随着人工智能的兴起，人们对包含矩阵运算在内的计算性能期望不断增长。传统处理器受到能耗和带宽的限制，逐渐无法满足庞大的数据处理需求。光学矩阵-向量乘法器有潜力显著降低信息传输能耗，提高设备运行速度，扩大信息传输容量，因此被视为高性能深度神经网络加速硬件的候选²。根据原理，目前被报道的片上乘法器可以被分为三大类³：波分复用型、马赫曾德干涉仪型和多平面光转换型。前者往往需要多个激光器以支持多波长，而后两者的矩阵加载需要借助算法。

时尧成团队首次提出并实现了基于模分复用的片上矩阵向量乘法器。该乘法器包括模式复用器、多模功率分束器、模式解复用器、调制器矩阵和合束器，其中模式（解）复用器为基于模式演化的反锥形耦合器（ADC）⁴、多模功率分束器采用了交叉波导交叉区域对角斜嵌入亚波长光栅（SWG）反射镜结构⁵。论文在绝缘体上硅平台研制了一个4×4规模的演示型乘法器，结合片外光源和探测器实验演示了2-bit调制精度下该器件的动态响应。原理和实验表明：该器件无需复杂光源、支持一对一的矩阵加载，且未来也可以结合波长实现更大的规模。

图1 | a, 提出的4×4模分复用乘法器示意图；b, 封装芯片图；c, 显微图。

图2 | a, ADC结构示意图；b, 不同模式的有效折射率；c, ADC模式转换效率和光场分布图。

图3 | a, SWG结构示意图；b, 不同模式的传输光谱和光场分布图；c, 1 ×4分束器的结构图。

图4 | 驱动电压波形与调制器、乘法器的响应。

On-chip optical matrix-vector multiplier based on mode division multiplexing¹

With the rise of artificial intelligence, people's expectations for computing performance, including matrix operations, continue to grow. Traditional processors are limited by energy consumption and bandwidth, gradually becoming unable to meet the demands of extensive data processing. Optical matrix-vector multipliers have the potential to significantly reduce information transmission energy consumption, enhance device operating speed, and expand information transmission capacity, making them candidates for high-performance deep neural network acceleration hardware².

Based on principles, currently reported on-chip multipliers can be categorized into three types³: wavelength division multiplexing, Mach-Zehnder interferometer, and multiplane light conversion. The former often requires multiple lasers to support multiple wavelengths, while the latter two necessitate matrix loading with the help of algorithms.

The team led by Yaocheng Shi proposed and implemented an on-chip matrix-vector multiplier based on mode division multiplexing. This multiplier includes a mode multiplexer, multimode power splitter, mode demultiplexer, modulator array, and combiner. The mode (de) multiplexer is based on the asymmetric directional coupler (ADC)⁴, and the multimode power splitter uses the structure of multimode waveguide crossing with an obliquely embedded subwavelength grating (SWG)⁵. The research realized a 4x4-scale demonstrator on a silicon-on-insulator platform, demonstrating the device's dynamic response with 2-bit modulation precision using off-chip light source and photodetectors. The principle and experiments show that this device does not require complex light sources, supports one-one matrix loading, and can be scaled further with wavelength combinations in the future.

参考文献：

1. Ling, Q. et al. On-chip optical matrix-vector multiplier based on mode division multiplexing. Chip 2, 100061 (2023).

2. Feldmann, J. et al. Parallel convolutional processing using an integrated photonic tensor core. Nature 589, 52–58 (2021).

3. Zhou, H. et al. Photonic matrix multiplication lights up photonic accelerator and beyond. Light Sci. Appl. 11, 30 (2022).

4. Dai, D. et al. 10-channel Mode (de) multiplexer with dual polarizations. Laser Photonics Rev. 12, 1700109 (2018).

5. Xu, H., Dai, D. & Shi, Y. Ultra-broadband on-chip multimode power splitter with an arbitrary splitting ratio. OSA Contin. 3, 1212 (2020).

文章链接：

https://www.sciencedirect.com/science/article/pii/S2709472323000242

作者简介

时尧成，浙江大学教授/博士生导师，国家优秀青年科学基金获得者。现为光电科学与工程学院副院长、国家光学仪器工程技术研究中心副主任。主要从事光通信器件，光互连、光传感用光子集成芯片的研究，在多通道片上高速光互连芯片、片上相控阵激光雷达扫描芯片、高性能波导微腔等领域取得了多项重要进展，近年来Nature Communications, Laser & Photonics Reviews、Optics Letters等国际权威期刊共发表SCI收录论文160余篇，被引用6000余次，h-index 42。受邀担任国际光电子与通信大会（OECC）、亚洲光通信与光电子学会议（ACP）、国际光通信与网络会议（ICOCN）等本领域重要国际会议的程序委员会共主席、组织委员会主席等10余次，同时，受邀在重要国际会议作特邀报告30余次，并获已授权国家发明专利12项。作为项目负责人承担国家自然科学基金重点项目、国家自然科学基金优秀青年科学基金、国家重点研发计划课题等国家级项目10余项，获浙江省自然科学一等奖、中国光学学会光学科技一等奖等奖项。

Yaocheng Shi received the Ph.D. degree from the Royal Institute of Technology (KTH), Stockholm, Sweden, in 2008. Then he joined the College of Optical Science and Engineering, Zhejiang University, where he is a professor. His current research interests include integrated photonics for optical communications, optical interconnections, and optical sensing. He has published about 150 refereed international journal papers and his papers have been cited by > 6000 times (Google Scholar).

关于Chip

Chip（ISSN：2772-2724，CN：31-2189/O4）是全球唯一聚焦芯片类研究的综合性国际期刊，已入选由中国科协、教育部、科技部、中科院等单位联合实施的「中国科技期刊卓越行动计划高起点新刊项目」，为科技部鼓励发表「三类高质量论文」期刊之一。

Chip期刊由上海交通大学出版，联合Elsevier集团全球发行，并与多家国内外知名学术组织展开合作，为学术会议提供高质量交流平台。

Chip秉承创刊理念: All About Chip，聚焦芯片，兼容并包，旨在发表与芯片相关的各科研领域尖端突破性成果，助力未来芯片科技发展。迄今为止，Chip已在其编委会汇集了来自14个国家的70名世界知名专家学者，其中包括多名中外院士及IEEE、ACM、Optica等知名国际学会终身会士（Fellow）。

Chip第二卷第四期将于2023年12月在爱思维尔Chip官网以金色开放获取形式（Gold Open Access）发布，欢迎访问阅读本期最新文章。

爱思唯尔Chip官网：

https://www.sciencedirect.com/journal/chip