Resilience Evaluation for Approximating SystemC Designs Using Machine Learning Techniques

As digital circuits have become more complicated than ever, abstract description languages such as SystemC have been introduced, allowing designers to work on more abstract levels during the design process. Design metrics such as performance and energy consumption are a central concern for designers at all levels of abstraction. Approximate computing is a promising way to optimize these criteria, sacrificing accuracy. Defining which parts of a design can be approximated (and to what degree) is a crucial and non-trivial design decision, which is usually connected to a larger programming effort, especially when exploring the design space manually. In this paper, we propose an automated approach based on machine learning techniques in order to detect the resilience of a given SystemC design's modules. This is used to identify components of the design that can be approximated. The effectiveness of the proposed method is evaluated using several SystemC benchmarks from various domains.