考虑边缘计算的轻量级网络硬件优化设计

Design of lightweight network hardware optimization considering edge computing

随着移动互联网、物联网的蓬勃发展,大量智能终端设备产生了海量数据,这需要在网络边缘进行实时的智能分析和处理。因此,研究轻量级神经网络的硬件优化方案,以实现边缘智能成为当下的研究热点。文章阐述了基于模型压缩与量化、定点计算替代浮点计算、数据流优化、存储优化与并行计算等方面的轻量级网络硬件设计与优化策略,在FPGA实现方面,采用流水线并行与BRAM利用提升了MobileNetV2的执行效率。结果表明,与原始模型相比,优化后的模型参数量、内存占用等资源利用指标显著降低,CPU利用率、推理速度等性能指标明显提升。实验研究验证了文章所提的各项优化方法,为将深度神经网络部署到边缘设备提供了参考。

With the booming development of mobile internet and the Internet of Things,a large number of intelligent terminal devices have generated massive amounts of data,which requires real-time intelligent analysis and processing at the edge of the network.Therefore,researching hardware optimization solutions for lightweight neural networks to achieve edge intelligence has become a current research hotspot.This article focuses on the design and optimization strategies of lightweight network hardware based on model compression and quantization,fixed-point computing replacing floating-point computing,data flow optimization,storage optimization,and parallel computing.In terms of FPGA implementation,the use of pipeline parallelism and BRAM improves the execution efficiency of MobileNetV2.The results show that compared to the original model,the optimized model significantly reduces resource utilization indicators such as parameter count and memory usage,while performance indicators such as CPU utilization and inference speed are significantly improved.The study validated the proposed optimization methods and provided a reference for deploying deep neural networks to edge devices.