Towards an Energy Consumption Index for Deep Learning Models: A Comparative Analysis of Architectures, GPUs, and Measurement Tools
<p>Images from the Imagenette dataset, a subset of ImageNet, showcasing 10 categories.</p> "> Figure 2
<p>Comparison of different deep learning architectures: (<b>a</b>) AlexNet, (<b>b</b>) VGG16, (<b>c</b>) ResNet18, (<b>d</b>) EfficientNet-B3, (<b>e</b>) Swin-T, (<b>f</b>) ConvNeXt-T.</p> "> Figure 3
<p>Schematic diagram of the OpenZmeter. It includes an ARM board, AC/DC converter, LiPo battery, and sensors to measure electrical parameters from the power source to the load. Image adapted from [<a href="#B10-sensors-25-00846" class="html-bibr">10</a>].</p> "> Figure 4
<p>CodeCarbon [<a href="#B61-sensors-25-00846" class="html-bibr">61</a>] energy measurement process. Energy consumption (<span class="html-italic">E</span>, in kWh) is estimated using RAM (3 W/8 GB), CPU (Intel-RAPL), and GPU (Nvml). Carbon intensity (<span class="html-italic">C</span>, in kgCO<sub>2</sub>eq/kWh) integrates data from cloud providers, country energy mixes, and world averages to compute CO<sub>2</sub>eq.</p> "> Figure 5
<p>Carbontracker [<a href="#B11-sensors-25-00846" class="html-bibr">11</a>] process for estimating energy consumption (<span class="html-italic">E</span>, kWh) using CPU (Intel RAPL), GPU (Nvml, TDP), and RAM data, adjusted by PUE. Carbon intensity (<span class="html-italic">C</span>, kgCO<sub>2</sub>eq/kWh) is obtained from regional APIs. Final emissions (CO<sub>2</sub>eq) are calculated as <math display="inline"><semantics> <mrow> <mi>C</mi> <msub> <mi>O</mi> <mn>2</mn> </msub> <mi>e</mi> <mi>q</mi> <mo>=</mo> <mi>C</mi> <mo>×</mo> <mi>E</mi> </mrow> </semantics></math>.</p> "> Figure 6
<p>Experimental framework for evaluating the performance and energy consumption of AlexNet, VGG16, ResNet18, EfficientNet-B3, Swin-T, and ConvNeXt-T measured with OpenZmeter, CodeCarbon and Carbontracker.</p> "> Figure 7
<p>Energy consumption (kWh) during the training of AlexNet, VGG16, ResNet18, EfficientNet-B3, Swin-T, and ConvNeXt-T on GPUs TITAN Xp and GTX 1080 Ti. The graph compares the results from three energy measurement tools: OpenZmeter (energy meter reference), CodeCarbon, and Carbontracker.</p> "> Figure 8
<p>Average Execution times (s) during the training of AlexNet, VGG16, ResNet18, EfficientNet-B3, Swin-T, and ConvNeXt-T on GPUs TITAN Xp and GTX 1080 Ti. The graph compares the results from three energy measurement tools: OpenZmeter (energy meter reference), CodeCarbon, and Carbontracker.</p> "> Figure 9
<p>CO<sub>2</sub> Emissions during the training of AlexNet, VGG16, ResNet18, EfficientNet-B3, Swin-T, and ConvNeXt-T on GPUs TITAN Xp and GTX 1080 Ti; results obtained using CodeCarbon.</p> "> Figure 10
<p>Active power consumption during the training of DL models on TITAN Xp using OpenZmeter. The solid lines show the average active power consumption over time, while the shaded areas indicate the variability of the data at each time point.</p> "> Figure 11
<p>Active power consumption during the training of DL Models on GTX 1080 Ti using OpenZmeter. The solid lines show the average active power consumption over time, while the shaded areas indicate the variability of the data at each time point.</p> "> Figure 12
<p><span class="html-italic">Kappa–Energy Index</span> for AlexNet, VGG16, ResNet18, EfficientNet-B3, Swin-T, and ConvNeXt-T on TITAN Xp and GTX 1080 Ti GPUs, comparing OpenZmeter (hardware reference meter) and CodeCarbon (software meter) results.</p> "> Figure 13
<p>Energy consumption by component (RAM, CPU, GPU) during the training of AlexNet, VGG16, ResNet18, EfficientNet-B3, Swin-T, and ConvNeXt-T on TITAN Xp and GTX 1080 Ti GPUs. The results are obtained using CodeCarbon.</p> "> Figure 14
<p>Energy consumption by component (RAM, CPU, GPU) during the inference of AlexNet, VGG16, ResNet18, EfficientNet-B3, Swin-T, and ConvNeXt-T on TITAN Xp and GTX 1080 Ti GPUs. The results are obtained using CodeCarbon.</p> ">
Abstract
:1. Introduction
- Proposal of an energy consumption index to standardize the evaluation of energy efficiency in DL models during training and inference stages, offering a metric for objective comparisons across architectures.
