Volume I

Introduction to Machine Learning Systems

Lecture decks for Volume I, which covers machine learning systems built on a single machine with one to eight accelerators. The decks run from the course overview to the conclusion. The larger chapters are taught in two lectures, Part 1 and Part 2, and each part is a complete class that can be taught on its own. Every lecture comes as a PDF for projection, a PowerPoint file whose notes pane holds the speaker notes, and the LaTeX source.

All PDFs (ZIP) All PowerPoint (ZIP) Source

Introduction

Course OverviewThe systems that train and serve one modelPDF PowerPoint Source

IntroductionWhy accurate models still fail as systemsRead the chapterPDF PowerPoint Source

Part I: Foundations

System ParadigmsFour deployment regimes and the limits that define themRead the chapterPDF PowerPoint Source

ML LifecycleSix stages from problem definition to monitoringRead the chapterPDF PowerPoint Source

Data EngineeringRead the chapter

Part 1Treating training data with the discipline of codePDF PowerPoint Source

Part 2Feeding the accelerator and detecting driftPDF PowerPoint Source

Part II: Development

Neural ComputationCounting the arithmetic and memory of a networkRead the chapterPDF PowerPoint Source

Network ArchitecturesRead the chapter

Part 1Dense, convolutional, and recurrent networksPDF PowerPoint Source

Part 2Attention, transformers, and recommender systemsPDF PowerPoint Source

ML FrameworksWhat the framework decides between Python and the acceleratorRead the chapterPDF PowerPoint Source

Model TrainingRead the chapter

Part 1The time and memory cost of trainingPDF PowerPoint Source

Part 2Fitting a training step on one acceleratorPDF PowerPoint Source

Part III: Optimization

Data SelectionChoosing the examples worth a training stepRead the chapterPDF PowerPoint Source

Model CompressionRead the chapter

Part 1Fitting a trained model inside a device budgetPDF PowerPoint Source

Part 2Quantizing weights and activations to fewer bitsPDF PowerPoint Source

Hardware AccelerationRead the chapter

Part 1Memory hierarchies, dataflow, and the rooflinePDF PowerPoint Source

Part 2Mapping a computation onto the chipPDF PowerPoint Source

BenchmarkingRead the chapter

Part 1Reading a benchmark number and its conditionsPDF PowerPoint Source

Part 2Deciding from noisy measurementsPDF PowerPoint Source

Part IV: Deployment

Model ServingRead the chapter

Part 1Latency budgets, queues, and batchingPDF PowerPoint Source

Part 2Serving a stream of tokensPDF PowerPoint Source

ML OperationsRead the chapter

Part 1Rebuilding, checking, and releasing a model safelyPDF PowerPoint Source

Part 2Monitoring a live model and deciding when to actPDF PowerPoint Source

Responsible EngineeringRead the chapter

Part 1Fairness requirements and how to test themPDF PowerPoint Source

Part 2Lifetime cost, carbon, and data obligationsPDF PowerPoint Source

Conclusion

ConclusionPredicting where a whole system will breakRead the chapterPDF PowerPoint Source

Read the matching chapter at mlsysbook.ai/vol1. The teaching guide explains the speaker notes, the activities, and how to rebuild or change a deck.

Tip

PPTX files are image-based (300 DPI) and visually identical to the PDF. Use them for PowerPoint presenter mode and slide annotations. For editable slides, download the LaTeX source.