Presentation
Democratize Chip Design --with Deep Learning and Automated Code Transformation
DescriptionIn the past six decades, electronic design automation (EDA) has done a remarkable job to improve the productivity of hardware designers. I suggest that the next phase of EDA is to enable many software programmers to design their own chips or accelerators for a wide range of applications for better performance and energy efficiency,. In this talk, I shall present our effort towards this goal. Coupled with our multi-decade research high-level synthesis (HLS), we developed and integrated multiple deep learning techniques, such as graph neural networks (GNNs) and large language models (LLMs), cross-modality learning, and agent-based design space exploration. For regular structures, such as systolic arrays, stencil computation, or even more general affine programs used almost all deep learning kernel, we can also use mathematical programming to achieve automated code transformation. Combining these techniques, we show very promising results of mapping software code to high-quality silicon implementations.
Presenter
Event Type
Keynote
TimeWednesday, June 259:00am - 10:00am PDT
Location3022, Level 3