08.03.2022
| Research |
With Machine Learning to Green IT
For battery-powered Internet of Things (IoT) devices, selecting the right hardware is a major challenge. Using an ML-based method, the power consumption of neural networks on graphics processors used in IoT devices can be determined at an early stage.
In the paper “ML-based Power Estimation of Convolutional Neural Networks on GPGPUs”, our research associate Christopher Metz, in cooperation with Mehran Goli (DFKI), has developed an approach to determine the power consumption of GPGPUs, when running Convolutional Neural Networks (CNNs) in early stages of development.
The work improves on the methodology of previous studies published by the authors at the DATE Friday Workshop
System-level Design Methods for Deep Learning on Heterogeneous Architectures (we reported
»here) and the
International Conference on Hardware/Software Codesign and System Synthesis (we reported
»here).
With the new approach it is possible to determine the power consumption of a Convolutional Neural Network on a General Purpose Computing on Graphics Processing Unit (GPGPU) using “only” seven predictors. This makes it possible to determine the power consumption without having to run the neural network on a real graphics card.
The work was done in cooperation with the Cyber-Physical Systems group of DFKInBremen and will be presentedat the
25th IEEE International Symposium on Design and Diagnostics of Electronic Circuits and Systems (DDECS) (
ddecs2022.fit.cvut.cz) held in Prague from 06.04. to 08.04.2022. Other authors are Dr.-Ing. Mehran Goli (DFKI/University of Bremen) and Prof. Rolf Drechsler (University of Bremen/DFKI).
About DDECS:
The
International Symposium on Design and Diagnostics of Electronic Circuits and Systems (DDECS) provides a forum for exchanging ideas, discussing research results, and presenting practical applications in the areas of design, test, and diagnosis of microelectronic digital, analog, and mixed-signal circuits and systems.
Updated by: Lena Steinmann
« back
08.03.2022 | Research
With Machine Learning to Green IT
For battery-powered Internet of Things (IoT) devices, selecting the right hardware is a major challenge. Using an ML-based method, the power consumption of neural networks on graphics processors used in IoT devices can be determined at an early stage.
In the paper “ML-based Power Estimation of Convolutional Neural Networks on GPGPUs”, our research associate Christopher Metz, in cooperation with Mehran Goli (DFKI), has developed an approach to determine the power consumption of GPGPUs, when running Convolutional Neural Networks (CNNs) in early stages of development.
The work improves on the methodology of previous studies published by the authors at the DATE Friday Workshop
System-level Design Methods for Deep Learning on Heterogeneous Architectures (we reported
»here) and the
International Conference on Hardware/Software Codesign and System Synthesis (we reported
»here).
With the new approach it is possible to determine the power consumption of a Convolutional Neural Network on a General Purpose Computing on Graphics Processing Unit (GPGPU) using “only” seven predictors. This makes it possible to determine the power consumption without having to run the neural network on a real graphics card.
The work was done in cooperation with the Cyber-Physical Systems group of DFKInBremen and will be presentedat the
25th IEEE International Symposium on Design and Diagnostics of Electronic Circuits and Systems (DDECS) (
ddecs2022.fit.cvut.cz) held in Prague from 06.04. to 08.04.2022. Other authors are Dr.-Ing. Mehran Goli (DFKI/University of Bremen) and Prof. Rolf Drechsler (University of Bremen/DFKI).
About DDECS:
The
International Symposium on Design and Diagnostics of Electronic Circuits and Systems (DDECS) provides a forum for exchanging ideas, discussing research results, and presenting practical applications in the areas of design, test, and diagnosis of microelectronic digital, analog, and mixed-signal circuits and systems.
Author: Lena Steinmann
Please contact us if you have any questions:
Christopher Metz
Research Associate
+49 (421) 218 - 63942
cmetz@uni-bremen.de
« back