Energy Complexity Model for Convolutional Neural Networks

0573373 - ÚI 2024 RIV CH eng C - Konferenční příspěvek (zahraniční konf.)
Šíma, Jiří - Vidnerová, Petra - Mrázek, V.
Energy Complexity Model for Convolutional Neural Networks.
Artificial Neural Networks and Machine Learning – ICANN 2023. Proceedings, Part X. Cham: Springer, 2023 - (Iliadis, L.; Papaleonidas, A.; Angelov, P.; Jayne, C.), s. 186-198. Lecture Notes in Computer Science, 14263. ISBN 978-3-031-44203-2. ISSN 0302-9743.
[ICANN 2023: International Conference on Artificial Neural Networks /32./. Heraklion (GR), 26.09.2023-29.09.2023]
Grant CEP: GA ČR(CZ) GA22-02067S
Institucionální podpora: RVO:67985807
Klíčová slova: Convolutional neural networks * Energy complexity * Dataflow
Obor OECD: Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)
https://dx.doi.org/10.1007/978-3-031-44204-9_16

The energy efficiency of hardware implementations of convolutional neural networks (CNNs) is critical to their widespread deployment in low-power mobile devices. Recently, a plethora of methods have been proposed providing energy-optimal mappings of CNNs onto diverse hardware accelerators. Their estimated power consumption is related to specific implementation details and hardware parameters, which does not allow for machine-independent exploration of CNN energy measures. In this paper, we introduce a simplified theoretical energy complexity model for CNNs, based on only two-level memory hierarchy that captures asymptotically all important sources of power consumption of different CNN hardware implementations. We calculate energy complexity in this model for two common dataflows which, according to statistical tests, fits asymptotically very well the power consumption estimated by the Time/Accelergy program for convolutional layers on the Simba and Eyeriss hardware platforms. The model opens the possibility of proving principal limits on the energy efficiency of CNN hardware accelerators.
Trvalý link: https://hdl.handle.net/11104/0343835