- Development of a combined methodology that uses both hardware and software energy meters to accurately measure energy consumption during the training and inference stages of DL models, enabling a precise evaluation of the energy-performance trade-off.
- Analysis of the impact of heterogeneous hardware on the energy consumption and performance of DL models, examining how differences in hardware architectures affect the energy efficiency during both the training and inference stages of these models.
- Validation and calibration of software energy meters to ensure accuracy in the assessment of the energy consumption during training and inference of DL models, aiming to achieve precise energy consumption measurements through software meters.
2. Materials and Methods
2.1. Data Description
2.2. Convolutional Neural Networks
2.3. Transformer-Based Neural Networks
2.4. Implemented Deep Learning Architectures
2.4.1. AlexNet
2.4.2. VGG16
2.4.3. ResNet18
2.4.4. EfficientNet-B3
2.4.5. Swin-T
2.4.6. ConvNeXt-T
2.5. Performance Evaluation Metrics
2.5.1. Model Performance Evaluation
2.5.2. Measuring Energy Consumption
2.5.3. DL Model Performance over Energy Consumption
2.5.4. Measuring CO2 Emissions
2.6. Energy Consumption Meters
- OpenZmeter [10] is a low-cost, open-source, intelligent hardware energy meter and power quality analyzer. It measures reactive, active, and apparent energy, frequency, Root Mean Square (RMS) voltage, RMS current, power factor, phase angle, voltage events, harmonics up to the 50th order, and total harmonic distortion (THD). It records energy consumption in kilowatt-hours (kWh). The device includes a web interface and an API for integration. It can be installed in electrical distribution panels and features Ethernet, Wi-Fi, and 4G connectivity. Additionally, it offers remote monitoring and real-time alerts. Figure 3 shows the OpenZmeter diagram.
- CodeCarbon [61] is an open-source tool to measure and reduce software programs’ carbon footprint. It tracks energy consumption in kilowatt-hours (kWh) during code execution, accounting for hardware and data center location to calculate CO2 emissions. Carbon intensity can vary hourly and adapt to users’ location when real-time data from sources like the CO2 Signal API are accessible. For cloud computing, CodeCarbon uses Google Cloud Platform data (Mountain View, CA, USA), although Amazon (Seattle, WA, USA) and Microsoft (Redmond, WA, USA) do not provide specific carbon details for their data centers. For private infrastructures, CodeCarbon draws from “Our World in Data” when available or defaults to the energy mix of the user’s country from “globalpetrolprices.com” adjusting carbon intensity accordingly. When specific data are absent, a global average of 475 gCO2/kWh, based on the International Energy Agency (IEA), is applied. The tool also offers an Application Programming Interface (API) and Python libraries to integrate carbon monitoring into projects, along with reports and visualizations that consider data center locations. Energy consumption measurement focuses on key system components, specifically the GPU, RAM, and CPU. However, the tool does not account for other elements, such as storage, network, or peripherals, which leads to underestimations of total consumption. Figure 4 shows the CodeCarbon diagram.
- Carbontracker [11] is an open-source software tool for energy management in the training of DL models, allowing users to track and predict energy consumption and carbon emissions. It facilitates the adjustment of epochs to monitor consumption and can track the entire training process to provide accurate estimates. Noteworthy research utilizing Carbontracker includes a study cited in [62], which discusses how combining federated learning with transfer learning can enhance the classification of medical images in an energy-efficient and privacy-preserving manner. Another investigation [63] examines the relationship between the quality of generative audio models and their energy consumption. Figure 5 shows the Carbontracker diagram.
