高算力芯片未来技术发展途径

doi:10.3981/j.issn.2097-0781.2022.03.010

前瞻科技 ›› 2022, Vol. 1 ›› Issue (3): 115-129.DOI: 10.3981/j.issn.2097-0781.2022.03.010

高算力芯片未来技术发展途径

姚鹏(), 宋昌明, 胡杨, 蔡坚, 尹首一, 吴华强()

清华大学集成电路学院,北京 100084

收稿日期:2022-08-23 修回日期:2022-09-02 出版日期:2022-09-20 发布日期:2022-11-04
通讯作者: 吴华强
作者简介:姚鹏,助理研究员。研究方向为忆阻器阵列及其应用,存算一体宏电路设计,存算一体AI芯片架构和算法等。在Nature、Nature Communications等学术期刊和ISSCC、IEDM、VLSI等学术会议上发表多篇论文。电子信箱： yaop14@tsinghua.org.cn。
吴华强,教授,博士研究生导师。现任清华大学集成电路学院院长。国家杰出青年科学基金获得者。长期从事新型存储器与存算一体技术研究,并开展了从系统、电路到器件的多层次创新研究。先后负责国家自然科学基金、国家“863”计划、国家“973”计划、国家科技重大专项、国家重点研发计划等国家重大科技任务40余项。2021年入选“全球顶尖前10万科学家”榜单。获首届“科学探索奖”。在Nature、Nature Electronics等学术期刊和ISSCC、IEDM等学术会议上发表论文100余篇。电子信箱： wuhq@tsinghua.edu.cn。
基金资助:
科技创新2030-重大项目(2021ZD0201200);国家重点研发计划(2018YFB2202600)

Future Technical Development Approach for High Computing Power Chips

YAO Peng(), SONG Changming, HU Yang, CAI Jian, YIN Shouyi, WU Huaqiang()

School of Integrated Circuits, Tsinghua University, Beijing 100084, China

Received:2022-08-23 Revised:2022-09-02 Online:2022-09-20 Published:2022-11-04
Contact: WU Huaqiang

摘要/Abstract

摘要：

面向未来高算力芯片需求,分析了国内外高算力芯片发展趋势,提出由数据互连、单位晶体管提供的算力、晶体管密度和芯片面积构成的芯片算力表达式。介绍了未来高算力芯片发展的关键技术,并结合算力表达式论述相关技术如何发挥作用。从新材料、新器件、先进工艺、新架构、集成封装等角度出发,探讨了集成电路先进制造工艺、单片三维集成技术、领域专用架构、粗粒度可重构架构、存算一体技术、芯粒（Chiplet）技术和晶圆级集成等国内外发展现状及其对芯片算力的提升效果,并深入分析了各项技术的发展和挑战。结合中国高算力芯片现状和集成电路先进制程发展受限,提出从“架构+集成+系统”出发,探索实现高算力芯片的一体化自主可控创新路径,可以采用成熟制程,结合粗粒度可重构和存算一体新型架构,采用基于先进集成的芯粒技术实现总算力突破。

关键词: 高算力芯片, 先进制程, 新型架构, 粗粒度可重构, 存算一体, 芯粒

Abstract:

Targeting the future demand for high computing power chips, this paper analyzes the development trends of high computing power chips in China and abroad and further presents an expression of the computing power of a chip involving data interconnection, computing power per transistor, transistor density per unit area, and chip area. It then describes the key technologies for the future development of high computing power chips and explains how these technologies make a difference in accordance with the proposed expression of computing power. Specifically, the current status of these technologies, including advanced integrated circuit manufacturing, monolithic three-dimensional integration, domain-specific architectures, coarse-grained reconfigurable architecture (CGRA), in-memory computing (IMC), chiplets, and wafer-scale integration, in China and abroad and their performance in elevating the computing power of chips are discussed from the perspectives of new materials, new devices, advanced processes, novel architectures, and integrated packaging. The prospects of these technologies and the challenges they face are examined in depth. Regarding the current situation of high computing power chips in China and the restrictions on the development of advanced integrated circuit manufacturing processes, this paper proposes proceeding from the perspective of “architecture + integration + system” to explore an integrated, independently controllable, and innovative approach to high computing power chips. Mature manufacturing processes and novel architectures of CGRA and IMC, together with chiplets based on advanced integration, can be leveraged to achieve breakthroughs in total computing power.

