Nvidia Partners with Georgia Tech for Contest to Develop Cutting-Edge AI Dataset for GPU Design

Nvidia, in collaboration with Georgia Institute of Technology and other key partners, has launched an innovative contest aimed at building a groundbreaking open-source dataset to revolutionize AI-driven GPU design. Despite their prowess in content generation, large language models (LLMs) have struggled with hardware design due to a lack of comprehensive training data. This contest seeks to bridge that gap by creating a high-quality dataset tailored for hardware design.

Haoxing (Mark) Ren, Nvidia's Director of Design Automation Research, unveiled the collaboration on X (formerly Twitter), emphasizing the critical need for high-quality hardware design data. The ICCAD Contest on LLM-Assisted Hardware Code Generation is designed to overcome current limitations where LLM-designed hardware often requires significant human intervention and results in impractical designs.

The contest consists of two phases. The first phase, running until August 10, 2024, involves collecting data samples starting with the existing Verilog code dataset. The second phase, from August 20 to October 1, 2024, focuses on refining the dataset by filtering out low-quality data, generating precise descriptions, and developing effective labeling strategies.

Winners of the LLM4HWDesign contest will be announced at the IEEE/ACM International Conference on Computer-Aided Design in late October 2024. Sponsored by Nvidia and the National Science Foundation, the contest promises valuable awards and significant recognition for participants.