2.7. Computational Resources
2.8. Experimental Setup
Algorithm 1: Pseudo-code for Performance and Energy Evaluation |
3. Results
3.1. Energy Consumption and Meter Precision in DL Models Training Experiments
Active Power Consumption During DL Models Training
3.2. Evaluation of Kappa–Energy Index for DL Models Training and Inference
3.2.1. Evaluation of Kappa–Energy Index for DL Models Training
3.2.2. Evaluation of Kappa–Energy Index for DL Models Inference
3.3. Evaluation of Hyperparameter Influence and Energy Scaling per CUDA Core
4. Discussion
5. Conclusions
Author Contributions
Funding
Institutional Review Board Statement
Informed Consent Statement
Data Availability Statement
Conflicts of Interest
References
- Foster, S.; Elzinga, D. The Role of Fossil Fuels in a Sustainable Energy System. 2023. Available online: https://www.un.org/es/chronicle/article/el-papel-de-los-combustibles-fosiles-en-un-sistema-energetico-sostenible (accessed on 29 November 2023).
- United Nations Framework Convention on Climate Change. Paris agreement. In Proceedings of the Report of the Conference of the Parties to the United Nations Framework Convention on Climate Change (21st Session, 2015: Paris); United Nations: Bonn, Germany, 2015; Volume 4, p. 2017. Available online: https://unfccc.int/sites/default/files/resource/parisagreement_publication.pdf (accessed on 5 August 2024).
- Yan, Y.; Qian, Y.; Sharif, H.; Tipper, D. A Survey on Smart Grid Communication Infrastructures: Motivations, Requirements and Challenges. IEEE Commun. Surv. Tutor. 2013, 15, 5–20. [Google Scholar] [CrossRef]
- Prieto, A.; Prieto, B. Las Tecnologías de la Información y de las Comunicaciones como parte del problema y de la solución del consumo energético. Actas XXIX Jorn. Sobre Ense Nanza Univ. Inform. 2023, 8, 9–23. [Google Scholar]
- Kuo, C.C.J.; Madni, A.M. Green learning: Introduction, examples and outlook. J. Vis. Commun. Image Represent. 2023, 90, 103685. [Google Scholar] [CrossRef]
- Cheng, H.; Liu, B.; Lin, W.; Ma, Z.; Li, K.; Hsu, C.H. A survey of energy-saving technologies in cloud data centers. J. Supercomput. 2021, 77, 13385–13420. [Google Scholar] [CrossRef]
- Mehonic, A.; Kenyon, A.J. Brain-inspired computing needs a master plan. Nature 2022, 604, 255–260. [Google Scholar] [CrossRef]
- Goodfellow, I. Deep Learning; MIT Press: Cambridge, MA, USA, 2016. [Google Scholar]
- Liu, Z.; Lin, Y.; Cao, Y.; Hu, H.; Wei, Y.; Zhang, Z.; Lin, S.; Guo, B. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. arXiv 2021, arXiv:2103.14030. [Google Scholar]
- Viciana, E.; Alcayde, A.; Montoya, F.G.; Baños, R.; Arrabal-Campos, F.M.; Zapata-Sierra, A.; Manzano-Agugliaro, F. OpenZmeter: An efficient low-cost energy smart meter and power quality analyzer. Sustainability 2018, 10, 4038. [Google Scholar] [CrossRef]
- Anthony, L.F.W.; Kanding, B.; Selvan, R. Carbontracker: Tracking and Predicting the Carbon Footprint of Training Deep Learning Models. arXiv 2020, arXiv:2007.03051. [Google Scholar]
- Montoya, F.G.; Baños, R.; Alcayde, A.; Arrabal, F. Efficient open-source power quality analyser and smart meter. In Proceedings of the 25th International Conference on Electricity Distribution (CIRED), Madrid, Spain, 3–6 June 2019. [Google Scholar]