Key words: high computing power chip, advanced manufacturing process, novel architecture, coarse-grained reconfigurable architecture, in-memory computing, chiplet

姚鹏, 宋昌明, 胡杨, 蔡坚, 尹首一, 吴华强. 高算力芯片未来技术发展途径[J]. 前瞻科技, 2022, 1(3): 115-129.

YAO Peng, SONG Changming, HU Yang, CAI Jian, YIN Shouyi, WU Huaqiang. Future Technical Development Approach for High Computing Power Chips[J]. Science and Technology Foresight, 2022, 1(3): 115-129.

图/表 11

图1 英伟达GPU算力与工艺制程的关系

表1 全球主要集成电路制造厂的量产工艺节点及晶体管密度

代工厂	国家/地区	量产工艺节点/nm	晶体管密度/（10⁶_• mm^-2）
台积电	中国台湾	5	196.6
中芯国际	中国大陆	14	30.0
英特尔	美国	7	106.1
三星	韩国	5	145.7

图2 工艺制程发展中的不同器件结构

图3 碳、硅和忆阻器单片三维异质集成系统

图4 传统计算芯片的架构特点

图5 基于忆阻器的存算一体技术原理

图6 芯粒系统芯片分解图

图7 冯·诺依曼架构瓶颈和“三墙”问题

图8 2.5D封装示意图

图9 3D封装示意图

图10 高算力芯片突破路径

参考文献 45

[1]	Hager G, Wellein G. Introduction to high performance computing for scientists and engineers[M]. Boca Raton: CRC Press, 2010.
[2]	Reuther A, Michaleas P, Jones M, et al. AI accelerator survey and trends[C]// 2021 IEEE High Performance Extreme Computing Conference (HPEC). Piscataway: IEEE Press, 2021, doi: 10.1109/HPEC49654.2021.9622867. DOI
[3]	Luo Y, Mirabbasi S. A 30-fps 192×192 CMOS image sensor with per-frame spatial-temporal coded exposure for compressive focal-stack depth sensing[J]. IEEE Journal of Solid-State Circuits, 2022, 57(6): 1661-1672. DOI URL
[4]	Waldrop M M. More than moore[J]. Nature, 2016, 530(7589): 144-148. DOI URL
[5]	Shalf J. The future of computing beyond Moore,s law[J]. Philosophical Transactions of the Royal Society A, 2020, doi: 10.1098/rsta.2019.0061. DOI
[6]	许海涛, 彭练矛. 碳基集成电路技术研究进展与展望[J]. 数据与计算发展前沿, 2021, 3(5): 4-27.
[7]	Yen A, Meiling H, Benschop J. Enabling manufacturing of sub-10 nm generations of integrated circuits with EUV lithography[C]// 2019 Electron Devices Technology and Manufacturing Conference (EDTM). Piscataway: IEEE Press, 2019, doi: 10.1109/EDTM.2019.8731058. DOI
[8]	Hisamoto D, Wen-chin L, Kedzierski J, et al. FinFET-a self-aligned double-gate MOSFET scalable to 20 nm[J]. IEEE Transactions on Electron Devices, 2000, 47(12): 2320-2325. DOI URL
[9]	Takato H, Sunouchi K, Okabe N, et al. High performance CMOS surrounding gate transistor (SGT) for ultra high density LSIs[C]// International Electron Devices Meeting. Piscataway: IEEE Press, 2002, doi: 10.1109/IEDM.1988.32796. DOI
[10]	Singh N, Agarwal A, Bera L K, et al. High-performance fully depleted silicon nanowire (diameter/spl les/5 nm) gate-all-around CMOS devices[J]. IEEE Electron Device Letters, 2006, 27(5): 383-386. DOI URL
[11]	Aly M M S, Gao M, Hills G, et al. Energy-efficient abundant-data computing: The N3XT 1,000x[J]. Computer, 2015, 48(12): 24-33.
[12]	Shulaker M M, Hills G, Park R S, et al. Three-dimensional integration of nanotechnologies for computing and data storage on a single chip[J]. Nature, 2017, 547(7661): 74-78. DOI URL
[13]	Bishop M D, Hills G, Srimani T, et al. Fabrication of carbon nanotube field-effect transistors in commercial silicon manufacturing facilities[J]. Nature Electronics, 2020, 3(8): 492-501. DOI URL
