Real-Valued MZI Mesh: A Simplified Solution for Efficient Optical Matrix Computing

Optical computing faces a challenge when it comes to implementing real-valued optical matrix-vector multiplication (MVM). While optical computing offers advantages like high bandwidth, low latency, and energy efficiency, existing methods for optical matrix computing are designed for complex-valued matrices. This leads to a redundancy of resources when dealing with real-valued matrices, which consumes extra energy and takes up more space. To address this issue, a new solution called the Real-Valued MZI Mesh for incoherent optical MVM has emerged.

The Significance of Real-Valued MZI Mesh

The Real-Valued MZI Mesh is a simplified and innovative approach to optical matrix computing. It reduces the number of phase shifters required to N^2 while maintaining an optical depth of N + 1. Instead of detecting the complex value of the output optical field, this method employs an extra port to perform optical power subtraction, resulting in a real-valued output. This not only simplifies the hardware requirements but also streamlines the detection process, overcoming the limitations of previous solutions.

The Performance of Real-Valued MZI Mesh

Extensive numerical evaluations utilizing particle swarm optimization (PSO) were conducted to assess the performance and viability of the Real-Valued MZI Mesh. The results demonstrated the mesh’s exceptional performance in benchmark tasks, highlighting its potential as an efficient solution for real-valued optical MVM in optical neural networks (ONNs). Furthermore, error analyses revealed the mesh’s resilience to fabrication errors, enhancing its reliability for practical applications.

To further emphasize the suitability of the mesh for large-scale ONNs, a matched on-chip nonlinear activation function was introduced in the study. With its space efficiency, energy efficiency, scalability, and robustness to fabrication errors, the Real-Valued MZI Mesh emerges as a promising solution to the challenges posed by real-valued optical matrix computing.

As the field of optical computing continues to evolve, this innovative approach holds significant promise for the future of large-scale ONNs and combination optimization problem solvers. It offers a more efficient and practical path forward.

Source link