Single-shot matrix-matrix photonic processor based on spatial-spectral hypermultiplexed parallel diffraction

Area of Science:

• Optoelectronics and Photonics
• Artificial Intelligence Hardware
• Computer Engineering

Background:

• Growing data demands necessitate high-speed, energy-efficient computing hardware.
• Analog optical neural network (ONN) processors offer advantages in bandwidth and power consumption.
• Existing ONN architectures face limitations in computational parallelism and scalability.

Purpose of the Study:

• To introduce a novel spatial-wavelength-temporal hyper-multiplexed ONN processor.
• To address the limitations of current ONN processors in terms of parallelism and scalability.
• To enable large-scale, high-performance optical tensor processing for deep learning.

Main Methods:

• Development of a hyper-multiplexed ONN architecture based on parallel diffractive beam routing.
• Demonstration of a 16x16 parallel diffractive beam routing system.
• Implementation of single-shot matrix-matrix multiplication for accelerating neural networks.

Main Results:

• Achieved a large-scale (16x16-by-16x16) optical tensor processor with high parallelism (4096 MACs/shot) and high speed (2 Gsa/s).
• Demonstrated benchmark image recognition using convolutional neural networks (CNNs) and deep neural networks (DNNs) in the optical domain.
• Operated with ultra-low optical energy consumption (≈20 attojoules/MAC) at 96.4% classification accuracy.

Conclusions:

• The proposed hyper-multiplexed ONN processor architecture is feasible for large-scale implementation.
• The system supports broad spectral and spatial bandwidths, enabling significant advancements in optical computing.
• This technology paves the way for highly efficient, large-scale optical computing for next-generation deep learning applications.