基于忆阻器的PIM结构实现深度卷积神经网络近似计算

发布时间：2018-01-13 08:03

本文关键词：基于忆阻器的PIM结构实现深度卷积神经网络近似计算　出处：《计算机研究与发展》2017年06期 　论文类型：期刊论文

【摘要】：忆阻器(memristor)能够将存储和计算的特性融合,可用于构建存储计算一体化的PIM(processing-in-memory)结构.但是,由于计算阵列以及结构映射方法的限制,基于忆阻器阵列的深度神经网络计算需要频繁的AD/DA转换以及大量的中间存储,导致了显著的能量和面积开销.提出了一种新型的基于忆阻器的深度卷积神经网络近似计算PIM结构,利用模拟忆阻器大大增加数据密度,并将卷积过程分解到不同形式的忆阻器阵列中分别计算,增加了数据并行性,减少了数据转换次数并消除了中间存储,从而实现了加速和节能.针对该结构中可能存在的精度损失,给出了相应的优化策略.对不同规模和深度的神经网络计算进行仿真实验评估,结果表明,在相同计算精度下,该结构可以最多降低90%以上的能耗,同时计算性能提升约90%.
[Abstract]:The memory device memristors can blend the features of storage and computing. It can be used to construct the PIMN processing-in-memory structure, however, due to the limitation of the computing array and the method of structure mapping. The computation of depth neural network based on resistive array requires frequent AD/DA conversion and a large amount of intermediate storage. A novel deep convolution neural network based on amnesizer is proposed to approximate calculate the PIM structure, which greatly increases the data density. The convolution process is decomposed into different forms of memory array, which increases the data parallelism, reduces the number of data conversion and eliminates the intermediate storage. In view of the possible loss of precision in the structure, the corresponding optimization strategy is given. The simulation results show that the neural network calculation with different sizes and depths is evaluated by simulation. Under the same calculation precision, the energy consumption of the structure can be reduced by more than 90%, and the computational performance can be improved by about 90 percent at the same time.
【作者单位】：西北工业大学计算机学院;嵌入式系统集成教育部工程研究中心(西北工业大学);
【基金】：国家自然科学基金项目(61472322) 中央高校基本科研业务费专项资金项目(3102015BJ(Ⅱ)ZS018)~~
【分类号】：TN60;TP183
【正文快照】： 新型器件忆阻器(memristor)具有能够将计算和存储功能相融合的特殊性质,因此,以忆阻器及其基本单元构造的PIM(processing-in-memory)结构被认为是访存瓶颈的有效解决途径[1].同时,由于欧姆定律和神经网络矩阵乘法运算的高度契合,并且忆阻器的高存储密度也能满足神经网络的巨大

【相似文献】