[14]	Liu Y, Zhang J, Peng L M. Three-dimensional integration of plasmonics and nanoelectronics[J]. Nature Electronics, 2018, 1(12): 644-651. DOI URL
[15]	Li Y, Tang J, Gao B, et al. Monolithic 3D integration of logic, memory and computing-in-memory for one-shot learning[C]// 2021 IEEE International Electron Devices Meeting. Piscataway: IEEE Press, 2022, doi: 10.1109/IEDM19574.2021.9720534. DOI
[16]	Choquette J, Gandhi W, Giroux O, et al. NVIDIA A100 tensor core GPU: Performance and innovation[J]. IEEE Micro, 2021, 41(2): 29-35.
[17]	Elster A C, Haugdahl T A. Nvidia hopper GPU and grace CPU highlights[J]. Computing in Science & Engineering, 2022, 24(2): 95-100.
[18]	Tu F, Wu Z, Wang Y, et al. A 28 nm 15.59 µJ/token full-digital bitline-transpose CIM-based sparse transformer accelerator with pipeline/parallel reconfigurable modes[C]// 2022 IEEE International Solid-State Circuits Conference (ISSCC). Piscataway: IEEE Press, 2022, doi: 10.1109/ISSCC42614.2022.9731645. DOI
[19]	Hennessy J L, Patterson D A. A new golden age for computer architecture[J]. Communications of the ACM, 2019, 62(2): 48-60. DOI
[20]	Liu L, Zhu J, Li Z, et al. A survey of coarse-grained reconfigurable architecture and design: Taxonomy, challenges, and applications[J]. ACM Computing Surveys (CSUR), 2019, 52(6): 1-39.
[21]	Tu F, Yin S, Ouyang P, et al. Deep convolutional neural network architecture with reconfigurable computation patterns[J]. IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 2017, 25(8): 2220-2233. DOI URL
[22]	Burstein I. Nvidia data center processing unit (DPU) architecture[C]// 2021 IEEE Hot Chips 33 Symposium (HCS). Piscataway: IEEE Press, 2021, doi: 10.1109/HCS52781.2021.9567066. DOI
[23]	Singh G, Chelini L, Corda S, et al. Near-memory computing: Past, present, and future[J]. Microprocessors and Microsystems, 2019, doi: 10.1016/j.micpro.2019.102868. DOI
[24]	Zhang W, Gao B, Tang J, et al. Neuro-inspired computing chips[J]. Nature Electronics, 2020, 3(7): 371-382. DOI URL
[25]	Knowles S. Graphcore[C]// 2021 IEEE Hot Chips 33 Symposium (HCS). Piscataway: IEEE Press, 2021, doi: 10.1109/HCS52781.2021.9567075. DOI
[26]	Lauterbach G. The path to successful wafer-scale integration: The cerebras story[J]. IEEE Micro, 2021, 41(6): 52-57. DOI URL
[27]	Ke L, Zhang X, So J, et al. Near-memory processing in action: Accelerating personalized recommendation with AxDIMM[J]. IEEE Micro, 2021, 42(1): 116-127.
[28]	Niu D, Li S, Wang Y, et al. 184 QPS/W 64 Mb/mm² 3D logic-to-DRAM hybrid bonding with process-near-memory engine for recommendation system[C]// 2022 IEEE International Solid-State Circuits Conference (ISSCC). Piscataway: IEEE Press, 2022, doi: 10.1109/ISSCC42614.2022.9731694. DOI
[29]	Lanza M, Sebastian A, Lu W D, et al. Memristive technologies for data storage, computation, encryption, and radio-frequency communication[J]. Science, 2022, 376(6597), doi: 10.1126/science.abj9979. DOI
[30]	Yao P, Wu H, Gao B, et al. Fully hardware-implemented memristor convolutional neural network[J]. Nature, 2020, 577(7792): 641-646. DOI URL
