Dynamic Frequency Scaling for Low-Power Operation of a Single-Core Processor: A Radial Basis Function Approach

Abstract

Dynamic frequency scaling (DFS) is one of the most important approaches to saving power in modern day processors. With ever-increasing complexity at system, circuit and device levels, the problem of achieving an efficient DFS boils down to multi-parametric nonlinear optimization. Therefore, it is imperative to explore ingenuous approaches to DFS which could identify an optimal underclocking frequency on-the-fly using an adaptive mechanism. This paper proposes an offline neural network approach to DFS of a ubiquitous single-core processor where several performance parameters of the processor were monitored under application of a number of clocking frequencies. The dataset thus generated was used to train two classifiers, viz. the radial basis function network and the probabilistic neural network. Under changing parametric conditions of the proposed network, the model was fit to performance-monitoring data while running 64-point and 1024-point FFT applications, and one benchmark application named basicmath. To demonstrate the generalization of the classifiers, the models were trained offline by the dataset obtained by clubbing the aforementioned applications. The performance of both the classifiers was found to be promising, and good generalization was obtained with all the datasets. The results indicate toward suitability of trained radial basis family of networks for on-chip deployment for implementing on-the-fly DFS.