Global
EN
Applications
Support
Support
With over a thousand cooperative customers and 17 years of service experience, we can provide you with everything from model selection to technical support
Development
Development
Our unyielding mission is to continuously innovate and lead the industry's progress.
News & Events
News & Events
We will share every little bit of our life with you at all times
About
About
Yinte Electronics integrates technology research and development, chip manufacturing, packaging and testing, sales, and service
Careers
Careers
Unleash potential together, shape a healthy future for humanity
News & Events
We will share every little bit of our life with you at all times
Corporate News Industry News Product Knowledge Downloads

Semiconductor Institute Develops Ultra-High Integration Optical Convolution Processor

Source:Yint Time:2023-06-04 Views:6679
Share:

近日,据中国科学院半导体研究所消息,半导体所集成光电子学国家重点实验室微波光电子课题组李明研究员-祝宁华院士团队研制出一款超高集成度光学卷积处理器。相关研究成果以“Compact optical convolution processing unit based on multimode interference”为题发表在《自然通讯》(Nature Communications)杂志上。

卷积神经网络是一种受生物视觉神经系统启发而发展起来的人工神经网络,它由多层卷积层、池化层和全连接层组成。作为卷积神经网络的核心组成部分,卷积层通过对输入数据进行局部感知和权值共享,提取出不同层次和抽象程度的特征。在一个完整的卷积神经网络中,卷积运算的运算量通常占整个网络运算量的80%以上。虽然卷积神经网络在图像识别等领域取得了巨大的成功,但是它也面临着巨大的挑战。传统的卷积神经网络主要基于冯·诺依曼架构的电学硬件实现,存储单元和处理单元是分立的,这导致了数据交换速度和能耗之间的固有矛盾。随着数据量和网络复杂度的增加,电子计算方案越来越难以满足海量数据实时处理对高速、低能耗的计算硬件的需求。

光学卷积处理器.jpg

光计算是一种利用光波作为载体进行信息处理的技术,它具有大带宽、低延时、低功耗等优点,提供了一种传输即计算,结构即功能的计算架构,有望避免冯·诺依曼计算范式中存在的数据潮汐传输问题。光计算在近年来受到了广泛关注,但大部分已报道的光计算方案中,光学元件的数量随着计算矩阵的规模呈二次增长趋势,这对光计算芯片规模扩展存在巨大挑战。

http://www.semi.ac.cn/xwdt/zhxw/202305/t20230530_6765011.html