[31]	Liu Q, Gao B, Yao P, et al. 33.2 A fully integrated analog ReRAM based 78.4 TOPS/W compute-in-memory chip with fully parallel MAC computing[C]// 2020 IEEE International Solid-State Circuits Conference (ISSCC). Piscataway: IEEE Press, 2020, doi: 10.1109/ISSCC19947.2020.9062953. DOI
[32]	Xia J, Cheng C, Zhou X, et al. Kunpeng 920: The first 7-nm chiplet-based 64-Core ARM SoC for cloud services[J]. IEEE Micro, 2021, 41(5): 67-75.
[33]	Kim J, Murali G, Park H, et al. Architecture, chip, and package co-design flow for 2.5D IC design enabling heterogeneous IP reuse[C]// 2019 56th ACE/IEEE Design Automation Conference. Piscataway: IEEE Press, 2019, doi: 10.1109/TVLSI.2020.3015494. DOI
[34]	Li T, Hou J, Yan J, et al. Chiplet heterogeneous integration technology—Status and challenges[J]. Electronics, 2020, doi: 10.3390/electronics9040670. DOI
[35]	Ma X, Wang Y, Wang Y,, et al. Survey on chiplets: Interface, interconnect and integration methodology[J]. CCF Transactions on High Performance Computing, 2022, 4(5): 43-52. DOI URL
[36]	Gomes W, Koker A, Stover P, et al. Ponte vecchio: A multi-tile 3D stacked processor for exascale comput- ing[C]// 2022 IEEE International Solid-State Circuits Conference (ISSCC). Piscataway: IEEE Press, 2022, doi: 10.1109/ISSCC42614.2022.9731673. DOI
[37]	蒋剑飞, 王琴, 贺光辉, 等. Chiplet技术研究与展望[J]. 微电子学与计算机, 2022, 39(1): 1-6.
[38]	Douglas C, Wang C T, Hsia H. Foundry perspectives on 2.5D/3D integration and roadmap[C]// 2021 IEEE International Electron Devices Meeting (IEDM). Piscataway: IEEE Press, 2022, doi: 10.1109/IEDM19574.2021.9720568. DOI
[39]	杨晖. 后摩尔时代Chiplet技术的演进与挑战[J]. 集成电路应用, 2020, 37(5): 52-54.
[40]	陈桂林, 王观武, 胡健, 等. Chiplet封装结构与通信结构综述[J]. 计算机研究与发展, 2022, 59(1): 22-30.
[41]	Flack W, Flores G. Lithographic manufacturing techniques for wafer scale integration[C]// Proceedings International Conference on Wafer Scale Integration. Piscataway: IEEE Press, 2002: 4-13.
[42]	Cheng R, Bai J, Liao L, et al. High-frequency self-aligned graphene transistors with transferred gate stacks[J]. Proceedings of the National Academy of Sciences, 2012, 109(29): 11588-11592. DOI URL
[43]	Wu F, Tian H, Shen Y, et al. Vertical MoS₂ transistors with sub-1-nm gate lengths[J]. Nature, 2022, 603(7900): 259-264. DOI URL
[44]	Si J, Zhong D, Xu H, et al. Scalable preparation of high-density semiconducting carbon nanotube arrays for high-performance field-effect transistors[J]. ACS Nano, 2018, 12(1): 627-634. DOI PMID
[45]	Pitner G, Zhang Z, Lin Q, et al. Sub-0.5 nm interfacial dielectric enables superior electrostatics: 65 mV/dec top-gated carbon nanotube FETs at 15 nm gate length[C]// 2022 IEEE International Electron Devices Meeting (IEDM). Piscataway: IEEE Press, 2021, doi: 10.1109/IEDM13553.2020.9371899. DOI

高算力芯片未来技术发展途径

Future Technical Development Approach for High Computing Power Chips

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 11

参考文献 45

相关文章 1

编辑推荐

Metrics

访问统计

友情链接

联系我们

官方微信公众号