CAS Researchers Advance Multidimensional Multiplexed Optical Computing Architecture
Researchers at the CAS Shanghai Institute of Optics and Fine Mechanics have developed a multidimensional multiplexed optical computing architecture published in Laser & Photonics Reviews. Building on their 'Meteor-1' integrated chip, the team achieved 108 parallel wavelength channels per core across a three-core system, delivering a combined compute parallelism of 324. This architecture demonstrates how fusing wavelength-division and space-division multiplexing can significantly scale optical computing density and parallelism for AI workloads. It offers a promising hardware pathway to overcome electronic interconnect bandwidth bottlenecks and energy constraints in complex tasks like real-time remote sensing image processing. The architecture allows multiple tasks to share front-end optical networks while using dedicated back-ends, demonstrating matrix network sharing across two distinct tasks. In optical convolution experiments across broad spectrums (Diagonal, Sobel-x, Sobel-y operators), the system maintained a root-mean-square error (RMSE) below 0.08 compared to electronic benchmarks and achieved an 84.5% accuracy in remote sensing image registration and recognition.
## BACKGROUND
Optical computing uses photons instead of electrons to perform mathematical operations such as matrix multiplication at high speeds with low energy consumption. Photonic tensor cores leverage multiplexing techniques, such as wavelength-division multiplexing, to process multiple data streams simultaneously on a single chip, offering higher throughput than traditional silicon microprocessors.