Tutorial on ML-Assisted Architecture Design (MLAD)

Abstract

Design space exploration for domain-specific architectures has increasingly relied on machine learning (ML) techniques offering a promising approach to managing complexity. However, utilizing ML for design space exploration presents several challenges. Firstly, selecting the most appropriate ML algorithm from a growing pool of methods is not a straightforward task. Secondly, effectively assessing the trade-offs between performance and sample efficiency across these methods remains inconclusive. Finally, the absence of a comprehensive framework for fair, reproducible, and objective comparisons hampers the adoption of ML-aided architecture design space exploration and inhibits the creation of useful artifacts such as datasets or data-driven architecture research. 

Tutorial Overview and Objectives

In this tutorial, we introduce ArchGym, an open-source gymnasium and easily extensible framework that connects various machine learning algorithms to architecture simulators. Next, we provide a hands-on experience to the audience through presenting the entire process of adding a hardware simulator to ArchGym and leveraging various machine learning algorithms (e.g. Reinforcement learning, Bayesian optimization, Genetic algorithm, etc.) for design space exploration. Through this process, we expect the participants to learn how to integrate an ML algorithm into their custom simulator, enabling efficient exploration of the design space. We will cover training techniques and methods for collecting datasets to enhance the learning process. Finally, we will showcase examples of integrating CFU-Playground and Astra-Sim, illustrating the practical implementation of ArchGym. By the end of the tutorial, attendees will have a solid understanding of ML-aided design space exploration and will possess the skills to build and integrate custom simulators with ML algorithms using the ArchGym framework.

What You'll Learn

• Challenges and Opportunities of ML-Assisted Architecture Design

• How to build a generic OpenAI environment from a custom simulator?

• How to integrate and use an ML algorithms with the custom simulator gym environment for design space exploration?

• How to collect standard datasets from ML-aided design space exploration?

• MICRO 2023 Workshop Focus: Hands on experience of ML-aided design space exploration with CFU-Playground and Astra-sim

Prerequisites 

• Curiosity and interest in this research area

• Ubuntu Laptop for hands-on tutorial