ExHero: Execution History-Aware Error-Rate Estimation in Pipelined Designs

Abstract

The increased variability renders nanometer devices prone to timing errors. Recent work focused on the development of error prediction models for either evaluating the effects of timing errors on applications or guiding the voltage/frequency settings. Such models may have considered the data-dependent excitation of paths, but they have neglected the impact of all the concurrently executed instructions on error occurrence; which may lead to inaccurate error estimation in pipelined designs. To investigate such a limitation, we develop ExHero, a fully automated framework that performs dynamic timing analysis considering the execution history of a number of in-flight instructions. Using ExHero, we first demonstrate that the order and type of instructions within sequences that have length equal to the pipeline depth significantly affect the error rate. When applied to a pipelined floating-point unit, ExHero reveals that existing approaches estimate on average 46.5% and 32.1% lower error rate and absolute error, respectively, than the actual ones.