基于软件定义的可重构卷积神经网络架构设计

Architecture design of re-configurable convolutional neural network on software definition

为满足卷积神经网络业务处理的灵活性和高性能需求,提出一种基于软件定义的可重构卷积神经网络架构。该架构采用归一化处理流程实现卷积层网络的动态重构与运算模式的加速。采用AHB和AXI的双总线架构,实现卷积神经网络的流水计算。通过软件定义在FPGA上实现了不同网络结构下的数据集实时处理。实验结果表明,所设计的FPGA电路能够实现两种网络模型的软件定义,网络模型与输入数据集相同的条件下,该架构的运算处理能力为CPU的10倍,运算能耗比为GPU的2倍。

In order to meet the flexibility and efficiency requirement in convolutional neural network(CNN),an architecture of re-configurable CNN based on software definition was proposed.In the architecture,the process of CNN could be normalized and the operation mode could be accelerated.The calculation pipeline was implemented by using dual bus architecture based on AHB and AXI protocols.By software definition,the proposed architecture,which could realize the real-time processing of data among different CNN structure,was implemented on FPGA.The result shows that at least 2 CNN models can be software defined on the FPGA circuit.The output measures an operation processing capacity of 10 times that of CPU,and an operation energy consumption ratio of 2 times that of GPU.