搜索

x

留言板

尊敬的读者、作者、审稿人, 关于本刊的投稿、审稿、编辑和出版的任何问题, 您可以本页添加留言。我们将尽快给您答复。谢谢您的支持!

姓名
邮箱
手机号码
标题
留言内容
验证码

基于忆阻器的脉冲神经网络硬件加速器架构设计

武长春 周莆钧 王俊杰 李国 胡绍刚 于奇 刘洋

引用本文:
Citation:

基于忆阻器的脉冲神经网络硬件加速器架构设计

武长春, 周莆钧, 王俊杰, 李国, 胡绍刚, 于奇, 刘洋

Memristor based spiking neural network accelerator architecture

Wu Chang-Chun, Zhou Pu-Jun, Wang Jun-Jie, Li Guo, Hu Shao-Gang, Yu Qi, Liu Yang
PDF
HTML
导出引用
  • 脉冲神经网络(spiking neural network, SNN)作为第三代神经网络, 其计算效率更高、资源开销更少, 且仿生能力更强, 展示出了对于语音、图像处理的优秀潜能. 传统的脉冲神经网络硬件加速器通常使用加法器模拟神经元对突触权重的累加. 这种设计对于硬件资源消耗较大、神经元/突触集成度不高、加速效果一般. 因此, 本工作开展了对拥有更高集成度、更高计算效率的脉冲神经网络推理加速器的研究. 阻变式存储器(resistive random access memory, RRAM)又称忆阻器(memristor), 作为一种新兴的存储技术, 其阻值随电压变化而变化, 可用于构建crossbar架构模拟矩阵运算, 已经在被广泛应用于存算一体(processing in memory, PIM)、神经网络计算等领域. 因此, 本次工作基于忆阻器阵列, 设计了权值存储矩阵, 并结合外围电路模拟了LIF (leaky integrate and fire)神经元计算过程. 之后, 基于LIF神经元模型实现了脉冲神经网络硬件推理加速器设计. 该加速器消耗了0.75k忆阻器, 集成了24k神经元和192M突触. 仿真结果显示, 在50 MHz的工作频率下, 该加速器通过部署三层的全连接脉冲神经网络对MNIST (mixed national institute of standards and technology)数据集进行推理加速, 其最高计算速度可达148.2 frames/s, 推理准确率为96.4%.
    Spiking neural network (SNN) as the third-generation artificial neural network, has higher computational efficiency, lower resource overhead and higher biological rationality. It shows greater potential applications in audio and image processing. With the traditional method, the adder is used to add the membrane potential, which has low efficiency, high resource overhead and low level of integration. In this work, we propose a spiking neural network inference accelerator with higher integration and computational efficiency. Resistive random access memory (RRAM or memristor) is an emerging storage technology, in which resistance varies with voltage. It can be used to build a crossbar architecture to simulate matrix computing, and it has been widely used in processing in memory (PIM), neural network computing, and other fields. In this work, we design a weight storage matrix and peripheral circuit to simulate the leaky integrate and fire (LIF) neuron based on the memristor array. And we propose an SNN hardware inference accelerator, which integrates 24k neurons and 192M synapses with 0.75k memristor. We deploy a three-layer fully connected network on the accelerator and use it to execute the inference task of the MNIST dataset. The result shows that the accelerator can achieve 148.2 frames/s and 96.4% accuracy at a frequency of 50 MHz.
      通信作者: 刘洋, yliu1975@uestc.edu.cn
    • 基金项目: 国家自然科学基金(批准号: 92064004)资助的课题.
      Corresponding author: Liu Yang, yliu1975@uestc.edu.cn
    • Funds: Project supported by the National Natural Science Foundation of China (Grant No. 92064004).
    [1]

    Redmon J, Farhadi A 2017 30th IEEE Conference on Computer Vision & Pattern Recognition Honolulu, HI, July 21–26, 2017 pp6517–6525

    [2]

    Silver D, Schrittwieser J, Simonyan K, Antonoglou I, Huang A, Guez A, Hubert T, Baker L, Lai M, Bolton A, Chen Y, Lillicrap T, Hui F, Sifre L, Van Den Driessche G, Graepel T, Hassabis D 2017 Nature 550 354Google Scholar

