Team UMD led by ECE Student Carsten Portner Places in International LLM Competition

ECE student Carsten Portner recently placed third in the Large Language Model-Assisted Hardware Design Code Generation (LLM4HW) Contest at the 2024 International Conference on Computer-Aided Design (ICCAD’24). The competition, sponsored by NVIDIA, aims to “harness community efforts to develop an open-source, large-scale, and high-quality dataset for hardware code generation, igniting an ImageNet-like revolution in LLM-based hardware code generation.”

The UMD team, comprised of Portner and Professor Cunxi Yu, focused on data preparation and generation techniques for fine-tuning large language models (LLMs) in the field of design automation. With two primary goals determined as the basis for the project, the team successfully completed both of their objectives.

The first objective was to develop novel methods for generating and collecting high-quality Verilog samples. This was completed by using high-level synthesis (HLS) compilers that allowed them to create a diverse set of hardware designs from high-level specifications, thus allowing the team to capture a wide range of design patterns and optimization strategies resulting in a diverse and robust dataset for LLM fine-tuning.

Secondly, by implementing advanced data cleaning and labeling techniques, the team was able to improve the quality of existing datasets. Techniques used included the systematic removal of noise, error correction, deduplication, and the addition of meaningful contextual labels. The results ensured that datasets were both accurate and better aligned with LLM training requirements.

As a third-year student, Portner is pursuing a combined Bachelor and Master degrees in Computer Engineering with a minor in Robotics and Autonomous Systems. Under the guidance of Professor Yu, he is studying electronic design automation with a focus in Verilog generation using large language models and machine learning for Boolean network optimization. He is a recipient of the prestigious UMD Banneker-Key Scholarship and is an Undergraduate Teaching Fellow for ENEE150, an intermediate course in Linux and C programming.

Published November 25, 2024