EGAN: A Framework for Exploring the Accuracy vs. Energy Efficiency Trade-off in Hardware Implementation of Error Resilient Applications

In this paper, we propose a framework, which is called EGAN, for exploring the trade-off between accuracy and energy efficiency in hardware implementation of error resilient applications. EGAN automatically extracts the Pareto frontier (PF) of approximate implementations of an error resilient application based on the data flow graph (DFG) of the application as well as the accuracy and energy consumption of the available approximate/exact components. The framework explores different implementation configurations heuristically to find the best energy efficient implementation of the input application under various output accuracies. The proposed framework, which works by generating some random configurations, clustering them and suggesting some neighboring configurations, reduces the search space considerably. As a result, EGAN achieves a significant reduction in the number of explored configurations compared to the exhaustive (exact) approach while achieving near-optimal results. The efficacy of the proposed framework is assessed using three DSP applications consisting of Sobel edge detector, Finite Inverse Response (FIR) filter and Discrete Cosine Transform (DCT). The studies show that in the worst-case (DCT application with 42 components) EGAN takes 89 hours to extract the PF whereas the exact approach takes 5 million years.