Jeff Dean (Google)
Title: A Sketch of How to Move Towards an End-to-End AI-Automated Chip Design Process
Abstract
Customized hardware that is purpose-built for particular applications can be orders of magnitude higher in performance and more efficient than general-purpose computational devices. However, because of the significant effort involved in creating new purpose-built chips (e.g. hundreds of person years in the design and implementation phase across one to two years of actual elapsed time before even getting to fabricating the customized chip), specialization is used only in a limited set of circumstances. In this talk, I'll highlight progress on using machine learning to automate various aspects of the chip design process, including computer architecture choices, logic design, verification, and chip floorplanning, and sketch out a vision of how we might be able to do more end-to-end automation of this process.
Bio
Jeff Dean is the Chief Scientist for Google Research and Google DeepMind, and co-leads the Gemini project. He has worked for many years at the intersection of computer systems for machine learning, including work in ML accelerators, low-level software and frameworks for machine learning, work on sparse model architectures, algorithms like distillation and neural architecture search, training of large language and multimodal models, and applications of machine learning to areas like ASIC design, healthcare, and translation. He is a recipient of the ACM Prize in Computing, the IEEE John von Neumann medal, the Mark Weiser Award, & best paper awards at NeurIPS, OSDI, OOPSLA, PLDI, SOSP, & MLSys. He is a Fellow of the ACM, and a member of the US National Academy of Engineering, and the American Academy of Arts and Sciences.