| dc.description.abstract | Emerging high-performance architectures are anticipated to contain unreliable components (e.g., ALUs) that offer low power consumption at the expense of soft errors. Some applications (such as multimedia processing, machine learning, and big data analytics) can often naturally tolerate soft errors and can therefore trade accuracy of their results for reduced energy consumption by utilizing these unreliable hardware components. We present and evaluate a technique for reliability-aware optimization of approximate computational kernel implementations. Our technique takes a standard implementation of a computation and automatically replaces some of its arithmetic operations with unreliable versions that consume less power, but may produce incorrect results with some probability. Our technique works with a developer-provided specification of the required reliability of a computation -- the probability that it returns the correct result -- and produces an unreliable implementation that satisfies that specification. We evaluate our approach on five applications from the image processing, numerical analysis, and financial analysis domains and demonstrate how our technique enables automatic exploration of the trade-off between the reliability of a computation and its performance. | en_US |
