Programmable matrix operation with reconfigurable time-wavelength plane manipulation and dispersed time delay

Area of Science:

• Optics and Photonics
• Computer Science
• Electrical Engineering

Background:

• Matrix manipulation is crucial for complex computations in fields like deep learning.
• Existing electronic computing architectures face limitations in speed and energy efficiency for massive parallel processing.

Purpose of the Study:

• To propose and validate a novel optical computing architecture for massive parallel matrix manipulation.
• To demonstrate the architecture's capability in high-speed signal processing and image analysis tasks.

Main Methods:

• Utilizing reconfigurable time-wavelength plane manipulation and dispersed time delay for optical computation.
• Implementing linear weighting methods in both wavelength and time domains.
• Performing autocorrelation, matrix-vector multiplication, and 2D convolution using the optical architecture.

Main Results:

• Achieved 1.18×10^11 multiplications and accumulations per second (MACs/s) for m-sequence autocorrelation.
• Demonstrated 2.69×10^9 MACs/s for 4x4 matrix-vector multiplication.
• Realized optical 2D convolution for edge extraction of 32x32 images at 5×10^8 MACs/s in simulation.

Conclusions:

• The proposed optical computing unit offers a promising building block for complex computational tasks and deep learning.
• This architecture is well-suited for future photonic neural network circuits, enabling significant advancements in processing speed and efficiency.