An Energy and Area Estimation Plugin for Accelerator Architecture Simulation

Abstract

Development of domain-specific hardware accelerators has been an important focus for high performance computing research in recent years, enabling significant gains in a variety of practical applications. Of particular interest is accelerator design for applications involving sparse data. Such accelerators inherently tend towards a diverse array of architecture designs, and often rely on custom simulators for evaluation. In addition to raw performance, energy consumption and chip area are both important considerations for evaluating accelerators. Accelergy is a tool that provides a good general framework for fine-grained energy and area estimation. However, output from simulation tools may not be compatible with Accelergy’s expected input format, which is the case for the custom simulator Accelsim. To address this gap, this work presents a streamlined plugin for processing Accelsim simulator output into Accelergy input, for the purpose of generating accurate and explainable energy consumption and area models for accelerator architectures. We demonstrate the plugin’s flexibility by performing energy and area estimates for two state-of-the-art hardware accelerators, ISOSceles and Trapezoid. Overall, this plugin is easy-to-use, self-contained, and supports a wide variety of configurable functionalities, making it an excellent general tool for running Accelergy on Accelsim output.