A Fast Precision Tuning Solution for Always-On DNN Accelerators

doi:10.1109/TCAD.2021.3089667

	A Fast Precision Tuning Solution for Always-On DNN Accelerators
	Wang, Ying 1,2; He, Yintao 1,2; Cheng, Long 1,2; Li, Huawei 1,2,3; Li, Xiaowei 1,2
	2022-05-01
发表期刊	IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS
ISSN	0278-0070
卷号	41 期号:5 页码:1236-1248
摘要	Due to the nonvolatility nature of resistive RAM (ReRAM), dynamic operations in the arrays contribute to a much larger portion of power in ReRAM-based neural networks than static power. To reduce the dynamic power of in-situ operations with neural parameters, precision-tuning is considered a viable approach of approximate computing to tradeoff excessive computation exactness for power and efficiency gains. However, the switching overhead of precision tuning in hardware severely impacts its effectiveness when the systems need to quickly react to the change of environment, user constraint or input quality. This work for the first time investigates the feasibility of agile precision tuning for neural network accelerators to benefit from approximate computing. The proposed computing in memory (CiM) CNN accelerators fully utilize the normally off characteristics of memristor crossbars to achieve instant network precision tuning without worrying about the model reloading penalty. The ReRAM-based accelerator, with the proposed neural parameter mapping policy and the novel mixed-model training method, induces negligible precision-switching latency and power consumption when compared with traditional variable precision accelerators. In evaluation with state-of-the-art workloads, the proposed ReRAM deep learning and neural network architecture saves 58.3%-62.47% area overhead over the baseline design. We also leverage the proposed ReRAM accelerator architecture to build a novel always-on key-word spotting (KWS) system. The KWS design can switch between different precision modes to capture the relevant sound with high accuracy. The experimental results show the precision-adjustable KWS architecture saves considerable operating energy when fed with realistic test-sets of audio data.
关键词	Computer architecture Neural networks Computational modeling Approximate computing Tuning Switches Microprocessors Always-on CNN computing-in-memory (CiM) resistive RAM
DOI	10.1109/TCAD.2021.3089667
收录类别	SCI
语种	英语
资助项目	National Natural Science Foundation of China[61874124] ; National Natural Science Foundation of China[61876173] ; Youth Innovation Promotion Association, CAS[2018138] ; National Key Research and Development Program of China[2018AAA0102700]
WOS研究方向	Computer Science ; Engineering
WOS类目	Computer Science, Hardware & Architecture ; Computer Science, Interdisciplinary Applications ; Engineering, Electrical & Electronic
WOS记录号	WOS:000784196800007
出版者	IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
引用统计	被引频次：1[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/2XEOYT63/18901
专题	中国科学院计算技术研究所期刊论文_英文
通讯作者	Li, Huawei; Li, Xiaowei
作者单位	1.Chinese Acad Sci, Inst Comp Technol, State Key Lab Comp Architecture, Beijing 100190, Peoples R China 2.Univ Chinese Acad Sci, Beijing 100190, Peoples R China 3.Peng Cheng Lab, Shenzhen, Peoples R China
推荐引用方式 GB/T 7714	Wang, Ying,He, Yintao,Cheng, Long,et al. A Fast Precision Tuning Solution for Always-On DNN Accelerators[J]. IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS,2022,41(5):1236-1248.
APA	Wang, Ying,He, Yintao,Cheng, Long,Li, Huawei,&Li, Xiaowei.(2022).A Fast Precision Tuning Solution for Always-On DNN Accelerators.IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS,41(5),1236-1248.
MLA	Wang, Ying,et al."A Fast Precision Tuning Solution for Always-On DNN Accelerators".IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS 41.5(2022):1236-1248.