On Pareto-frontier Approximate Computing for Many-core Systems

Abstract

Approximate computing is an emerging paradigm that aggressively improves performance or reduces energy consumption by sacrificing computation quality for error forgiving applications. Various approximate techniques, including loop truncation, approximate communication, etc. have been proposed. Previous works focus on optimization using only one approximation knob. However, we have observed that simultaneously optimizing with multiple approximation knobs leads to a large search space and is more likely to find better solutions. Therefore, in this paper, we first develop application models for performance, error, and power, followed by formulation of an optimization problem to maximize system performance under error and power constraints, using three approximation knobs, which are loop truncation, data dropping, and computational precision scaling. In order to solve the problem efficiently, a lightweight algorithm based on interior point algorithm is proposed. Experimental results show that, compared to state-of-the-art approximate approaches, the proposed scheme can reduce the execution time by as much as 33.1%. The overhead of the proposed method is low, making it a suitable approximate scheme for future many-core systems.