OMica-based High-Precision, Large-Scale Optoelectronic Parallel Matrix Computing System

Update time: 2023-06-12

Researchers have proposed a novel architecture, called the Optical Multi-Imaging-casting architecture (OMica), for high-precision, large-scale optoelectronic parallel matrix computing accelerators using diffraction-based beam splitter components.

This new architecture enables true parallel computing for optical matrix convolution, vector-matrix multiplication, and matrix multiplication, and has the potential for significant applications in large-scale matrix parallel computing acceleration for specialized purposes.

The research was led by Dr. Junjie Yu from the Shanghai Institute of Optics and Fine Mechanics (SIOM) of Chinese Academy of Sciences, and Prof Changhe Zhou from Jinan university. Results were published in Photonics Research on February 1st 2023 on and Optics Letter on March 31st 2023, respectively.

“The mainstream optoelectronic computing architectures can be roughly divided into two categories: planar integrated and free-space interconnected. Among them, planar integrated optoelectronic computing can only achieve one-dimensional vector-matrix multiplication and is limited by the difficulty of integrating photonic unit devices to expand computing power,” said Yu, one of the correspond ding authors of this paper. “Free-space interconnected optoelectronic computing has the inherent ability to control millions of pixels and thus it is expected to achieve much higher computing power.”

The research team creatively used high-quality beam splitting elements, Dammann gratings, to successfully construct a large-scale, high-precision optoelectronic matrix computing architecture and demonstrated matrix convolution of 10 × 10 and 20 × 20 with a computational precision of about 8 bits. Furthermore, the team optimized a Dammann grating with a beam splitting ratio of 182 × 224 and preliminarily demonstrated the potential of this method in acceleration of large-scale optical matrix convolution.

The team further studied temporal and spatial sequence coding methods to enable negative and complex number calculations, and implemented optical convolutional neural network inference tasks based on spatial sequence coding methods.

In addition, based on OMica, the researchers proposed an optical implementation architecture for acceleration of multiple channels vector-matrix multiplication, or matrix multiplication, and the team successfully demonstrated parallel computation acceleration for 8 × 4 and 4 × 8 matrix multiplication in the proof-of-concept experiment.

Furthermore, this OMica accelerator provides the possibility to operate under white light illumination conditions. Thus, it is expected to achieve direct processing of optical images directly from most of optical imaging system for real natural scenes, thereby breaking through the bottleneck of the electronic-optical conversion, at least at the input end. At the same time, this architecture can be further integrated through a planar waveguide optical system, and is expected to be practically applied in dedicated purpose computing acceleration for image data processing, machine vision, target recognition and other scenarios.

The researchers are now working to improve computing scale to realize optical computer with TOPS (Tera operations per seconds) computing power and are also seeking to further reduce volume and energy consumption of the whole system.


(Image by SIOM)

Article website:
https://doi.org/10.1364/OL.487676

Contact:
WU Xiufeng
General Administrative Office
Shanghai Institute of Optics and Fine Mechanics, CAS
Email: xfwu@siom.ac.cn
Web: http://english.siom.cas.cn/

附件下载:
Baidu
map