CNN-SIM: A Detailed Arquitectural Simulator of CNN Accelerators

Abstract

In this work we provide a quick overview of our ongoing effort to derive an open-source framework for detailed architectural simulation of the inference procedure of CNN hardware accelerators. Our tool, called CNN-SIM, exposes the values computed during the inference procedure of any CNN model using real inputs, which allows the investigation of architectural techniques for optimized inference. As a use case, we show the percentage of communicated zero values for two possible dataflows.