LightSim: A Flexible, Fast Modeling Framework for Holistic Photonic Computing Systems

Photonic computing presents a promising alternative for accelerating neural network (NN) computation. However, the effective evaluation and design of these next-generation systems necessitate a thorough understanding of the complex interplay between their electronic and photonic components. In this work, we propose LightSim , a holistic and flexible modeling framework aimed at guiding the design and optimization of photonic computing systems. LightSim features a hierarchical and modular structure that provides users with a customizable and user-friendly interface. Its photonic-specific models offer in-depth insights into critical aspects like dataflow, mapping, and system performance. Furthermore, its fast modeling capability facilitates rapid exploration of the design space. LightSim supports co-design and fair architectural comparisons across all levels, thereby fostering the agile, iterative design process crucial for the development of high-performance and efficient photonic computing hardware. Validation across both incoherent and coherent photonic computing designs demonstrates a high correlation between our results and the reference data.