Automated Design Space Exploration of Approximation-Enabled Accelerators: A Systematic Literature Mapping

This paper provides a systematic mapping of the literature regarding automated Design Space Exploration (DSE) for error-tolerant systems, where Approximate Computing (AxC) can be included to provide gains in energy efficiency or performance. Starting from the Scopus aggregator and IEEE Xplore library, 1302 texts were analyzed, from which 43 peer-reviewed research works were selected to map the state of the art in AxC-enabled DSE for the design of hardware accelerators. The selected papers were used to catalog the most commonly used search strategies, the methods for integrating approximate units and techniques in the design space of a given system, and challenges and opportunities for the usage of automated DSE in AxC-enabled workflows.The results highlight the usage of genetic and evolutionary algorithms, and identify a trend for tree search based exploration, with a focus on arithmetic units and functional approximation. This study also shows the challenge of integrating a DSE stage in existing workflows, particularly due to the difficulty of analyzing error and electrical characteristics with feasible runtimes.