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.
CITATION STYLE
Muñoz-Martínez, F., Abellán, J. L., & Acacio, M. E. (2020). CNN-SIM: A Detailed Arquitectural Simulator of CNN Accelerators. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 11997 LNCS, pp. 720–724). Springer. https://doi.org/10.1007/978-3-030-48340-1_56
Mendeley helps you to discover research relevant for your work.