基于ARM+DLP+SRIO的嵌入式智能计算系统研究

doi:10.12263/DZXB.20200415

电子学报 ›› 2021, Vol. 49 ›› Issue (3): 443-453.DOI: 10.12263/DZXB.20200415

基于ARM+DLP+SRIO的嵌入式智能计算系统研究

赵二虎, 吴济文, 查晶晶, 郭振, 徐勇军

中国科学院计算技术研究所专项技术研究中心, 北京 100190

收稿日期:2020-05-06 修回日期:2020-08-03 出版日期:2021-03-25
- 作者简介:
- 赵二虎 男,1985年生于河北邢台.现为中国科学院计算技术研究所高级工程师,专项技术研究中心智算平台研究组组长,在读博士研究生,主要研究方向为嵌入式智能计算系统.E-mail:zhaoerhu@ict.ac.cn;吴济文 男,1987年生于江西上饶.现为中国科学院计算技术研究所工程师,主要研究方向为嵌入式智能计算系统的软硬件协同优化.E-mail:wujiwen@ict.ac.cn;查晶晶 女,1994年生于河南周口.现为中国科学院计算技术研究所工程师,主要研究方向为边缘异构智能计算系统中的算法优化与加速.E-mail:zhajingjing@ict.ac.cn;郭振男,1992年生于陕西商洛.现为中国科学院计算技术研究所工程师,主要研究方向为嵌入式智能计算系统的算法移植与优化.E-mail:guozhen@ict.ac.cn;徐勇军 男,1979年生于安徽安庆.现为中国科学院计算技术研究所正研级高级工程师、博士生导师、专项技术研究中心主任,主要研究方向为数据智能.E-mail:xyj@ict.ac.cn
- 基金资助:
- "十三五"领域基金 (No.61403120111）

Embedded AI Computing System Based on ARM+DLP+SRIO

ZHAO Er-hu, WU Ji-wen, ZHA Jing-jing, GUO Zhen, XU Yong-jun

Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China

Received:2020-05-06 Revised:2020-08-03 Online:2021-03-25 Published:2021-03-25
- Supported by:
- 13th Five-year Plan Field Fund (No.61403120111)

摘要/Abstract

摘要： 以x86+GPU为代表的当前主流AI计算平台，受限于功耗、体积、带宽、环境适应性等因素，无法适用于物端及边缘智能计算场景.提出并研究了一种基于ARM+DLP+SRIO的嵌入式智能计算系统，从AI算力、能效比、IO带宽三个方面分析了所提嵌入式智能计算系统的设计思路和技术优势，并实验验证了该系统的功能及性能指标.实验结果表明：基于ARM+DLP+SRIO的嵌入式智能计算系统AI峰值算力达到114.9TOPS，能效比达到1.03TFLOPS/W，IO带宽达到20Gbps.在智能计算系统领域，其能效比优于国内其它已知同类板卡或系统，嵌入式环境适应能力优于传统台式机和服务器，可作为物端及边缘环境下AI计算任务的通用硬件加速平台.

关键词: 人工智能, 深度学习处理器, 嵌入式智能计算系统, 串行RapidIO, 能效比

Abstract: The existing artificial intelligent (AI) computing platform represented by x86+GPU, limited by power consumption, dimension, bandwidth, environmental adaptability, and other factors, cannot be well adapted to the things and edge intelligent computing scenarios. We proposed an embedded AI computing system based on ARM (Advanced RISC Machine) + DLP (Deep Learning Processor) + SRIO (Serial RapidIO), and elaborated the design methods and technical advantages. In study, three aspects of the system were dissertated: AI computing performance, power efficiency, and IO bandwidth, and the function and performance of the system were verified by experiments. The results show that the peak performance of the embedded AI computing system based on ARM+DLP+SRIO is up to 114.9TOPS, the energy efficiency is up to 1.03TFLOPS/W, and the IO bandwidth is up to 20Gbps. In the field of AI computing systems, its energy efficiency is better than other similar boards or systems in China, and its embedded environmental adaptability is better than that of traditional desktops and servers, so it can provide a general hardware acceleration platform for AI computing tasks in things and edge computing scenarios.

