My research focuses on mitigating the rapidly growing energy cost of artificial intelligence computation by developing a new ultra-low power compute approach using photonic integrated circuits (PICs). PICs – microchips that let light flow instead of electricity – have become increasingly popular since light can carry higher data bandwidths while consuming far less power. As a result, they offer a promising path towards powerful but energy-efficient AI accelerators. We minimize energy consumption by using programmable phase change material that can store data on-chip at zero energy, and organize PCM-based compute cells in a specialized architecture (systolic array) for efficient data reuse. In the past year we developed integrated neural network and chip modeling tools to optimize system parameters, which shows that our approach can outperform state-of-the-art cloud-based processors. We are currently designing and fabricating a photonic-only version of our chip, and will be demonstrating its benefits before moving on to designing a single chip that integrates photonics and CMOS transistors.
Advisor: Sajjad Moazeni – Electrical & Computer Engineering