ADEPT: Automatic Differentiable DEsign of Photonic Tensor Cores
TimeThursday, July 14th10:52am - 11:15am PDT
Location3005, Level 3
Event Type
Research Manuscript
AI/ML Design: Circuits and Architecture
DescriptionSilicon photonics is a promising hardware platform that represents a paradigm shift in efficient AI with its ultra-low latency and high efficiency. Composed of passive/active optical components, photonic tensor cores (PTCs) are designed to achieve ultra-fast tensor operations for neuromorphic computing. Current PTCs are based on empirical designs with regular structures. To our best knowledge, there does not exist automatic design methodology to explore the design space of PTC circuit topology. Therefore, in this work, we present a fully-differentiable design framework that can efficiently generate PTC designs that demonstrate high expressivity and good variation robustness under various hardware constraints.