Design space navigation for neighboring power-performance efficient microprocessor configurations

Abstract

Microprocessor design is a considerably complex task. First, microprocessors include many resources that may be configured in different ways. This leads to a time consuming multi-objective optimization problem. Second, currently the designs must take into account not only performance but also power consumption thus making the optimization goal more complex. Third, different types of applications have different demands but producing several different microprocessors would not be cost effective. This paper proposes an efficient algorithm to explore the design space: design space navigation. With this algorithm it is possible to obtain optimal configurations by starting from a baseline and "navigating" on the design space. Different configurations tailored for different applications, but derived from the same baseline, are called neighboring configurations. Experimental results show that navigation finds designs that achieve better power-performance efficiency for a fraction of the time required by other design space exploration algorithms. Also, the algorithm is used to obtain four neighboring configurations for four types of applications: multimedia, integer and floating-point scientific, and database workloads. The results showed that the navigation configuration achieves a power-performance improvement of 30% to 118% depending on the workload. Using different workloads for navigation and execution may result in a loss of efficiency of as much as 94%. © Springer-Verlag Berlin Heidelberg 2005.