SmartApprox: learning-based configuration of approximate memories for energy-efficient execution

Abstract

Approximate memories reduce power and increase energy efficiency, at the expense of errors in stored data. These errors may be tolerated, up to a point, by many applications with negligible impact on the quality of results. Uncontrolled errors in memory may, however, lead to crashes or broken outputs. Error rates are determined by fabrication and operation parameters, and error tolerance depends on algorithms, implementation, and inputs. An ideal configuration features parameters for approximate memory that minimize energy while allowing applications to produce acceptable results. This work introduces SmartApprox, a framework that configures approximation levels based on features of applications. In SmartApprox, a training phase executes a set of applications under different approximation settings, building a knowledge base that correlates application features (e.g., types of instructions and cache efficiency) with suitable approximate memory configurations. At runtime, features of new applications are sampled and approximation knobs are adjusted to correspond to the predicted error tolerance, according to existing knowledge and the current error scenario, in consonance with hardware characterization. In this work, we list and discuss sets of features that influence the approximation results and measure their impact on the error tolerance or applications. We evaluate SmartApprox on different voltage-scaled DRAM scenarios using a knowledge base of 26 applications, wherein energy savings of 36% are possible with acceptable output. An evaluation using a combined energy and quality metric shows that SmartApprox scores 97% of an exhaustive search for ideal configurations, with significantly lower effort and without application-specific quality evaluation.