Meng Xiangyan, Shi Nuannuan, Zhang Guojie, Li Junshen, Jin Ye, Sun Shiyou, Shen Yichen, Li Wei, Zhu Ninghua, Li Ming
Key Laboratory of Optoelectronic Materials and Devices, Institute of Semiconductors, Chinese Academy of Sciences, Beijing, 100083, China.
College of Materials Science and Opto-Electronic Technology, University of Chinese Academy of Sciences, Beijing, 100049, China.
Light Sci Appl. 2025 Jan 3;14(1):27. doi: 10.1038/s41377-024-01706-9.
The burgeoning volume of parameters in artificial neural network models has posed substantial challenges to conventional tensor computing hardware. Benefiting from the available optical multidimensional information entropy, optical intelligent computing is used as an alternative solution to address the emerging challenges of electrical computing. These limitations, in terms of device size and photonic integration scale, have hindered the performance of optical chips. Herein, an ultrahigh computing density optical tensor processing unit (OTPU), which is grounded in an individual microring resonator (MRR), is introduced to respond to these challenges. Through the independent tuning of multiwavelength lasers, the operational capabilities of an MRR are orchestrated, culminating in the formation of an optical tensor core. This design facilitates the execution of tensor convolution operations via the lightwave and microwave multidomain hybrid multiplexing in terms of the time, wavelength, and frequency of microwaves. The experimental results for the MRR-based OTPU show an extraordinary computing density of 34.04 TOPS/mm. Additionally, the achieved accuracy rate in recognizing MNIST handwritten digits was 96.41%. These outcomes signify a significant advancement toward the realization of high-performance optical tensor processing chips.
人工神经网络模型中参数数量的迅速增加给传统张量计算硬件带来了巨大挑战。受益于可用的光学多维信息熵,光学智能计算被用作解决新兴的电计算挑战的替代解决方案。这些在器件尺寸和光子集成规模方面的限制阻碍了光学芯片的性能。在此,引入了一种基于单个微环谐振器(MRR)的超高计算密度光学张量处理单元(OTPU)来应对这些挑战。通过对多波长激光器的独立调谐,精心编排了MRR的运算能力,最终形成了一个光学张量核心。这种设计有助于通过光波和微波在时间、波长和频率方面的多域混合复用执行张量卷积运算。基于MRR的OTPU的实验结果显示出34.04 TOPS/mm的非凡计算密度。此外,在识别MNIST手写数字时达到的准确率为96.41%。这些成果标志着在实现高性能光学张量处理芯片方面取得了重大进展。