Chongfan Technology
News
27
2025
-
03
Light | Cloud and Light: High-Density & Multi-Task Optical In-Memory Computing
Author:
Introduction
On-chip Optical Neural Network(on-chip ONN) has been extensively studied due to its high-bandwidth, low-loss, and low-latency computational characteristics and CMOS-compatible manufacturing process. However, challenges remain in simultaneously achieving high-density integration and multi-tasking capabilities.
The results are published inUltra-compact multi-task processor based on in-memory optical computinginLight: Science & ApplicationsTsinghua Universitydoctoral student Liu Wencanis the first author of the paper, and Tsinghua UniversityProfessor Chen Hongweiis the corresponding author.
Ultra-compact on-chip optical in-memory computing architecture
The research team, through innovative design of the diffractive optical computing chip, ensured that a single optical chip is no longer for a specific task, but aims to be a general processing unit for various task inputs. The optical chip serves as an in-memory computing unit with fixed weights, paired with a lightweight backend network to handle multi-tasking scenarios, achieving an "optical fixed weight + electrical lightweight network" computing paradigm. The team integrated most of the trained neural network weights as subwavelength diffractive elements into the chip. Therefore, during computation, only the input data needs to be modulated; accessing network weights from memory is unnecessary, and computation is completed during light propagation.
To achieve higher integration and computing density in the diffractive chip, while reducing energy consumption and improving model accuracy, the research team used a deep regression neural network in model simulation to replace traditional modeling of light propagation using Maxwell's equations. The team utilized a deep regression neural network with complex-valued parameters to simulate and predict the propagation of the light field in the chip. This significantly increased prediction speed while achieving accuracy comparable to simulation, and enabled a computing integration density as high as 60,000 parameters/mm².
Figure 1: Overall system processing flow chart and application schematic in edge computing architecture
For Multi-task Processing Applications
The research team demonstrated the system's performance on multi-classification and multi-regression tasks. For multi-classification tasks, classifying three independent datasets, the system achieved performance comparable to electronic networks while reducing electronic power consumption parameters by 90%. For multi-regression tasks, the system could highly fit multiple groups of convolutional kernels, and achieved higher computational accuracy with fewer electronic parameters in convolutional neural networks. Moreover, the proposed system can be used for edge applications, processing multi-task inputs through a series of internal optical chips for high-speed and low-power data processing. The processed data is then transmitted to a cloud database, and further lightweight computations are performed on various terminals based on specific task requirements.
Figure 2: System processing flowchart and experimental results comparison for multi-classification tasks
Summary and Outlook
This research proposes a novel on-chip optical multi-tasking processor architecture, improving the integration and reconfigurability of in-memory optical diffractive chips. By fixing most parameters within an ultra-compact, in-memory optical diffractive chip structure, it achieves multi-tasking capabilities while increasing the parameter integration density and computational capacity of current photonic chips by an order of magnitude. This work not only advances optical computing architectures but also holds great potential for future large-scale, high-performance computing systems.
Paper Information
Liu, W., Huang, Y., Sun, R., et al. Ultra-compact multi-task processor based on in-memory optical computing. Light Science & Applications 14, 134 (2025).
LATEST NEWS
Thank you for visiting the official website of Chongfan Technology. If you have cooperation intentions or suggestions, please contact us through the following methods, and we will reply as soon as possible, thank you!
Address: Room 403, Building 6, Phase III of R&D, No. 36 Xiyong Avenue, High tech Zone, Chongqing, China.
Telephone: +86-13658337211
E-mail: Sales@cfkeji.net
Website: www.cfkeji.net
Mobile Version
