A systematic approach for architecture-level energy estimation of accelerator designs

Abstract

With Moore's law slowing down and Dennard scaling over, energy-ecient domain-specific accelerators have become a promising way for hardware designers to continue bringing energy eciency improvements to data and computation-intensive applications. To enable the fast exploration of the accelerator design space, architecture-level energy estimators, which perform energy estimations without requiring complete hardware description of the designs, are critical to designers. However, it is difficult to use existing architecture-level energy estimators to obtain accurate estimates for accelerator designs, as accelerator designs are diverse and sensitive to data patterns. This thesis presents Accelergy, a generally applicable energy estimation methodology for accelerators that allows flexible specification of designs comprised of user-defined high-level compound components and user-defined low-level primitive components, which can be characterized by third-party energy estimation plugins. We have provided primitive and compound components for modeling deep neural network (DNN) accelerator designs as applications of the proposed methodology. The proposed Accelergy energy estimation framework, which consists of the Accelergy energy estimator and multiple estimation plugins, is validated on Eyeriss, a well-known DNN accelerator design. Overall, with its rich collections of action types and components, Accelergy can achieve 95% accuracy comparing to energy obtained from post-layout simulation in terms of total energy consumption and provide accurate energy breakdowns for components at dierent levels of granularity.