Chongfan Technology

News

All Categories >

Home page

2025

Light | High-performance Density Optical Tensor Processing Core

Author:

Introduction

Optical computing, as a novel computing paradigm, offers significant advantages such as high bandwidth, low latency, and low power consumption. However, the large unit device size and limited chip area of photonic integration restrict the chip's computing power. On the other hand, high-speed analog optical computing faces a serious bandwidth mismatch problem with low-speed digital electrical computing, and requires the use of high-power, high-speed analog-to-digital/digital-to-analog converters, leading to serious power consumption issues.

To address these issues,the research team of Professor Ming Li at the Institute of Semiconductors, Chinese Academy of Sciencesproposed a high-computing-density optical tensor processing unit architecture based on wavelength tuning and optical-wave microwave fusion multiplexing. This architecture achieves a record-breaking computing density of 34.04 TOPS/mm based on a single micro-ring resonator through wavelength tuning, taking a solid step towards the realization of high-computing-power optical computing chips; in addition, the team introduced microwave subcarriers through optical-wave microwave fusion multiplexing technology, dividing the large bandwidth of high-speed optical computing into several sub-bands. This leverages the advantages of optical computing's large bandwidth while improving the data input dimension, solving the problem of optical-electrical bandwidth mismatch and avoiding the use of high-power, high-speed analog-to-digital/digital-to-analog converters, effectively promoting the practical application of optical computing chips.²This achievement is titled “

High-integrated photonic tensor core utilizing high-dimensional lightwave and microwave multidomain multiplexing” and published inLight: Science & Applications。The Institute of Semiconductors, Chinese Academy of SciencesProfessor Ming LiAssociate Professor Nuan Nuan Shi、are the co-corresponding authors,Postdoctoral fellow Xiangyan Mengis the first author.Tensors, as carriers of high-dimensional data, form the computational basis of artificial neural networks. With the rapid evolution of deep learning technology, especially the breakthrough progress in generative artificial intelligence, the inter-layer connection dimensions of neural network models continue to expand, and the scale of parameters grows exponentially. In this context, the traditional electronic computing architecture, which separates storage and computing units, is facing systemic challenges such as high latency and high energy consumption. These hardware limitations have become key factors hindering the iteration of artificial intelligence technology.

The research team pioneered an optical tensor processing unit (OTPU) based on wavelength tuning and optical-wave microwave fusion multiplexing of a single micro-ring resonator

Figure 1，achieving ultra-high computing density while solving the problem of optical-electrical bandwidth mismatch. Figure 1: Optical Tensor Processing Unit

(1) Ultra-high computing density optical tensor computing

This architecture utilizes the periodic notch filter characteristics of micro-ring resonators (MRRs). It innovatively uses MRRs without thermal tuning to reconstruct convolution kernels via wavelength tuning. A single MRR can achieve the function of a traditional MRR array, thus achieving an ultra-high computing density of 34.04 TOPS/mm

Figure 2²). In addition, since the architecture uses MRRs without thermal tuning, compared with traditional thermally tuned MRR arrays, it avoids the serious thermal crosstalk problem in the process of thermal tuning MRRs, significantly reducing the complexity of chip control and reducing chip power consumption. Figure 2: (a) Physical image; (b) Microscopic image of the processing unit(2) Bandwidth-matched optoelectronic fusion tensor computing

This architecture uses optical-wave microwave fusion multiplexing and channelization technology to divide the broadband optical frequency spectrum (~hundreds of GHz) into several narrowband microwave sub-bands (~GHz). Each microwave sub-band serves as a data input channel for low-speed data input, thus effectively utilizing the advantages of optical computing's large bandwidth while solving the bandwidth mismatch problem between high-speed optical computing and low-speed electrical computing. This avoids the use of high-power, high-speed analog-to-digital/digital-to-analog converters and expands the data input dimension, enabling optical tensor operations.

Summary and Outlook

The research team proposed a novel high-computing-density optical tensor processing unit (OTPU) using MRRs that do not rely on thermal tuning as the core unit. Through wavelength division multiplexing technology, flexible weight control is achieved using wavelength tuning, achieving an ultra-high computing density of 34.04 TOPS/mm

This explores a new scheme for high-computing-power optical computing. By introducing optical-wave microwave fusion multiplexing technology, the data input dimension is improved to achieve optical tensor operations, solving the bandwidth mismatch problem between high-speed optical computing and low-speed electrical computing, and avoiding the dependence on high-power, high-speed analog-to-digital/digital-to-analog converters. This research result not only highlights the huge potential of optical computing in high-performance computing and high-dimensional data processing, but also provides a more efficient and energy-saving solution for complex computing tasks in artificial intelligence and data science.

Paper Information²Meng, X., Shi, N., Zhang, G. et al. High-integrated photonic tensor core utilizing high-dimensional lightwave and microwave multidomain multiplexing. Light Sci Appl 14, 27 (2025).

https://doi.org/10.1038/s41377-024-01706-9

Meng， X.， Shi， N.， Zhang， G. 等人。利用高维光波和微波多域多路复用的高集成光子张量核心。光科学应用14， 27 （2025）。

https://doi.org/10.1038/s41377-024-01706-9