    [3]

    McCulloch W S, Pitts W 1943 Bull. Math. Biophys. 5 115Google Scholar

    [4]

    Hodgkin A L, Huxley A F 1952 J. Physiol. 116 449Google Scholar

    [5]

    Gerstner W 1995 Phys. Rev. E:Stat. Phys. Plasmas Fluids Relat. Interdisciplin. Top. 51 738

    [6]

    Maass W 1997 Neural Networks 10 1659Google Scholar

    [7]

    Roy K, Jaiswal A, Panda P 2019 Nature 575 607Google Scholar

    [8]

    陈怡然, 李海, 陈逸中, 陈凡, 李思成, 刘晨晨, 闻武杰, 吴春鹏, 燕博南 2018 人工智能 13 46

    Chen Y R, Li H, Chen Y Z, Chen F, Li S C, Liu C C, Wen W J, Wu C P, Yan B N 2018 Artif. Intell. View 13 46

    [9]

    Schuman C D, Potok T E, Patton R M, Birdwell J D, Dean M E, Rose G S, Plank J S2017 arXiv:1705.06963

    [10]

    Mahapatra N R, Venkatrao B 1999 Crossroads 5 2

    [11]

    von Neumann J 1993 IEEE Ann. Hist. Comput. 15 27Google Scholar

    [12]

    Chen T, Du Z, Sun N, Wang J, Wu C, Chen Y, Temam O 2014 Acm Sigplan Notices 49 269

    [13]

    Benjamin B V, Gao P, Mcquinn E, Chou D Hary S, Chandrasekaran A R, Bussat J, Alvarez-Icaza R, Arthur J V, Merolla P A, Boahen K 2014 Proc. IEEE 102 699Google Scholar

    [14]

    Pei J, Deng L, Song S, Zhao M G, Zhang Y H, Wu S, Wang G R, Zou Z, Wu Z Z, He W, Chen F, Deng N, Wu S, Wang Y, Wu Y J, Yang Z Y, Ma C, Li G Q, Han W T, Li H L, Wu H Q, Zhao R, Xie Y, Shi L P 2019 Nature 572 106Google Scholar

    [15]

    Davies M, Srinivasa N, Lin T H, Chinya G, Cao Y, Choday S H, Dimou G, Joshi P, Imam N, Jain S 2018 IEEE Micro 38 82Google Scholar

    [16]

    Akopyan F, Sawada J, Cassidy A, Alvarez-Icaza R, Arthur J, Merolla P, Imam N, Nakamura Y, Datta P, Nam G J 2015 IEEE Trans. Comput. Aided Des. Integr. Circuits Syst. 34 1537Google Scholar

    [17]

    Furber S B, Galluppi F, Temple S, Plana L A 2014 Proc. IEEE 102 652Google Scholar

    [18]

    李锟, 曹荣荣, 孙毅, 刘森, 李清江, 徐晖 2019 微纳电子与智能制造 1 87

    Li K, Cao R R, Sun Y, Liu S, Li Q J, Xu H 2019 Micro/nano Electron. Intell. Manuf. 1 87

    [19]

    Xia Q F, Yang J J 2019 Nat. Mater. 18 309Google Scholar

    [20]

    邓亚彬, 王志伟, 赵晨晖, 李琳, 贺珊, 李秋红, 帅建伟, 郭东辉 2021 计算机应用研究 38 2241

    Deng Y B, Wang Z W, Zhao C H, Li L, He S, Li Q H, Shuai J W, Guo D H 2021 Appl. Res. Comput. 38 2241

    [21]

    Burr G W, Shelby R M, Sidler S, Nolfo C D, Jang J, Boybat I, Shenoy R S, Narayanan P, Virwani K, Giacometti E U 2015 IEEE Trans. Electron Devices 62 3498Google Scholar

    [22]

    Moro F, Hardy M, Fain B, Dalgaty T, Clemencon P, De Pra A, Esmanhotto E, Castellani N, Blard F, Gardien F, Mesquida T, Rummens F, Eseni D, Casas J, Indiveri G, Payvand M, Vianello E 2022 Nat. Commun. 13 3506Google Scholar

