Design Space Exploration of Approximate Computing Techniques

Nowadays, the rising energy consumption of smartphones and portable devices creates an energy efficiency challenge. To address this problem, Approximate Computing (AxC) techniques are becoming popular since they sacrifice computation accuracy for enhanced performance, energy efficiency, and area reduction. However, selecting suitable AxC techniques for target applications remains intricate. Design Space Exploration (DSE) approaches can be employed to systematically explore all different possible approximate versions of an application and select the most suitable versions. This paper proposes a DSE approach that models the target application computations and the approximation-induced errors using Interval Arithmetic. The experimental results show the efficiency of the proposed approach in quickly evaluating different approximate versions of an application eliminating the time-consuming task of executing each approximate version. Also, using Artificial intelligence, such as Reinforcement Learning approaches, is proposed to explore the design space automatically.