Conference: Imagenets4EDA: Towards Open-Source Datasets for AI in Chip Design

New York University, New York NY

Investigators

Abstract

Machine learning (ML) tools and Large Language Models (LLMs) have shown great promise in making the complex task of designing a computer chip design easier, faster and better. However, modern ML tools and LLMs rely on vast quantities of high-quality training data. Unfortunately, much of the data that can be used to train ML models for chip design are held by semiconductor companies who cannot release this data publicly. The paucity of large and good datasets poses a fundamental challenge to progress in this area. Such a challenge was encountered during the early years of the ML boom. At the time, a public dataset of images, ImageNet, played the role of catalyzing progress in the area of ML for image recognition. To mirror this effort for chip design, this project will convene leading experts from academia, industry and government at an “Imagenets4EDA” workshop. Via a sequence of panels, talks, and brainstorming sessions, the goal is to put forth a concrete agenda on how industry, academia and government can work together to achieve the common goal of building the equivalent of the ImageNet dataset for chip design. The outcome will be a concrete action plan that participants will commit to pursue. The organizers will make every effort to ensure that a broad range of voices, including participants from under-represented groups, are invited and heard at the workshop. Participants will also be encouraged to think about ways to achieve geographic, institutional, and demographic diversity in the group of students and researchers involved in data collection efforts and benchmarking competitions. The project will fund the participation of US-based researchers and students in the workshop. Recently, ML methods, like generative AI, LLMs and reinforcement learning (RL), have shown remarkable ability in performing a wide range of tasks in hardware design. Their applications in the design of computing stacks promise to revolutionize hardware code generation, system-level, and the electronic design automation (EDA) flow. Yet there is a critical need for datasets and benchmarks to realize this promise. By bringing together a community of experts in this area representing all key stakeholders, the ImageNets4EDA workshop agenda will pursue three synergistic goals. First, gaps in existing datasets will be identified via discussions and analyses of existing datasets, and pinpointing areas where current datasets are lacking. The second goal is an open call to the community---academia, industry and government---to contribute datasets in ways that protect the intellectual property rights of companies, while still providing sufficient quality of data that would enable training of foundation models for EDA. The third direction is a plan to organize benchmarking competitions by deciding one step in the EDA flow, for instance, physical design. ML tools will cut chip design lifecycles, improve productivity of semiconductor designers, and result in faster and lower chips, providing large benefits to US economy and society. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.