基于边缘部署低功耗的神经网络加速器
Low Power Neural Network Accelerators Based on Edge Deployment
周诗云 1钱松荣 2卫少东 1郑鑫1
作者信息
- 1. 贵州大学 机械工程学院,贵阳 550025
- 2. 贵州大学 公共大数据国家重点实验室,贵阳 550025
- 折叠
摘要
卷积神经网络作为一种处理网络数据的深度学习模型,广泛的应用于自动驾驶、航空航天等行业.而随着数据量的增长,卷积网络的结构也变得越来越复杂,对于卷积网络这种计算和资源密集型网络如何部署在低功耗、资源少的边缘设备上就成为了一种困难.而FPGA由于其具有高的并行性和低功耗,可以作为一种边缘部署的设备.在这基础上,提出了一种针对于LeNet-5轻量网络的加速器,利用流水线并行加速和循环展开对FPGA的并行计算最大化,然后使用Vitis HLS将高级编程语言转变为硬件描述语言,再利用Vitis IDE进行软件驱动的编写.实验结果表明,相对于在CPU、GPU上进行网络推理,在ZYNQ上FPGA进行网络推理,在检测速率相近的情况下,功耗减少了8倍,这使得神经网络的边缘部署多了一种选择.
Abstract
Convolutional neural networks,as a deep learning model for processing network data,are widely used in in-dustries such as autonomous driving and aerospace.And with the growth of data volume,the structure of convolution-al network becomes more and more complex,for the convolutional network this kind of computation and resource-in-tensive network how to be deployed on the edge devices with low power consumption and few resources becomes a kind of difficulty.And FPGA can be used as an edge deployment device due to its high parallelism and low power consumption.On this basis,a gas pedal for LeNet-5 lightweight networks is proposed to maximize parallel computa-tion on FPGA using pipelined parallel acceleration and loop unfolding,and then use Vitis HLS to transform the high-level programming language into a hardware description language,and then use the Vitis IDE to write the software driver.Experimental results show that network inference on FPGA on ZYNQ reduces power consumption by a factor of 8 with similar detection rates,relative to network inference on CPU and GPU,which makes it an additional option for edge deployment of neural networks.
关键词
卷积神经网络/边缘部署/低功耗/FPGA/流水线/循环展开/HLSKey words
convolutional neural network/edge deployment/low power consumption/FPGA/pipeline/loop unrolling/HLS引用本文复制引用
基金项目
贵州光电子信息与智能化应用国际联合研究中心项目(黔科合平台人才20195802号)
出版年
2024
NETL
NSTL
万方数据