Fully parallel optical matrix-matrix multiplication

Yufeng Zhang,Hao Yan,Kaizhi Wang

2023-09-19

Abstract:In recent years, with the rapid development of electro-optic modulators, optical computing has become a potential excellent candidate for various computing tasks. New structures and devices for optical computing are emerging one after another, but the computing method is still the optical vector-matrix multiplication method that was decades ago. Here, we propose a novel optical computing paradigm that can parallelly implement matrix-matrix multiplication operation, which can directly replace existing vector-matrix multiplication, greatly improving computational efficiency. This preprint presents theoretical analysis, and we will supplement experimental results and conclusions in the future.

Optics

What problem does this paper attempt to address?

The paper attempts to address the problem of achieving full parallelization of Matrix-Matrix Multiplication (MMM) in optical computing to significantly enhance computational efficiency. Specifically:

1. **Background and Motivation**:
- Matrix-Matrix Multiplication (MMM) is one of the core operations in computational applications, widely used in fields such as signal processing, image processing, and deep learning.
- Due to its time complexity of O(N³), MMM becomes the most time-consuming operation in various computational tasks.
- Optical computing is considered an ideal alternative to traditional digital computing due to its low latency and low power consumption characteristics.

2. **Limitations of Existing Methods**:
- Although early researchers proposed various methods to achieve Optical Vector-Matrix Multiplication (OVMM), these methods are still based on decades-old technology.
- These technologies, while improving device performance and combination strategies, have not fundamentally proposed more efficient optical computing principles.

3. **Innovations**:
- The paper proposes a novel fully parallelized optical matrix-matrix multiplication (POMMM) paradigm, completely transforming the traditional OVMM methods.
- The new method does not require pre-coding or pre-processing of the matrix and completes the computation through the propagation of light from a single light source and a single wavelength.
- This method has a simple yet sophisticated architecture, making it very suitable for accelerating Optical Neural Networks (ONNs) and other optical computing operations.

4. **Application Prospects**:
- POMMM can be used for parallel training and inference of multiple samples and neurons, adding an extra computational dimension to ONNs.
- Besides OVMM and 2-D convolution operations, this method also has the potential to become a new parallel computing paradigm in Fourier optics.

In summary, the paper aims to significantly improve the efficiency and performance of optical computing by proposing a novel fully parallelized optical matrix-matrix multiplication method.