On Accuracy-Performance Tradeoff Frameworks for Energy Saving: Models and Review

Abstract

A class of research, which explores energy saving via the accuracy-performance tradeoff ability of applications with soft outputs, is emerging. Many such applications implement various notions of precision improvement in their iterative computations and output approximated results. Computing more precisely often requires greater computational effort; hence there is a tradeoff space between precision and cost as long as the output accuracy remains within a targeted bound. This paper reviews the state-of-the-art accuracy tradeoff frameworks, and presents abstract model representations to facilitate our comparison of these frameworks in terms of robustness, performance, profiling, run-time feedback control, and specification of accuracy metrics.