基于光梳的单片光子电子加速器用于自注意计算

IF 4.3 2区工程技术 Q1 ENGINEERING, ELECTRICAL & ELECTRONIC IEEE Journal of Selected Topics in Quantum Electronics Pub Date : 2024-07-09 DOI:10.1109/JSTQE.2024.3425456

Tzu-Chien Hsueh;Yeshaiahu Fainman;Bill Lin

{"title":"基于光梳的单片光子电子加速器用于自注意计算","authors":"Tzu-Chien Hsueh;Yeshaiahu Fainman;Bill Lin","doi":"10.1109/JSTQE.2024.3425456","DOIUrl":null,"url":null,"abstract":"This paper adopts advanced monolithic silicon-photonics integrated-circuits manufacturing capabilities to realize system-on-chip photonic-electronic linear-algebra accelerators for self-attention computation in various applications of deep-learning neural networks and Large Language Models. With the features of holistic co-design approaches, optical comb-based broadband modulations, and consecutive matrix-multiplication architecture, the system/circuit/device-level simulations of the proposed accelerator can achieve 2.14-TMAC/s/mm\n<sup>2</sup>\n computation density and 27.9-fJ/MAC energy efficiency with practical considerations of power/area overhead due to photonic-electronic on-chip conversions, integrations, and calibrations.","PeriodicalId":13094,"journal":{"name":"IEEE Journal of Selected Topics in Quantum Electronics","volume":"30 5: Microresonator Frequency Comb Technologies","pages":"1-17"},"PeriodicalIF":4.3000,"publicationDate":"2024-07-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Optical Comb-Based Monolithic Photonic-Electronic Accelerators for Self-Attention Computation\",\"authors\":\"Tzu-Chien Hsueh;Yeshaiahu Fainman;Bill Lin\",\"doi\":\"10.1109/JSTQE.2024.3425456\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"This paper adopts advanced monolithic silicon-photonics integrated-circuits manufacturing capabilities to realize system-on-chip photonic-electronic linear-algebra accelerators for self-attention computation in various applications of deep-learning neural networks and Large Language Models. With the features of holistic co-design approaches, optical comb-based broadband modulations, and consecutive matrix-multiplication architecture, the system/circuit/device-level simulations of the proposed accelerator can achieve 2.14-TMAC/s/mm\\n<sup>2</sup>\\n computation density and 27.9-fJ/MAC energy efficiency with practical considerations of power/area overhead due to photonic-electronic on-chip conversions, integrations, and calibrations.\",\"PeriodicalId\":13094,\"journal\":{\"name\":\"IEEE Journal of Selected Topics in Quantum Electronics\",\"volume\":\"30 5: Microresonator Frequency Comb Technologies\",\"pages\":\"1-17\"},\"PeriodicalIF\":4.3000,\"publicationDate\":\"2024-07-09\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"IEEE Journal of Selected Topics in Quantum Electronics\",\"FirstCategoryId\":\"5\",\"ListUrlMain\":\"https://ieeexplore.ieee.org/document/10591337/\",\"RegionNum\":2,\"RegionCategory\":\"工程技术\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"ENGINEERING, ELECTRICAL & ELECTRONIC\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"IEEE Journal of Selected Topics in Quantum Electronics","FirstCategoryId":"5","ListUrlMain":"https://ieeexplore.ieee.org/document/10591337/","RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"ENGINEERING, ELECTRICAL & ELECTRONIC","Score":null,"Total":0}

引用次数: 0

摘要

本文采用先进的单片硅光子集成电路制造能力，实现了片上系统光子电子线性代数加速器，用于深度学习神经网络和大型语言模型等各种应用中的自注意计算。利用整体协同设计方法、基于光梳的宽带调制和连续矩阵乘法架构等特点，通过系统/电路/器件级仿真，拟议加速器可实现 2.14-TMAC/s/mm2 的计算密度和 27.9-fJ/MAC 的能效，并实际考虑了光子-电子片上转换、集成和校准带来的功率/面积开销。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

查看原文

微信好友朋友圈 QQ好友复制链接

本刊更多论文

Optical Comb-Based Monolithic Photonic-Electronic Accelerators for Self-Attention Computation

This paper adopts advanced monolithic silicon-photonics integrated-circuits manufacturing capabilities to realize system-on-chip photonic-electronic linear-algebra accelerators for self-attention computation in various applications of deep-learning neural networks and Large Language Models. With the features of holistic co-design approaches, optical comb-based broadband modulations, and consecutive matrix-multiplication architecture, the system/circuit/device-level simulations of the proposed accelerator can achieve 2.14-TMAC/s/mm ² computation density and 27.9-fJ/MAC energy efficiency with practical considerations of power/area overhead due to photonic-electronic on-chip conversions, integrations, and calibrations.

求助全文

通过发布文献求助，成功后即可免费获取论文全文。去求助

来源期刊

IEEE Journal of Selected Topics in Quantum Electronics 工程技术-工程：电子与电气

CiteScore

10.60

自引率

2.00%

发文量

212

审稿时长

3 months

期刊介绍： Papers published in the IEEE Journal of Selected Topics in Quantum Electronics fall within the broad field of science and technology of quantum electronics of a device, subsystem, or system-oriented nature. Each issue is devoted to a specific topic within this broad spectrum. Announcements of the topical areas planned for future issues, along with deadlines for receipt of manuscripts, are published in this Journal and in the IEEE Journal of Quantum Electronics. Generally, the scope of manuscripts appropriate to this Journal is the same as that for the IEEE Journal of Quantum Electronics. Manuscripts are published that report original theoretical and/or experimental research results that advance the scientific and technological base of quantum electronics devices, systems, or applications. The Journal is dedicated toward publishing research results that advance the state of the art or add to the understanding of the generation, amplification, modulation, detection, waveguiding, or propagation characteristics of coherent electromagnetic radiation having sub-millimeter and shorter wavelengths. In order to be suitable for publication in this Journal, the content of manuscripts concerned with subject-related research must have a potential impact on advancing the technological base of quantum electronic devices, systems, and/or applications. Potential authors of subject-related research have the responsibility of pointing out this potential impact. System-oriented manuscripts must be concerned with systems that perform a function previously unavailable or that outperform previously established systems that did not use quantum electronic components or concepts. Tutorial and review papers are by invitation only.