Key words: artificial intelligent, deep learning processor, embedded AI computing system, serial RapidIO, power efficiency

中图分类号:

TP389.1

赵二虎, 吴济文, 查晶晶, 等. 基于ARM+DLP+SRIO的嵌入式智能计算系统研究[J]. 电子学报, 2021, 49(3): 443-453.

ZHAO Er-hu, WU Ji-wen, ZHA Jing-jing, et al. Embedded AI Computing System Based on ARM+DLP+SRIO[J]. Acta Electronica Sinica, 2021, 49(3): 443-453.

参考文献

[1] 彭晓晖,张星洲,王一帆等.Web使能的物端计算系统[J].计算机研究与发展,2018,55(3):572-584. Peng Xiaohui,Zhang Xingzhou,Wang Yifan,et al.Web enabled things computing system[J].Journal of Computer Research and Development,2018,55(3):572-584.(in Chinese)
[2] 彭晓晖,徐志伟.基于边缘计算的物端系统挑战与愿景[J].中兴通讯技术,2019,25(3):31-36,57. Peng Xiaohui,Xu Zhiwei.Challenges and vision of things system in edge computing[J].ZTE Technology Journal,2019,25(3):31-36,57.(in Chinese)
[3] 徐志伟,曾琛,朝鲁,彭晓辉.面向控域的体系结构:一种智能万物互联的体系结构风格[J].计算机研究与发展,2019,56(1):90-102. Xu Zhiwei,Zeng Chen,Chao Lu,Peng Xiaohui.Zone-oriented architecture:an architectural style for smart web of everything[J].Journal of Computer Research and Development,2019,56(1):90-102.(in Chinese)
[4] 赵梓铭,刘芳,蔡志平,肖侬.边缘计算:平台、应用与挑战[J].计算机研究与发展,2018,55(2):327-337. Zhao Ziming,Liu Fang,Cai Zhiping,Xiao Nong.Edge computing:platforms,applications and challenges[J].Journal of Computer Research and Development,2018,55(2):327-337.(in Chinese)
[5] Shi W,Cao J,Zhang Q,et al.Edge computing:vision and challenges[J].Internet of Things Journal,IEEE,2016,3(5):637-646.
[6] 施巍松,孙辉,曹杰,张权,刘伟.边缘计算:万物互联时代新型计算模型[J].计算机研究与发展,2017,54(5):907-924. Shi Weisong,Sun Hui,Cao Jie,Zhang Quan,Liu Wei.Edge computing-an emerging computing model for the internet of everything era[J].Journal of Computer Research and Development,2017,54(5):907-924.(in Chinese)
[7] 陈云霁,李玲,李威,郭崎,杜子东.智能计算系统[M].北京:机械工业出版社,2020.
[8] MLU100简介[EB/OL].http://forum.cambricon.com/index.php?m=content&c=index&a=show&catid=50&id=127,2019-07-11.
[9] X Zhang,X Zhou,M Lin,J Sun.ShuffleNet:an extremely efficient convolutional neural network for mobile devices[A].2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition[C].Piscataway,NJ:IEEE Press,2018.6848-6856.
[10] A G Howard,M Zhu,B Chen,D Kalenichenko,W Wang,T Weyand,M Andreetto,H Adam.MobileNets:Efficient Convolutional Neural Networks for Mobile Vision Applications[EB/OL].https://arxiv.org/abs/1704.04861,2020-07-27.
[11] 王齐.PCI Express体系结构导读[M].北京:机械工业出版社,2010.
[12] PCI Express Base Specification Revision 3.1a[EB/OL].https://pcisig.com/specifications/pciexpress/,2015-10-07.
[13] 王勇,林粤伟,吴冰冰.RapidIO嵌入式系统互连[M].北京:电子工业出版社,2006.
[14] RapidIO Specification 2.2[EB/OL].http://www.rapidio.org/rapidio-specifications/,2011-05.
[15] 刘军伟.多种高速串行总线的对比研究与分析[J].电子测试,2016,(3):43-45. Liu Junwei.Comparative study and analysis of multiple high speed serial bus[J].Electronic Test,2016,(3):43-45.(in Chinese)
[16] 李超.RapidIO高速互联技术研究[J].现代导航,2017,(5):385-390. Li Chao.Research of high-speed RapidIO interconnection technology[J].Modern Navigation,2017,(5):385-390.(in Chinese)
[17] 寒武纪端云一体人工智能开发平台白皮书[EB/OL].http://forum.cambricon.com/list-79-1.html,2020-07-29.
[18] 寒武纪软件开发环境[EB/OL].http://forum.cambricon.com/index.php?m=content&c=index&a=show&catid=50&id=128,2019-07-11.
[19] 刘彦,付彬,李仁发.高性能嵌入式计算[M].北京:机械工业出版社,2016.
[20] 赵二虎,徐勇军,吴济文,安竹林,李超.用于人工智能处理器的数据互联方法、系统、芯片和装置[P].中国专利申请号:201910826850.X,2020-01-17.
