基于神经网络的多核功耗预测策略

Neural Network Based Power Prediction Strategy for Multi-core Architecture

多/众核处理器是计算机发展的趋势。在多/众核处理器的设计过程中,如何从庞大的设计空间中找出满足条件的设计结构,成为了关键和难点。为了解决传统软件模拟技术开销大、效率低等问题,提出了基于神经网络的模型来预测多核处理器的性能和功耗,建立了BP与RBF两种神经网络预测模型,利用SESC模拟器进行CPI与POWER模拟,并比较分析了两种预测模型的预测精度和可靠性。模拟结果表明,采用神经网络预测模型平均误差控制在1.6%~6.6%,较传统的软件模拟等方法,能更有效地节省时间、提高效率,其中,RBF神经网络预测模型具有更好的预测精度。

Multi/many-core processor is a trend of computer development.In the design process of multi-core processor,how to find the design structure from a large design space to meet the conditions has become a critical and challenging task.In order to solve the high overhead and low efficiency problem of traditional software simulation technology,this paper proposed a model based on neural network to predict the performance and power of multi-core processors,established two kinds of BP and RBF neural network prediction models,used SESC simulator to simulate CPI and power,and compared and analyzed the prediction accuracy and reality of two kinds of prediction models.Simulation results show that the neural network prediction model average error is 1.6%~6.6%in control.Compared to general software simulation method,it can also save time and improve efficiency.More importantly,RBF neural network prediction model has better prediction accuracy.