This paper demonstrates a novel approach to recurrent neural networks using the natural spatiotemporal dynamics of light in multimode optical fibers. The system processes temporal data with high energy efficiency by leveraging passive light propagation and inherent optical nonlinearities, achieving competitive performance across diverse temporal and spatiotemporal learning tasks without trainable optical parameters.

This paper introduces SKYLIGHT, a scalable 3D photonic in-memory tensor core architecture that overcomes existing limitations in photonic computing through innovations in topology, wavelength routing, accumulation, and programming. It demonstrates superior throughput and energy efficiency for real-time AI inference compared to traditional electronic and prior photonic accelerators.