Title: Harnessing Large Language Models Towards Customized Hardware Design Automation

Yongan (Luke) Zhang

Ph.D. Student

School of Computer Science, Colleague of Computing

Georgia Institute of Technology

https://luke-avionics.github.io/

 

Date: Thursday, October 10th, 2024

Time: 3:00 PM - 5:00 PM EDT

Location (in-person): Klaus Advanced Computing Building - 3100

Zoom: https://gatech.zoom.us/j/6669781434?pwd=OG4xZk1XakJyUWVTRWJIdEFveUFNQT09

 

Committee

Dr. Yingyan (Celine) Lin (Advisor), College of Computing, Georgia Institute of Technology 

Dr. Hyesoon Kim, College of Computing, Georgia Institute of Technology

Dr. Haoxing (Mark) Ren, Design Automation Research, Nvidia

 

Abstract:

The rapid advancement of Artificial Intelligence (AI) has created a growing demand for customized hardware accelerators. However, the design process for these accelerators is time-consuming and requires extensive hardware expertise, hindering the adoption of customized hardware for many applications. My thesis research aims to democratize customized hardware design by leveraging large language models (LLMs) to automate the process. I propose a set of pioneering domain-specific adaptation techniques that enable LLMs to tackle hardware design tasks more effectively. These techniques include in-context learning flows with automatically retrieved hardware design demonstrations, an automated framework for generating high-quality hardware datasets to fine-tune specialized LLM, and LLM-friendly problem decomposition techniques for tackling complex hardware design flows. To support the practical implementation of LLMs in hardware design, I have also organized community efforts to create infrastructures that assist relevant research on LLMs for hardware design. Furthermore, I dedicate ongoing efforts to apply my proposed hardware domain-specific adaptation techniques to enhance LLMs' reasoning capabilities, enabling them to tackle more challenging and practical hardware design problems across different stages of the design flow, akin to a human designer. By harnessing the power of LLMs and developing domain-specific adaptation techniques, my thesis research aims to revolutionize the customized hardware design process, making it more accessible and efficient for a wider range of developers and applications.