Light-based Tensor Processing Enables Scalable Artificial Intelligence Performance.

Parallel Matrix Matrix Multiplication (POMMM) achieves fully parallel tensor processing via coherent light propagation, scaling performance with data dimension and demonstrating consistency with GPU-based matrix multiplication across real and complex domains. This adaptable and scalable approach enhances convolutional and vision transformer neural networks, offering improved theoretical power and efficiency.

The increasing demand for computational power in areas such as artificial intelligence and data analysis necessitates exploration beyond conventional electronic architectures. Current processing methods, largely optimised for scalar operations, encounter limitations when applied to the multi-dimensional data structures known as tensors, which are fundamental to modern machine learning. Researchers at Shanghai Jiao Tong University, the Chinese Academy of Sciences, and Aalto University, led by Yufeng Zhang, Xiaobing Liu, and Zhipei Sun et al, present a novel approach to address this challenge in their article, “Direct tensor processing with coherent light”. Their work details Parallel Optical Matrix Multiplication (POMMM), a paradigm utilising the propagation of coherent light to achieve fully parallel tensor processing, offering potential improvements in bandwidth, parallelism, and energy efficiency compared to existing electronic methods. The team demonstrates the consistency of POMMM with conventional GPU-based matrix multiplication and highlights its adaptability to complex neural network architectures, including convolutional and vision transformer networks.

Parallel Matrix Matrix Multiplication (POMMM) presents a novel approach to tensor processing, utilising the inherent parallelism of light to perform fully parallel computation via single coherent light propagation. This directly addresses limitations within current computational methods, which, while effective for scalar operations, struggle with the multi-dimensional nature of tensors increasingly vital for modern artificial intelligence and data analytics. Tensors are multi-dimensional arrays, generalisations of vectors and matrices, and are fundamental to representing and manipulating data in machine learning.

Validation confirms a high degree of consistency between POMMM and conventional Graphics Processing Unit (GPU)-based matrix multiplication, across both real and complex valued datasets. GPUs are specialised electronic circuits designed to rapidly manipulate and render images, and are now widely used for general-purpose computation due to their parallel processing capabilities. This consistency establishes POMMM as a viable alternative to established electronic computation, and its adaptability has been demonstrated through successful implementation in convolutional and vision transformer neural networks, showcasing its potential for broad application within machine learning. Convolutional neural networks are a type of deep learning model commonly used for image recognition, while vision transformers apply the transformer architecture, originally developed for natural language processing, to image data.

The system’s capacity to support multi-wavelength operation and large-scale expansion enhances its potential for future development and integration into advanced computing architectures. Multi-wavelength operation allows for increased computational density and throughput by utilising different wavelengths of light to perform separate computations simultaneously. This capability paves the way for more powerful and efficient tensor processing systems.

Researchers demonstrate that POMMM’s performance scales with data dimension, a significant advantage over existing methods. This scalability is validated by achieving high consistency between POMMM and GPU-based matrix multiplication, across both real and complex valued datasets. This consistency confirms the accuracy and reliability of the optical approach, and the method’s adaptability further showcases successful implementation in convolutional and vision transformer neural networks, demonstrating its versatility beyond basic matrix operations.

The Shanghai Institute of Optics and Fine Mechanics, part of the Chinese Academy of Sciences, provided crucial institutional support, specifically access to a multi-wavelength tunable filter, essential for the optical implementation. The research team comprised Y.Z., K.W., Y.S., Z.S., and X.G., who conceptualised the project, and Y.Z. and X.L., who developed the methodology. Investigation, visualisation, validation, software development, and writing were collaborative efforts involving several team members, highlighting a comprehensive and interdisciplinary approach to the research.

Theoretical analysis reveals that POMMM exhibits superior power efficiency and scalability compared to existing paradigms. Future work will focus on optimising the optical components and system architecture to maximise performance and minimise energy consumption. Exploration of advanced modulation techniques and novel optical materials will be crucial for achieving even greater computational density and speed.