摘要
长短期记忆(LSTM,Long Short-Term Memory)网络广泛应用于各个领域,天机系列芯片是的神经形态芯片。天机芯Ⅱ是最新版本,两步映射方法是研究LSTM网络到天机芯Ⅱ的映射工具链。LSTM网络计算比一般神经网络复杂,一般四步运算,映射到神经形态芯片也是四步,四步运算耗时很高,对于很多要求高速度的运算要求满足不了,如军工、金融等领域。为了提高速度满足高速度的人工智能中LSTM网络的需求,提出了一种两步LSTM网络的映射解决方案,进行了相关的计算推导,并在实验仿真平台进行了仿真验证。虽然Core资源开销是四步映射的几十、几百倍,但两步映射方法使得速度提高了一倍,在对速度要求高对硬件开销不高的领域,益处极大。
LSTM network is widely used in various fields. Tianji series chip is the neural form chip developed and designed by the Tsinghua University brain research center. TianjiⅡ is the latest version. This paper studies the mapping tool chain from LSTM network to sky 2. LSTM network computing is more complex than the general neural network. Generally, the four step operation, mapping to the neural morphology chip is also four steps, the four step operation is very time-consuming, for many demands for high speed operation, such as military, financial and other fields. In order to improve the speed to meet the needs of LSTM network in high speed artificial intelligence, I put forward a mapping solution of two step LSTM network, and carried out the related calculation and deduction, and the simulation verification was carried out in the experimental simulation platform. Although the Core resource overhead is a few hundred times the four step map, it has doubled the speed and is very useful in areas with high speed requirements and low hardware overhead.
作者
何峰
王冠睿
裴京
HE Feng;WANG Guanrui;PEI Jing(Department of precision instruments,Tsinghua University,Beijing 100084,China.)
出处
《集成电路应用》
2018年第7期14-17,共4页
Application of IC
基金
国家自然科学基金项目(NSFC 61327902)
清华大学自主科研项目(2016SZ0102)
关键词
集成电路设计
神经形态芯片
长短期记忆
两步映射
integrated circuit design
neuromorphic chip
long and short term memory
twostep mapping
作者简介
何峰，清华大学精密仪器专业，研究方向：精密仪器与电子技术。;通信作者：裴京，清华大学精密仪器专业，研究方向：电子技术，通信邮箱：peij@mail．tsinghua．edu．cn。
