Dynamic frequency scaling (DFS) is one of the most important approaches for on-the-fly power optimization in modern-day processors. Owing to the trend of chip size shrinkage and increasing the complexity of system design, the problem of achieving an efficient DFS depends upon multi-parametric, non-linear optimization. Hence, it becomes extremely important to identify an optimal underclocking frequency on-the-fly, which depends upon numerous parameters that do not share direct relationship amongst each other. This paper proposes a machine learning approach to DFS of a ubiquitous single-core processor. Several performance parameters of the processor were monitored under an application of a number of clocking frequencies. The dataset thus generated was used to train four artificial neural networks (ANNs) viz. generalized regression (GRNN), decision tree classifier, random forest classifier and backpropagation technique. Under changing parametric conditions of the proposed network, the modes were fit to data while running three applications, i.e. 64- and 1024-point fast fourier transform (FFT) and basicmath applications. The performance of all ANNs was found to be promising and good generalization was obtained with all datasets. In the view of optimizing both speed and power of a system, the results indicate towards suitability of trained GRNN for on-chip deployment for implementing DFS.

中文翻译：

---

使用机器学习范例的单核处理器的动态频率缩放

动态频率缩放（DFS）是现代处理器中动态功率优化的最重要方法之一。由于芯片尺寸缩小的趋势和系统设计复杂性的增加，实现有效DFS的问题取决于多参数非线性优化。因此，实时确定最佳的超频频率变得极为重要，该频率取决于众多彼此之间没有直接关系的参数。本文提出了一种针对无所不在的单核处理器的DFS的机器学习方法。在许多时钟频率的应用下，监视处理器的几个性能参数。这样生成的数据集被用来训练四个人工神经网络（ANN）。广义回归（GRNN），决策树分类器，随机森林分类器和反向传播技术。在所提出的网络的变化的参数条件下，这些模式在运行三个应用程序（即64点和1024点快速傅里叶变换（FFT）和基本数学应用程序）时适合数据。发现所有人工神经网络的性能都令人鼓舞，并且所有数据集都获得了良好的概括。考虑到优化系统的速度和功能，结果表明受过训练的GRNN是否适合于在芯片上部署以实施DFS。发现所有人工神经网络的性能都令人鼓舞，并且所有数据集都获得了良好的概括。考虑到优化系统的速度和功能，结果表明受过训练的GRNN是否适合于在芯片上部署以实施DFS。发现所有人工神经网络的性能都令人鼓舞，并且所有数据集都获得了良好的概括。考虑到优化系统的速度和功能，结果表明受过训练的GRNN是否适合于在芯片上部署以实施DFS。