Breaking the Interaction Wall: A DLPU-Centric Deep Learning Computing System

doi:10.1109/TC.2020.3044245

	Breaking the Interaction Wall: A DLPU-Centric Deep Learning Computing System
	Du, Zidong 1,2; Guo, Qi 1; Zhao, Yongwei 1,2; Zeng, Xi 1,2,3; Li, Ling 4; Cheng, Limin 4; Xu, Zhiwei 1,3; Sun, Ninghui 1,3; Chen, Yunji 1,3,5
	2022
发表期刊	IEEE TRANSACTIONS ON COMPUTERS
ISSN	0018-9340
卷号	71 期号:1 页码:209-222
摘要	Due to the broad successes of deep learning, many CPU-centric artificial intelligent computing systems employ specialized devices such as GPUs, FPGAs, and ASICs, which can be named as Deep Learning Processing Units (DLPUs), for processing computation-intensive deep learning tasks. The separation between the scalar control operations mapped on CPUs and the vector computation operations mapped on DLPUs causes the frequent and costly interactions between CPUs and DLPUs, leading to the Interaction Wall. Moreover, the increasing algorithm complexity and DLPU computation speed would further aggravate the interaction wall substantially. To break the interaction wall, we propose a novel DLPU-centric deep learning computing system consisting of an exception-oriented programming (EOP) model and the architectural support of CPULESS DLPU. The EOP model processes scalar control operations of a deep learning task as exception handlers to maximally avoid stalling the crucial and dominated vector computation operations. Together with the CPULESS DLPU which integrates a scalar processing unit (SPU) for scalar control operations and the parallel processing unit (PPU) for vector computation operations into a fused pipeline, the proposed DLPU-centric system can cost-effectively leverage the EOP model to execute the two kinds of operations simultaneously without disturbing each other. Compared with a state-of-the-art commodity CPU-centric system with discrete V100 GPU via PCIe bus, experimental results show that our DLPU-centric system achieves 10.30x better performance and 92.99 percent energy savings, respectively. Moreover, compared with a CPU-centric version of DLPU system where the SPU serves as the host with integrated PPU, the proposed DLPU-centric system still achieves 15.60 percent better performance from avoided interactions.
关键词	Deep learning Central Processing Unit Process control Task analysis Computational modeling Pipelines Runtime Neural net accelerators system architectures interaction wall
DOI	10.1109/TC.2020.3044245
收录类别	SCI
语种	英语
资助项目	National Key Research and Development Program of China[2018AAA0103300] ; National Key Research and Development Program of China[2017YFA0700900] ; National Key Research and Development Program of China[2017YFA0700902] ; National Key Research and Development Program of China[2017YFA0700901] ; National Key Research and Development Program of China[2019AAA0103802] ; National Key Research and Development Program of China[2020AAA0103802] ; NSF of China[61532016] ; NSF of China[61732007] ; NSF of China[61672491] ; NSF of China[61732002] ; NSF of China[61925208] ; NSF of China[61702478] ; NSF of China[61732020] ; NSF of China[61906179] ; NSF of China[U19B2019] ; NSF of China[U20A20227] ; Beijing Natural Science Foundation[JQ18013] ; Key Research Projects in Frontier Science of Chinese Academy of Sciences[QYZDB-SSW-JSC001] ; Strategic Priority Research Program of Chinese Academy of Science[XDB32050200] ; Strategic Priority Research Program of Chinese Academy of Science[XDC01020000] ; Strategic Priority Research Program of Chinese Academy of Science[XDC05010300] ; Strategic Priority Research Program of Chinese Academy of Science[XDC08040102] ; Youth Innovation Promotion Association CAS ; Beijing Academy of Artificial Intelligence (BAAI) ; Beijing Nova Program of Science and Technology[Z191100001119093] ; Guangdong Science and Technology Program[2019B090909005]
WOS研究方向	Computer Science ; Engineering
WOS类目	Computer Science, Hardware & Architecture ; Engineering, Electrical & Electronic
WOS记录号	WOS:000730414800017
出版者	IEEE COMPUTER SOC
引用统计	被引频次：2[WOS] [WOS记录] [WOS相关记录]
文献类型	期刊论文
条目标识符	http://119.78.100.204/handle/2XEOYT63/17974
专题	中国科学院计算技术研究所期刊论文_英文
通讯作者	Zhao, Yongwei
作者单位	1.Chinese Acad Sci, Inst Comp Technol, State Key Lab Comp Architecture, Beijing 100190, Peoples R China 2.Cambricon Technol, Beijing 100191, Peoples R China 3.Univ Chinese Acad Sci, Beijing 100049, Peoples R China 4.Chinese Acad Sci, Inst Software, Beijing 100190, Peoples R China 5.Shanghai Res Ctr Brain Sci & Brain Inspired Intel, Inst Brain Intelligence Technol, Zhangjiang Lab BIT ZJLab, CAS Ctr Excellence Brain Sci & Intelligence Techn, Beijing 200031, Peoples R China
推荐引用方式 GB/T 7714	Du, Zidong,Guo, Qi,Zhao, Yongwei,et al. Breaking the Interaction Wall: A DLPU-Centric Deep Learning Computing System[J]. IEEE TRANSACTIONS ON COMPUTERS,2022,71(1):209-222.
APA	Du, Zidong.,Guo, Qi.,Zhao, Yongwei.,Zeng, Xi.,Li, Ling.,...&Chen, Yunji.(2022).Breaking the Interaction Wall: A DLPU-Centric Deep Learning Computing System.IEEE TRANSACTIONS ON COMPUTERS,71(1),209-222.
MLA	Du, Zidong,et al."Breaking the Interaction Wall: A DLPU-Centric Deep Learning Computing System".IEEE TRANSACTIONS ON COMPUTERS 71.1(2022):209-222.