[21] 赵二虎,徐勇军,吴济文,李超,安竹林.面向物端数据处理的嵌入式智能计算机架构[P].中国专利申请号:201910605382.3,2020-01-24.
[22] 许家麟,韩思齐,孙宁霄,吴琼之.硬件加速系统中的PCIe-SRIO桥技术[J].电子设计工程,2017,25(15):189-193. Xu Jialin,Han Siqi,Sun Ningxiao,Wu Qiongzhi.PCIe-RapidIO bridge for hardware acceleration systems[J].Electronic Design Engineering,2017,25(15):189-193.(in Chinese)
[23] 李红兵.Linux系统下PCIE to RapidIO桥驱动设计与实现[J].雷达与对抗,2018,38(2):55-58,68. Li Hongbing.Driver design and implementation of PCIE to RapidIO bridge under Linux operating system[J].Radar & ECM,2018,38(2):55-58,68.(in Chinese)
[24] K He,X Zhang,S Ren,J Sun.Deep residual learning for image recognition[A].2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)[C].Piscataway,NJ:IEEE Press,2016.770-778.
[25] P Molchanov,S Tyree,T Karras,T Aila,J Kautz.Pruning convolutional neural networks for resource efficient inference[A].Proceedings of the 5th International Conference on Learning Representations[C].Toulon,France:ICLR,2017.
[26] K Simonyan,A Zisserman.Very deep convolutional networks for large-scale image recognition[A].Proceedings of the 3rd International Conference on Learning Representations[C].San Diego,CA,USA:ICLR,2015.
[27] J Redmon,A Farhadi.Yolov3:An Incremental Improvement[EB/OL].https://arxiv.org/abs/1804.02767,2018/04/08.
[28] 云燧T10人工智能训练加速卡[EB/OL].http://site-static.enflame-tech.com/enflame_back/public/uploads/c_video_img/%E4%BA%91%E7%87%A7T10%E5%8D%95%E9%A1%B52020-4-7.pdf,2020/07/29.
[29] 天数智芯边缘计算系统板硬件说明书[EB/OL].http://iluvatar.ai/filedownload/34977,2019/11/19.
[30] Habana GOYA Inference Card Specifications[EB/OL].https://habana.ai/wp-content/uploads/2019/06/Goya-Datasheet-HL-10x.pdf,2019/07.
[31] Tearing Apart GOOGLE's TPU 3.0 AI Coprocessor[EB/OL].https://developer.nvidia.com/deep-learning-performance-training-inference,2020/07/07.
[32] PreferredNetworks MN-Core Overview[EB/OL].https://projects.preferred.jp/mn-core/en/#mn-core,2020/05/06.
[33] 昇腾310 AI处理器技术规格[EB/OL].https://e.huawei.com/cn/products/cloud-computing-dc/atlas/ascend-310,2020/05/06.
[34] 地平线自动驾驶边缘AI芯片Journey征程[EB/OL].https://www.horizon.ai/product/journey,2020/05/06.
[35] 地平线AIoT边缘AI芯片Sunrise旭日[EB/OL].https://www.horizon.ai/product/sunrise,2020/05/06.
[36] 昇腾910 AI处理器规格参数[EB/OL].https://e.huawei.com/cn/products/cloud-computing-dc/atlas/ascend-910,2020/05/06.
[37] 百度自研AI芯片昆仑解读[EB/OL].http://www.elecfans.com/d/1196758.html,2020/04/03.
[38] 百度云端AI芯片昆仑[EB/OL].https://cloud.baidu.com/video-center/video.html?id=213,2020/05/06.
[39] L Gwennap.Groq Rocks Neural Networks Startup Creates New Architecture:One Core,1,000 TOPS[EB/OL].https://www.linleygroup.com/mpr/article.php?id=12245,2020/05/06.
[40] Z Jia,B Tillman,M Maggioni,D P Scarpazza.Dissecting thegraphcore IPU architecture via microbenchmarking[EB/OL].https://arxiv.org/abs/1912.03413v1,2019/10/07.
[41] 比特大陆Sophon AI计算加速卡SC5+[EB/OL].https://sophon.cn/product/introduce/sc5-plus.html,2020/05/06.
[42] NVIDIA Data Center Deep Learning Product Performance[EB/OL].https://developer.nvidia.com/deep-learning-performance-training-inference,2020/07/07.
[43] 刘升财.ATCA平台上的RapidIO链路设计与分析[D].成都:电子科技大学,2014.
[44] NVIDIA. NVIDIA Tesla P100 GPU Accelerator Datasheet[EB/OL]. https://www.nvidia.cn/content/dam/en-zz/Solutions/Data-Center/tesla-p100/pdf/nvidia-tesla-p100-datasheet.pdf,2016-10-16.