    [23]

    方旭东, 吴俊杰 2020 计算机工程与科学 42 1929Google Scholar

    Fang X D, Wu J J 2020 Comput. Eng. Sci. 42 1929Google Scholar

    [24]

    Peng Y, Wu H, Gao B, Eryilmaz S B, Qian H 2017 Nat. Commun. 8 15199Google Scholar

    [25]

    Huang L, Diao J T, Nie H S, Wang W, Li Z W, Li Q J, Liu H J 2021 Front. Neurosci. 15 639526Google Scholar

  • 图 1  基于LIF模型的全连接脉冲神经网络结构图

    Fig. 1.  Structure diagram of fully connected spiking neural network based on LIF model.

    图 2  (a)神经元原理图; (b)神经元计算原理图

    Fig. 2.  (a) Neuron schematic diagram; (b) schematic diagram of neuron computation.

    图 3  基于突触复用技术及权值共享技术的LIF神经元模型架构

    Fig. 3.  The LIF neuron model architecture based on synapse multiplexing technology and weight sharing technology.

    图 4  (a) 忆阻器的建立/擦除示意图; (b) 基于忆阻器的crossbar阵列

    Fig. 4.  (a) The set/reset operation of the memristor; (b) the crossbar structure based on the memristor.

    图 5  权值共享技术对精度的影响

    Fig. 5.  The influence of weight sharing technology on accuracy.

    图 6  基于忆阻器阵列的计算核架构

    Fig. 6.  Computing core architecture based on resistive random access memory matrix.

    图 7  基于硬件加速器的应用架构图

    Fig. 7.  Application architecture diagram based on hardware accelerator.

    图 8  推理网络模型到硬件加速的映射示意图

    Fig. 8.  Mapping diagram of network to hardware accelerator.

    图 9  硬件加速器对不同规模神经网络计算的加速效率

    Fig. 9.  Acceleration efficiency of hardware accelerator for different scale neuron networks.

    表 1  本次工作与其他类似工作的对比

    Table 1.  Comparison of this work with other works.

    Burr
    et al.[21]
    Peng
    et al.[24]
    Huang
    et al.[25]
    This work
    神经元模型MPMPMPLIF
    突触数161k1k~4.8M192M
    神经元数385128~6.4k24k
    忆阻器开销330k1k1.11M0.75k
    测试数据集MNISTMNISTMNIST
    准确率/%94.0—97.091.793.496.4
    下载: 导出CSV
  • [1]

    Redmon J, Farhadi A 2017 30th IEEE Conference on Computer Vision & Pattern Recognition Honolulu, HI, July 21–26, 2017 pp6517–6525

    [2]

    Silver D, Schrittwieser J, Simonyan K, Antonoglou I, Huang A, Guez A, Hubert T, Baker L, Lai M, Bolton A, Chen Y, Lillicrap T, Hui F, Sifre L, Van Den Driessche G, Graepel T, Hassabis D 2017 Nature 550 354Google Scholar

    [3]

    McCulloch W S, Pitts W 1943 Bull. Math. Biophys. 5 115Google Scholar

    [4]

    Hodgkin A L, Huxley A F 1952 J. Physiol. 116 449Google Scholar

    [5]

    Gerstner W 1995 Phys. Rev. E:Stat. Phys. Plasmas Fluids Relat. Interdisciplin. Top. 51 738

    [6]

    Maass W 1997 Neural Networks 10 1659Google Scholar

    [7]

    Roy K, Jaiswal A, Panda P 2019 Nature 575 607Google Scholar

    [8]

    陈怡然, 李海, 陈逸中, 陈凡, 李思成, 刘晨晨, 闻武杰, 吴春鹏, 燕博南 2018 人工智能 13 46

    Chen Y R, Li H, Chen Y Z, Chen F, Li S C, Liu C C, Wen W J, Wu C P, Yan B N 2018 Artif. Intell. View 13 46

    [9]

    Schuman C D, Potok T E, Patton R M, Birdwell J D, Dean M E, Rose G S, Plank J S2017 arXiv:1705.06963

    [10]

    Mahapatra N R, Venkatrao B 1999 Crossroads 5 2

    [11]

    von Neumann J 1993 IEEE Ann. Hist. Comput. 15 27Google Scholar

    [12]