- Prieto, B.; Escobar, J.J.; Gómez-López, J.C.; Díaz, A.F.; Lampert, T. Energy Efficiency of Personal Computers: A Comparative Analysis. Sustainability 2022, 14, 12829. [Google Scholar] [CrossRef]
- Lannelongue, L.; Grealey, J.; Inouye, M. Green algorithms: Quantifying the carbon footprint of computation. Adv. Sci. 2021, 8, 2100707. [Google Scholar] [CrossRef]
- García-Martín, E.; Lavesson, N.; Grahn, H.; Casalicchio, E.; Boeva, V. How to measure energy consumption in machine learning algorithms. In Proceedings of the ECML PKDD 2018 Workshops: Nemesis 2018, UrbReas 2018, SoGood 2018, IWAISe 2018, and Green Data Mining 2018, Dublin, Ireland, 10–14 September 2018; Proceedings 18. Springer: Cham, Switzerland, 2019; pp. 243–255. [Google Scholar]
- LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, 521, 436–444. [Google Scholar] [CrossRef] [PubMed]
- de Vries, A. The growing energy footprint of artificial intelligence. Joule 2023, 7, 2191–2194. [Google Scholar] [CrossRef]
- Meng, F.; Wang, X.; Shao, F.; Wang, D.; Hua, X. Energy-Efficient Gabor Kernels in Neural Networks with Genetic Algorithm Training Method. Electronics 2019, 8, 105. [Google Scholar] [CrossRef]
- Gordillo, A.; Calero, C.; Moraga, M.Á.; García, F.; Fernandes, J.P.; Abreu, R.; Saraiva, J. Programming languages ranking based on energy measurements. Softw. Qual. J. 2024, 32, 1539–1580. [Google Scholar] [CrossRef]
- Lacoste, A.; Luccioni, A.; Schmidt, V.; Dandres, T. Quantifying the Carbon Emissions of Machine Learning. arXiv 2019, arXiv:1910.09700. [Google Scholar]
- Díaz, A.F.; Prieto, B.; Escobar, J.J.; Lampert, T. Vampire: A smart energy meter for synchronous monitoring in a distributed computer system. J. Parallel Distrib. Comput. 2024, 184, 104794. [Google Scholar] [CrossRef]
- Mattson, P.; Cheng, C.; Diamos, G.; Coleman, C.; Micikevicius, P.; Patterson, D.; Tang, H.; Wei, G.Y.; Bailis, P.; Bittorf, V.; et al. Mlperf training benchmark. Proc. Mach. Learn. Syst. 2020, 2, 336–349. [Google Scholar]
- Reddi, V.J.; Cheng, C.; Kanter, D.; Mattson, P.; Schmuelling, G.; Wu, C.J.; Anderson, B.; Breughe, M.; Charlebois, M.; Chou, W.; et al. Mlperf inference benchmark. In Proceedings of the 2020 ACM/IEEE 47th Annual International Symposium on Computer Architecture (ISCA), Valencia, Spain, 30 May–3 June 2020; pp. 446–459. [Google Scholar]
- Mattson, P.; Reddi, V.J.; Cheng, C.; Coleman, C.; Diamos, G.; Kanter, D.; Micikevicius, P.; Patterson, D.; Schmuelling, G.; Tang, H.; et al. MLPerf: An Industry Standard Benchmark Suite for Machine Learning Performance. IEEE Micro 2020, 40, 8–16. [Google Scholar] [CrossRef]
- Fan, X.; Weber, W.D.; Barroso, L.A. Power provisioning for a warehouse-sized computer. ACM SIGARCH Comput. Archit. News 2007, 35, 13–23. [Google Scholar] [CrossRef]
- Jain, D.; Shukla, R.K.; Tomar, M.S.; Sharma, P. A Study of the Impact of Programming Language Selection on CO2 emission—A Green IT Initiative. In Proceedings of the 2nd International Conference on Data, Engineering and Applications (IDEA), Bhopal, India, 28–29 February 2020; pp. 1–5. [Google Scholar]
- Couto, M.; Pereira, R.; Ribeiro, F.; Rua, R.; Saraiva, J. Towards a green ranking for programming languages. In Proceedings of the 21st Brazilian Symposium on Programming Languages, Fortaleza, CE, Brazil, 21–22 September 2017; pp. 1–8. [Google Scholar]
