Dynamic Allocation of Datapath Resources for Low Power

Dmitry Ponomarev, Gurhan Kucuk, Kanad Ghose
Workshop on Complexity-Effective Design, 28^th Int'l. Symposium on Computer Architecture (ISCA-28), Goteborg, Sweden, June 2001.

Abstract

We show by profiling the execution of SPEC 95 benchmarks that the usage of datapath resources in a modern superscalar processor is highly dyanmic and correlated. The one-size-fits-all philosophy used for permanently allocating datapath resources in a modern superscalar CPU is thus complexity-ineffective due to the overcommitment of resources in general. We propose a strategy to dynamically and simultaneously adjust the sizes of two such correlated resources - the dispatch buffer (also known as an issue queue) and the reorder buffer - to reduce power dissipation in the datapath without significant impact on the performance. We also show how the resizing technqiue can be augmented with dynamic adaptation of dispatch rate. Representative results show reduction in power dissipation of 69% for the dispatch buffer and of 52% for the reorder buffer with an average IPC loss below 8.5%.

Back to Main