基于BP神经网络的CPU电压噪声预测

BP neural network based supply noise prediction on chip processor

为了解决CPU和SOC由于集成度越来越高导致的功耗密度问题,低电压已经成为了一种趋势。但与此同时,这种趋势却使由电压噪声可能导致的电压紧急情况问题更为凸显。传统的电压噪声预测方法首先需要构建精细的CPU供电网络物理模型,然后再进行耗时的RTL层面仿真。为了避免复杂的物理建模过程和进行快速准确地预测,这篇文章提出了一种基于神经网络的CPU电压噪声预测方法。实验表明,基于神经网络的预测方法能够完成CPU的电压噪声预测。并且与传统方法相比,该方法能够完成绝对误差为3.9 mV,相对误差为5.5%的准确预测。

With the increasing integration level, to reduce the power density of modern CPU and SOC, low supply noise level has become a tendency. But at the same time, it makes the voltage emergency caused by potential supply noise even more prominent. To get the supply noise, traditional method first builds a detailed physical model for the CPU power delivery network, then predicts the supply noise by complex RTL level simulation. To avoid the detailed physical modeling and decrease the cost of RTL level simulation, a BP Neural Network based supply noise prediction method is proposed. Experiment results show that the BP Neural Network based supply noise prediction method is able to accomplish the supply noise prediction of CPU. Compared with the traditional method, BP Neural Network based supply noise prediction method is able to achieve a supply noise prediction with a 3.9mV absolute prediction error, 5.5% relative prediction error.