    Chen T, Du Z, Sun N, Wang J, Wu C, Chen Y, Temam O 2014 Acm Sigplan Notices 49 269

    [13]

    Benjamin B V, Gao P, Mcquinn E, Chou D Hary S, Chandrasekaran A R, Bussat J, Alvarez-Icaza R, Arthur J V, Merolla P A, Boahen K 2014 Proc. IEEE 102 699Google Scholar

    [14]

    Pei J, Deng L, Song S, Zhao M G, Zhang Y H, Wu S, Wang G R, Zou Z, Wu Z Z, He W, Chen F, Deng N, Wu S, Wang Y, Wu Y J, Yang Z Y, Ma C, Li G Q, Han W T, Li H L, Wu H Q, Zhao R, Xie Y, Shi L P 2019 Nature 572 106Google Scholar

    [15]

    Davies M, Srinivasa N, Lin T H, Chinya G, Cao Y, Choday S H, Dimou G, Joshi P, Imam N, Jain S 2018 IEEE Micro 38 82Google Scholar

    [16]

    Akopyan F, Sawada J, Cassidy A, Alvarez-Icaza R, Arthur J, Merolla P, Imam N, Nakamura Y, Datta P, Nam G J 2015 IEEE Trans. Comput. Aided Des. Integr. Circuits Syst. 34 1537Google Scholar

    [17]

    Furber S B, Galluppi F, Temple S, Plana L A 2014 Proc. IEEE 102 652Google Scholar

    [18]

    李锟, 曹荣荣, 孙毅, 刘森, 李清江, 徐晖 2019 微纳电子与智能制造 1 87

    Li K, Cao R R, Sun Y, Liu S, Li Q J, Xu H 2019 Micro/nano Electron. Intell. Manuf. 1 87

    [19]

    Xia Q F, Yang J J 2019 Nat. Mater. 18 309Google Scholar

    [20]

    邓亚彬, 王志伟, 赵晨晖, 李琳, 贺珊, 李秋红, 帅建伟, 郭东辉 2021 计算机应用研究 38 2241

    Deng Y B, Wang Z W, Zhao C H, Li L, He S, Li Q H, Shuai J W, Guo D H 2021 Appl. Res. Comput. 38 2241

    [21]

    Burr G W, Shelby R M, Sidler S, Nolfo C D, Jang J, Boybat I, Shenoy R S, Narayanan P, Virwani K, Giacometti E U 2015 IEEE Trans. Electron Devices 62 3498Google Scholar

    [22]

    Moro F, Hardy M, Fain B, Dalgaty T, Clemencon P, De Pra A, Esmanhotto E, Castellani N, Blard F, Gardien F, Mesquida T, Rummens F, Eseni D, Casas J, Indiveri G, Payvand M, Vianello E 2022 Nat. Commun. 13 3506Google Scholar

    [23]

    方旭东, 吴俊杰 2020 计算机工程与科学 42 1929Google Scholar

    Fang X D, Wu J J 2020 Comput. Eng. Sci. 42 1929Google Scholar

    [24]

    Peng Y, Wu H, Gao B, Eryilmaz S B, Qian H 2017 Nat. Commun. 8 15199Google Scholar

    [25]

    Huang L, Diao J T, Nie H S, Wang W, Li Z W, Li Q J, Liu H J 2021 Front. Neurosci. 15 639526Google Scholar

