Designing High Performant Systems for AI

Catalog description

The objective of the course is to give a broad overview of the challenges and opportunities that exist in designing high performance AI systems. In addition, a course project will allow students to delve deeper into a topic of their interest. The course is designed to cater to two types of audiences: students working on data science projects who want to understand how to perform faster training or inference of their AI/ML models, or students working on parallel algorithms, or hardware acceleration, who want to understand modern techniques for accelerating data science applications. On the theory side, the course will cover basics and some recent advances in improving the performance of state-of-the-art AI/ML techniques including Convolutional Neural Networks (CNN), Graph Machine Learning (GML), and Transformer based Natural Language Models (NLM). Additionally, a high-level discussion of recently developed custom AI accelerators such as Microsoft's NPU, or Cerebras will be covered. On the practical side, the course will cover programming models and frameworks for accelerating these models. These will include parallel programming techniques in PyTorch (for CNN and NLM acceleration), Framework for Graph ML such as Deep Graph Library, and heterogeneous computing frameworks such as OpenMP, and DPC++. The focus will be primarily on algorithmic optimizations as opposed to device specific optimizations. While the course lectures will cover the breadth of the domain, students will be able to explore the depth of a single topic of their choice by a course project. Offered as CSDS 375 and CSDS 475 . Prereq: CSDS 310 .