文章摘要
于启航* **,文渊博*,杜子东* ***.Bi-SCNN:二值随机混合神经网络加速器[J].高技术通讯(中文),2024,34(12):1243~1255
Bi-SCNN:二值随机混合神经网络加速器
Bi-SCNN: a binary-stochastic hybrid neural network accelerator
  
DOI:10. 3772 / j. issn. 1002-0470. 2024. 12. 001
中文关键词: 二值神经网络(BNN); 随机计算(SC); 神经网络加速器
英文关键词: binary neural network (BNN), stochastic computing (SC), deep learning accelerator
基金项目:
作者单位
于启航* ** (*中国科学院计算技术研究所处理器芯片国家重点实验室北京 100190) (**中国科学院大学北京 100049) (***上海处理器技术创新中心上海 201203) 
文渊博*  
杜子东* ***  
摘要点击次数: 21
全文下载次数: 30
中文摘要:
      二值神经网络(BNN)具有硬件友好的特性,但为了保证计算精度,在输入层仍需要使用浮点或定点计算,增加了硬件开销。针对该问题,本文将另一种同样具有硬件友好特性的随机计算方法应用于BNN,实现了BNN输入层的高效计算,并设计了二值随机混合计算架构Bi-SCNN。首先,在BNN输入层使用高精度的随机运算单元,实现了与定点计算近似的精度;其次,通过在处理单元(PE)内和PE间2个层次对随机数生成器进行复用,并优化运算单元,有效降低了硬件开销;最后,根据输入数据的特性对权值配置方式进行优化,进而降低了整体计算延迟。相比于现有性能最优的BNN加速器,Bi-SCNN在保证计算精度的前提下,实现了2.4倍的吞吐量、12.6倍的能效比和2.2倍的面积效率提升,分别达到2.2 TOPS、7.3 TOPS·W-1和1.8 TOPS·mm-2。
英文摘要:
      Binary neural networks (BNNs) possess hardware-friendly characteristics, yet to ensure computational accuracy, floating-point or fixed-point calculations are still required at the input layer, increasing hardware overhead. To address this issue, this paper applies another hardware-friendly stochastic computing method to BNNs, achieving efficient computation at the BNN input layer and designing a binary stochastic computing neural network (Bi-SCNN) architecture. Firstly, high-precision stochastic computing units are used in the BNN input layer, achieving accuracy comparable to fixed-point computation. Secondly, by reusing random number generators within and between processing elements (PEs) and optimizing the computing units, Bi-SCNN effectively reduces hardware overhead. Lastly, the paper optimizes weight configuration methods based on input data characteristics, thereby reducing overall computational latency. Compared with the existing best-performing BNN accelerators, Bi-SCNN achieves a 2.4-fold increase in throughput, a 12.6-fold increase in energy efficiency, and a 2.2-fold improvement in area efficiency, reaching 2.2 TOPS, 7.3 TOPS·W-1, and 1.8 TOPS·mm-2 respectively.
查看全文   查看/发表评论  下载PDF阅读器
关闭

分享按钮