Challenges and Trends of Nonvolatile In-Memory-Computation Circuits for AI Edge Devices

被引：36
作者：
Hung, Je-Min [1 ]
Jhang, Chuan-Jia [1 ]
Wu, Ping-Chun [1 ]
Chiu, Yen-Cheng [1 ]
Chang, Meng-Fan [1 ]
机构：
[1] Institute of Electrical Engineering, National Tsing Hua University, Hsinchu
来源：
IEEE Open Journal of the Solid-State Circuits Society | 2021年 / 1卷
关键词：
Artificial intelligence; CIM; computation-in-memory; nonvolatile-memory; nvCIM; NVM;
D O I：
10.1109/OJSSCS.2021.3123287
中图分类号：
学科分类号：
摘要：
Nonvolatile memory (NVM)-based computing-in-memory (nvCIM) is a promising candidate for artificial intelligence (AI) edge devices to overcome the latency and energy consumption imposed by the movement of data between memory and processors under the von Neumann architecture. This paper explores the background and basic approaches to nvCIM implementation, including input methodologies, weight formation and placement, and readout and quantization methods. This paper outlines the major challenges in the further development of nvCIM macros and reviews trends in recent silicon-verified devices. © 2021 Institute of Electrical and Electronics Engineers Inc.. All rights reserved.
引用
收藏
页码：171 / 183
页数：12
相关论文
共 81 条
  • [1] Xue C.-X., Et al., 16.1 A 22nm 4Mb 8b-precision ReRAM computingin-memory macro with 11.91 to 195.7TOPS/W for tiny AI edge devices, Proc. IEEE Int. Solid-State Circuits Conf. (ISSCC), pp. 245-247, (2021)
  • [2] Xue C.-X., Et al., A CMOS-integrated compute-in-memory macro based on resistive random-access memory for AI edge devices, Nat. Electron., 4, pp. 81-90, (2021)
  • [3] Hu M., Et al., Dot-product engine for neuromorphic computing: Programming 1T1M crossbar to accelerate matrix-vector multiplication, Proc. 53nd ACM/EDAC/IEEE Design Autom. Conf. (DAC), pp. 1-6, (2016)
  • [4] Wang Z., Et al., An all-weights-on-chip DNN accelerator in 22nm ULL featuring 24×1 Mb eRRAM, Proc. IEEE Symp. VLSI Circuits, pp. 1-2, (2020)
  • [5] Yoon J.-H., Chang M., Khwa W.-S., Chih Y.-D., Chang M.-F., Raychowdhury A., 29.1 A 40nm 64Kb 56.67TOPS/W readdisturb-tolerant compute-in-memory/digital RRAM macro with activefeedback-based read and in-situ write verification, Proc. IEEE Int. Solid-State Circuits Conf. (ISSCC), pp. 404-406, (2021)
  • [6] Shafiee A., Et al., ISAAC: A convolutional neural network accelerator with in-situ analog arithmetic in crossbars, Proc. ACM/IEEE 43rd Annu. Int. Symp. Comput. Architect. (ISCA), pp. 14-26, (2016)
  • [7] Eryilmaz S.B., Kuzum D., Yu S., Wong H.S.P., Device and system level design considerations for analog-non-volatilememory based neuromorphic architectures, IEEE IEDM Tech. Dig., (2015)
  • [8] Yu S., Et al., Binary neural network with 16 Mb RRAM macro chip for classification and online training, Proc. IEEE Int. Electron Devices Meeting (IEDM), pp. 1621-1624, (2016)
  • [9] Bocquet M., Et al., In-memory and error-immune differential RRAM implementation of binarized deep neural networks, IEEE IEDM Tech. Dig., (2018)
  • [10] Adam G.C., Hoskins B.D., Prezioso M., Merrikh-Bayat F., Chakrabarti B., Strukov D.B., 3-D memristor crossbars for analog and neuromorphic computing applications, IEEE Trans. Electron Devices, 64, 1, pp. 312-318, (2017)
123456789