Recent advances in generative AI have created a pressing need for next-generation computing technologies capable of processing massive volumes of data with both high speed and energy efficiency. At the heart of many AI applications, particularly deep learning, lies matrix–vector multiplication (MVM)—a fundamental operation that strongly influences the overall performance of AI systems. While the continued scaling of electronic circuits has historically driven improvements in computational power, the diminishing returns of Moore’s Law have raised concerns about the long-term viability of electronic processors as the sole foundation for future AI workloads. In response, optical computing has emerged as a promising alternative, offering inherently low transmission losses and the ability to perform high-throughput, parallel signal processing—qualities that are well suited to accelerating MVM operations, a key bottleneck in AI processing pipelines.
To date, optical MVM circuits have primarily relied on wavelength or mode multiplexing to scale performance. However, these approaches face practical challenges, such as significant optical losses and inter-channel crosstalk, which limit their scalability in large-scale systems.
In this study, the researchers demonstrated a novel architecture that employs multiport photodetectors to sum signals from multiple optical waveguides—without relying on wavelength or mode multiplexing. This approach enables a waveguide-multiplexed optical MVM circuit with a simpler design and greater scalability. It demonstrates the feasibility of using spatial multiplexing across optical waveguides—an underutilized degree of freedom in previous optical computing architectures. Proof-of-concept demonstrations were performed on a compact silicon photonic chip fabricated by a commercial photonic foundry.
Looking ahead, the integration of waveguide multiplexing with existing wavelength and mode multiplexing techniques is expected to enable massively parallel optical processors. This work represents a critical step toward the development of advanced computing platforms that combine the speed of optics with the scalability of silicon photonics, paving the way for future information processing systems that are both high-performance and energy-efficient.
Schematic diagram of the waveguide-multiplexed photonic matrix-vector multiplication processor
Papers
Journal: Optica
Title: Waveguide-multiplexed photonic matrix-vector multiplication processor using multiport photodetectors
Authors: Rui Tang, Makoto Okano, Chao Zhang, Kasidit Toprasertpong, Shinichi Takagi, and Mitsuru Takenaka*