  • [1] 王梦蛟, 杨琛, 贺少波, 李志军. 一种新型复合指数型局部有源忆阻器耦合的Hopfield神经网络. 物理学报, 2024, 73(13): 130501. doi: 10.7498/aps.73.20231888
    [2] 王英, 黄慧香, 黄香林, 郭婷婷. 光电协同调控下HfOx基阻变存储器的阻变特性. 物理学报, 2023, 72(19): 197201. doi: 10.7498/aps.72.20230797
    [3] 郭慧朦, 梁燕, 董玉姣, 王光义. 蔡氏结型忆阻器的简化及其神经元电路的硬件实现. 物理学报, 2023, 72(7): 070501. doi: 10.7498/aps.72.20222013
    [4] 方波浪, 王建国, 冯国斌. 基于物理信息神经网络的光斑质心计算. 物理学报, 2022, 71(20): 200601. doi: 10.7498/aps.71.20220670
    [5] 黄颖, 顾长贵, 杨会杰. 神经网络超参数优化的删除垃圾神经元策略. 物理学报, 2022, 71(16): 160501. doi: 10.7498/aps.71.20220436
    [6] 丁大为, 卢小齐, 胡永兵, 杨宗立, 王威, 张红伟. 分数阶忆阻耦合异质神经元的多稳态及硬件实现. 物理学报, 2022, 71(23): 230501. doi: 10.7498/aps.71.20221525
    [7] 古亚娜, 梁燕, 王光义, 夏晨阳. NbOx忆阻神经元的设计及其在尖峰神经网络中的应用. 物理学报, 2022, 71(11): 110501. doi: 10.7498/aps.71.20220141
    [8] 朱佳雪, 张续猛, 王睿, 刘琦. 面向神经形态感知和计算的柔性忆阻器基脉冲神经元. 物理学报, 2022, 71(14): 148503. doi: 10.7498/aps.71.20212323
    [9] 张亚君, 蔡佳林, 乔亚, 曾中明, 袁喆, 夏钶. 基于磁性隧道结的群体编码实现无监督聚类. 物理学报, 2022, 71(14): 148506. doi: 10.7498/aps.71.20220252
    [10] 王童, 温娟, 吕康, 陈健中, 汪亮, 郭新. 仿生生物感官的感存算一体化系统. 物理学报, 2022, 71(14): 148702. doi: 10.7498/aps.71.20220281
    [11] 任宽, 张握瑜, 王菲, 郭泽钰, 尚大山. 基于忆阻器阵列的下一代储池计算. 物理学报, 2022, 71(14): 140701. doi: 10.7498/aps.71.20220082
    [12] 周正, 黄鹏, 康晋锋. 基于非挥发存储器的存内计算技术. 物理学报, 2022, 71(14): 148507. doi: 10.7498/aps.71.20220397
    [13] 刘玉东, 王连明. 基于忆阻器的spiking神经网络在图像边缘提取中的应用. 物理学报, 2014, 63(8): 080503. doi: 10.7498/aps.63.080503
    [14] 李志军, 曾以成, 李志斌. 改进型细胞神经网络实现的忆阻器混沌电路. 物理学报, 2014, 63(1): 010502. doi: 10.7498/aps.63.010502
    [15] 杨金, 周茂秀, 徐太龙, 代月花, 汪家余, 罗京, 许会芳, 蒋先伟, 陈军宁. 阻变存储器复合材料界面及电极性质研究. 物理学报, 2013, 62(24): 248501. doi: 10.7498/aps.62.248501
    [16] 俞阿龙. 基于遗传小波神经网络的机器人腕力传感器动态建模研究. 物理学报, 2008, 57(6): 3385-3390. doi: 10.7498/aps.57.3385
    [17] 王永生, 孙 瑾, 王昌金, 范洪达. 变参数混沌时间序列的神经网络预测研究. 物理学报, 2008, 57(10): 6120-6131. doi: 10.7498/aps.57.6120
    [18] 王瑞敏, 赵 鸿. 神经元传输函数对人工神经网络动力学特性的影响. 物理学报, 2007, 56(2): 730-739. doi: 10.7498/aps.56.730
    [19] 俞阿龙. 基于小波神经网络的振动速度传感器幅频特性补偿研究. 物理学报, 2007, 56(6): 3166-3171. doi: 10.7498/aps.56.3166
    [20] 常胜江, 刘, 张文伟, 申金媛, 翟宏琛, 张延. 适用于神经元状态非等概率分布的神经网络模型及其光学实现. 物理学报, 1998, 47(7): 1101-1109. doi: 10.7498/aps.47.1101
计量
  • 文章访问数:  9303
  • PDF下载量:  396
  • 被引次数: 0
出版历程
  • 收稿日期:  2022-01-13
  • 修回日期:  2022-05-29
  • 上网日期:  2022-07-12
  • 刊出日期:  2022-07-20

/

返回文章
返回