Architecture Lecture NPU Architecture
4 CUDA Programming Lecture CUDA Programming Through the Lens of GPU Architecture
5 CUDA Programming Lecture CUDA Programming as Hardware–Software Co-design
6 GPU Kernel Development In-class Hackathon Agentic CUDA Kernel Development and Optimization
7 NPU Kernel Development In-class Hackathon Agentic NPU Kernel Development and Optimization
8 NPU Kernel Development In-class Hackathon Agentic FlashAttention Development and Optimization on NPUs
9 LLM Training Lecture LLM Training Systems A–Z: Data, Tensor, Pipeline, and Expert Parallelism
10 LLM Training Lecture Advanced Topics in Large-scale LLM Training Systems
11 LLM Training In-class Hackathon Agentic Communication-kernel Development and Optimization
12 LLM Training In-class Hackathon Agentic Parallel-training Infrastructure Development and Optimization
13 LLM Inference Lecture LLM Inference Systems A–Z: Performance Optimization
14 LLM Inference In-class Hackathon Agentic LLM-inference Infrastructure Development and Optimization

Assessment

Assessment emphasizes engineering ability, quantitative evidence, reproducibility, and clear technical communication.

Component Weight Primary evidence
In-class hackathons / homework / quizzes 50% Correct solutions and implementations, quantitative evidence, engineering logs, and concise reflections
Course project 50% System design, measured improvement over a baseline, reproducibility package, final report, and demonstration

To be finalized: submission rules, late-work policy, and individual/team contribution requirements will be confirmed before the course begins.

Readings and tools

The course does not rely on a single required textbook. Readings will be drawn from research papers, architecture and programming manuals, compiler and runtime documentation, technical reports, and instructor notes.

Course expectations

AI and automation tools. Their use is expected when appropriate, but students remain responsible for correctness, safety, and performance claims. Submissions should disclose material tool use and preserve the prompts, configurations, generated artifacts, and validation evidence needed to reproduce the work.

Teamwork and attribution. Collaboration may be permitted for hackathons and the semester project. Each submission must identify contributors, external sources, reused code, and individual responsibilities.

Reproducible and responsible computing. Students should retain code, configurations, environment details, experiment logs, and baselines. Credentials, restricted data, and shared computing resources must be handled according to institutional and platform policies.

The schedule is tentative and may change in response to platform availability, project partnerships, and developments in the field.