Synergistic architecture and programming model support for approximate micropower computing

Energy consumption is a major constraining factor for embedded multi-core systems. Using aggressive voltage scaling can reduce power consumption, but memory operations become unreliable. Several embedded applications exhibit inherent tolerance to computation approximation, for which indicating parts that can tolerate errors has proven a viable way to reduce energy consumption. In this work we propose an extension to OpenMP to specify what regions of code and data are tolerant to approximation. A compiler pass places data into memory regions with different reliability guarantees according to their tolerance to errors. The voltage supply level is dynamically adjusted according to tolerance policies, with the overall goal of minimizing energy in full compliance with precision constraints.