MEMS Optical Circuit Switching for AI Applications
ABSTRACT
For over a decade, Google has invested in the development of TPU chips for machine learning and AI applications. Starting with TPUv4, we introduced the use of MEMS optical circuit switches to produce a reconfigurable 3D torus networking layer that significantly improves availability and performance. In this presentation, I will discuss how MEMS optical switches have enabled TPU AI clusters at-scale by interconnecting thousands of TPU chips to create an exascale supercomputer.
BIOGRAPHY
Kevin Yasumura is a Principal Engineer/Director in Google’s Platforms Optics Team. At Google he is the Technical Lead for MEMS and optical switching development. His optical switches are used worldwide for both datacenter networking and machine-learning applications and are critical to Google’s entire business model. He has over 25 years of experience developing MEMS-based products and systems. Before joining Google, he was a founding member and Manager of Micro-machining Development at Iolon Corporation where they developed MEMS tunable lasers, tunable filters, and optical switches. He was also Senior Director of Product Development at Formfactor where he led the development of MEMS-based integrated circuit test solutions. He holds a bachelor’s degree in physics from U.C. Berkeley and an M.S. and Ph.D. in applied physics from Stanford University.