基于ARM+DLP+SRIO的嵌入式智能计算系统研究

Embedded AI Computing System Based on ARM+DLP+SRIO

PDF

可视化

摘要/Abstract

引用本文

使用本文

参考文献

相关文章 10

编辑推荐

Metrics

[1]	张亚洲, 俞洋, 朱少林, 陈锐, 戎璐, 梁辉. 一种量子概率启发的对话讽刺识别网络模型[J]. 电子学报, 2022, 50(8): 1885-1893.
[2]	张笑宇, 沈超, 蔺琛皓, 李前, 王骞, 李琦, 管晓宏. 面向机器学习模型安全的测试与修复[J]. 电子学报, 2022, 50(12): 2884-2918.
[3]	孙鹏浩, 兰巨龙, 申涓, 胡宇翔. 一种基于深度增强学习的智能路由技术[J]. 电子学报, 2020, 48(11): 2170-2177.
[4]	王寿成, 严迎建, 徐进辉. 基于流体系结构的高效能分组密码处理器研究[J]. 电子学报, 2017, 45(4): 937-943.
[5]	王鹏, 张长胜, 张斌, 吴嘉轩, 刘婷婷. 一种基于双空间密度的多目标进化算法[J]. 电子学报, 2017, 45(10): 2343-2347.
[6]	徐晓滨;文成林;刘荣利. 基于随机集理论的多源信息统一表示与建模方法[J]. 电子学报, 2008, 36(6): 1174-1181.
[7]	钟义信. 机制主义:人工智能的统一理论[J]. 电子学报, 2006, 34(2): 317-321.
[8]	林亚平, 刘云中, 周顺先, 陈治平, 蔡立军. 基于最大熵的隐马尔可夫模型文本信息抽取[J]. 电子学报, 2005, 33(2): 236-240.
[9]	强宇, 刘宗田, 林炜, 时百胜, 李云. 模糊概念格在知识发现的应用及一种构造算法[J]. 电子学报, 2005, 33(2): 350-353.
[10]	王继成. 基于认知模拟的符号-神经网络系统的研究[J]. 电子学报, 2000, 28(11): 99-101.