haouo/mit-6.5903-1-walkthrough

This repository contains bilingual study guides for MIT 6.5930/1, a course on hardware architectures for deep learning taught at MIT by Joel Emer and Vivienne Sze. No video recordings of the lectures exist online, so the author reconstructed each lecture as a textbook-style chapter based on the original slide decks, with key slide images extracted and embedded inline. All 13 lectures are covered, in both English and Traditional Chinese. Each walkthrough follows the same structure: a short summary, learning objectives, conceptual chapters (each with embedded slide figures and a note on why the concept matters), a self-study guide with a mastery checklist, self-check questions, exercises, and common pitfalls, followed by a glossary, key takeaways, and a map of how each slide connects to the chapter structure. The topics progress from an overview of why specialized hardware matters for AI, through the mathematical notation used to describe neural network computations (called Einsums), to how data is organized and moved through hardware, how sparse computations can reduce wasted work, and how numerical precision affects efficiency. The final lecture covers how to formally calculate the amount of data movement required for a given hardware mapping. The source slide PDFs are stored in the repository. A shell script converts specific pages from those PDFs into PNG images that both the English and Chinese walkthroughs reference. An automated check runs on every update to confirm that the English and Chinese files stay in sync, that all embedded images resolve correctly, and that each file contains all required sections. The walkthroughs are unofficial study materials created by the repository author, not by MIT or the course instructors.