Monolithically integrated asynchronous optical recurrent accelerator

Abstract

Computing with light is widely recognized as a promising paradigm for overcoming the energy and latency limitations of electronic computing. However, the energy consumption and latency in current optical computing hardware predominantly arise in the electrical domain rather than the optical domain, primarily due to frequent signal conversions between optical (analog) and electrical (digital) formats. Furthermore, as the operating frequency of optical computing surpasses the GHz range, the synchronization of parallel electrical signals and the management of optical delays become increasingly critical. These challenges exacerbate energy consumption and latency, particularly in recurrent optical operations. To address these limitations, we propose a novel asynchronous computing paradigm for on-chip optical recurrent accelerators based on wavelength encoding, effectively mitigating synchronization challenges. By leveraging the intrinsic causality of wavelength relay, our approach eliminates the need for rigorous temporal alignment. To demonstrate the flexibility and efficacy of this asynchronous paradigm, we present two advanced recurrent models—an optical hidden Markov model and an optical recurrent neural network—monolithically integrated for the first time. These models incorporate hundreds of linear and nonlinear computing units densely packed into a compact footprint of just 10 mm2. Experimental evaluations on various benchmark tasks underscore the superior energy efficiency and low latency of the proposed asynchronous optical accelerators. This innovation enables the efficient processing of large-scale parallel signals and positions optical processors as a pivotal technology for applications such as autonomous driving and intelligent robotics.