Power Modeling and Exploration of Dynamic and Partially Reconfigurable Systems

Abstract

Although fairly known for a long time, the vast potential of Dynamic and Partial Reconfigu-ration (DPR) for high energy efficiency is still difficult to exploit, for reasons that are more methodological than purely technical. This work addresses this problem and provides a contri-bution by seeking to improve energy efficient deployment and analysis for embedded hetero-geneous multiprocessor platforms representative of current and upcoming systems. This paper explores the potential energy efficiency improvements of DPR on the concrete implementation of a H.264/AVC video decoder. The methodology used to explore the different implementations is presented and formalized. This formalization is based on pragmatic power consumption models of all the tasks of the application that are derived from real measurements. Results allow to identify low energy / high performance mappings, and by extension, condi-tions at which partial reconfiguration can achieve energy efficient application processing. The improvements are expected to be of 57% in energy and 37% in performance over pure software execution, corresponding also to 16% energy savings over static implementation of the same accelerators for 10% less performance.