文章摘要
张士长* **,王郁杰*,肖航* **,许浩博* **,李佳骏* **,王颖* **,张浩天* **,李晓维* **,韩银和* **.支持CNN与LSTM的二值权重神经网络芯片[J].高技术通讯(中文),2021,31(2):122~128
支持CNN与LSTM的二值权重神经网络芯片
Binary-weight neural network chip supporting CNN and LSTM
  
DOI:10. 3772/j. issn. 1002-0470. 2021. 02. 002
中文关键词: 卷积神经网络(CNN); 长短期记忆(LSTM); 神经网络加速器; 二值权重; 片上系统(SoC)
英文关键词: convolutional neural network (CNN), long short term memory (LSTM), neural network accelerator, binary-weight, system on chip (SoC)
基金项目:
作者单位
张士长* **  
王郁杰*  
肖航* **  
许浩博* **  
李佳骏* **  
王颖* **  
张浩天* **  
李晓维* **  
韩银和* **  
摘要点击次数: 1157
全文下载次数: 770
中文摘要:
      深度神经网络在图像分类、语音识别、视频检测等领域都取得了巨大的成功,这些领域主要采用了卷积神经网络(CNN)、长短期记忆(LSTM)中的一种或者两种网络类型。由于CNN和LSTM网络结构的差异使得现有深度神经网络加速器无法同时高效支持这两种网络类型。权重二值化使得加速器对于CNN和LSTM的同时支持更加高效,同时使得计算复杂度和访存量大幅降低,使得神经网络加速器能够获得更高的能效,并且二值权重对中小规模神经网络模型的精度损失的影响非常有限。本文提出了一种高效支持CNN与LSTM的二值权重神经网络加速器设计结构,该结构在运行CNN和LSTM网络模型时,其核心运算单元利用率超过已有加速器,并且该加速器通过了片上系统(SoC)芯片验证,经过芯片实测,该加速器芯片能效在SoC系统级别达到了6.43 TOPS/W。
英文摘要:
      Deep neural networks (DNNs) have achieved great success in image classification, speech recognition, video detection and other fields in which convolution neural network (CNN), long short term memory (LSTM) or both are mainly used. Existing DNN accelerators are inefficient to support both CNN and LSTM due to the differences of CNN and LSTM. We find that the binarization of weight is more applicable to support both CNN and LSTM and improve the energy efficiency of the neural network accelerator. The binarization of weight makes the requirements of computations and memory access decreased dramatically, which makes it more energy efficient to process the neural networks. Meanwhile, binary-weight neural networks lead to negligible accuracy losses on non-large-scale neural network models. In this letter, we propose a binary-weight CNN-LSTM system on chip (SoC) accelerator. It precedes existing accelerators supporting both CNN and LSTM in core computation cell utilization and achieves 6.43 TOPS/W at SoC level.
查看全文   查看/发表评论  下载PDF阅读器
关闭

分享按钮