基于人工神经网络的NoC智能动态链路管理方法

NoC Intelligent Dynamic Link Management Strategy Based on Artificial Neural Network

摘要:

功耗是片上网络(NoC)主要限制因素,链路状态的选择性开/关切换算法可降低电路级和系统级的链路功耗,这些算法大多集中于一个简单的静态阈值触发机制,该机制决定了是否应该打开或关闭链路。为解决上述触发机制存在诸多限制,提出了一种针对NoC的人工神经网络(Artificial Neutral Network,ANN)作为动态链路功耗管理方法,该方法基于对系统状态的有监督在线学习,通过使用小型可扩展的神经网络来关闭和打开链路,从而提高预测能力。基于人工神经网络的模型利用了非常低的硬件资源,并且可以集成在大型网状和环面NoC中。通过对不同网络拓扑上各种综合流量模型的仿真结果表明,与静态阈值计算相比,该方法在较低的硬件支出下可以节省功耗。可为解决链路管理NoC中的功耗问题提供思路。

Abstract:

Power consumption is the main limiting factor of network on chip (NOC). The selective on/off switching algorithm of link state can reduce the link power consumption at circuit level and system level. Most of these algorithms focus on a simple static threshold trigger mechanism, which determines whether the link should be turned on or off. In order to solve many limitations of the above trigger mechanism, an artificial neural network (ANN) for NOC is proposed as a dynamic link power consumption management method. This method is based on supervised online learning of system state, and uses a small scalable neural network to close and open the link, so as to improve the prediction ability. The model based on artificial neural network makes use of very low hardware resources and can be integrated in large mesh and torus NOC. The simulation results of various comprehensive traffic models on different network topologies show that compared with static threshold calculation, this method can save power consumption with low hardware overhead. It can provide ideas for solving the power consumption problem in link management NOC.