- Pereira, R.; Couto, M.; Ribeiro, F.; Rua, R.; Cunha, J.; Fernandes, J.P.; Saraiva, J. Energy efficiency across programming languages: How do energy, time, and memory relate? In Proceedings of the 10th ACM SIGPLAN International Conference on Software Language Engineering, SLE 2017, Vancouver, BC, Canada, 23–24 October 2017; pp. 256–267. [Google Scholar] [CrossRef]
- Alizadeh, N.; Castor, F. Green AI: A Preliminary Empirical Study on Energy Consumption in DL Models Across Different Runtime Infrastructures. In Proceedings of the IEEE/ACM 3rd International Conference on AI Engineering—Software Engineering for AI, CAIN 2024, Lisbon, Portugal, 14–15 April 2024; Volume 3, pp. 134–139. [Google Scholar] [CrossRef]
- Henderson, P.; Hu, J.; Romoff, J.; Brunskill, E.; Jurafsky, D.; Pineau, J. Towards the systematic reporting of the energy and carbon footprints of machine learning. J. Mach. Learn. Res. 2020, 21, 1–43. [Google Scholar]
- Bharany, S.; Sharma, S.; Khalaf, O.I.; Abdulsahib, G.M.; Al Humaimeedy, A.S.; Aldhyani, T.H.; Maashi, M.; Alkahtani, H. A systematic survey on energy-efficient techniques in sustainable cloud computing. Sustainability 2022, 14, 6256. [Google Scholar] [CrossRef]
- Bouza, L.; Bugeau, A.; Lannelongue, L. How to estimate carbon footprint when training deep learning models? A guide and review. Environ. Res. Commun. 2023, 5, 115014. [Google Scholar] [CrossRef] [PubMed]
- Escobar, J.J.; Rodríguez, F.; Prieto, B.; Kimovski, D.; Ortiz, A.; Damas, M. A distributed and energy-efficient KNN for EEG classification with dynamic money-saving policy in heterogeneous clusters. Computing 2023, 105, 2487–2510. [Google Scholar] [CrossRef]
- García-Martín, E.; Rodrigues, C.F.; Riley, G.; Grahn, H. Estimation of energy consumption in machine learning. J. Parallel Distrib. Comput. 2019, 134, 75–88. [Google Scholar] [CrossRef]
- Bhuiyan, M. Carbon Footprint Measurement and Mitigation Using AI. SSRN Electronic Journal. 2024. Available online: https://ssrn.com/abstract=4746446 (accessed on 9 October 2024).
- Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; et al. Pytorch: An imperative style, high-performance deep learning library. In Proceedings of the Advances in Neural Information Processing Systems 32, Vancouver, BC, Canada, 8–14 December 2019. [Google Scholar]
- Abadi, M.; Agarwal, A.; Barham, P.; Brevdo, E.; Chen, Z.; Citro, C.; Corrado, G.S.; Davis, A.; Dean, J.; Devin, M.; et al. TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems; Tensorflow: Mountain View, CA, USA, 2015. [Google Scholar]
- Young, S.R.; Devineni, P.; Parsa, M.; Johnston, J.T.; Kay, B.; Patton, R.M.; Schuman, C.D.; Rose, D.C.; Potok, T.E. Evolving Energy Efficient Convolutional Neural Networks. In Proceedings of the 2019 IEEE International Conference on Big Data (Big Data), Los Angeles, CA, USA, 9–12 December 2019; pp. 4479–4485. [Google Scholar] [CrossRef]
- Aquino Brítez, D.; Ortiz, A.; Ortega Lopera, J.; Escobar Pérez, J.J.; Formoso, M.; Gan, J.Q.; Escobar Pérez, J.J. Optimization of Deep Architectures for EEG Signal Classification: An AutoML Approach Using Evolutionary Algorithms. Sensors 2021, 21, 2096. [Google Scholar] [CrossRef] [PubMed]
- Cai, E.; Juan, D.C.; Stamoulis, D.; Marculescu, D. NeuralPower: Predict and Deploy Energy-Efficient Convolutional Neural Networks. arXiv 2017, arXiv:1710.05420. [Google Scholar]
- Koppula, S.; Orosa, L.; Yağlıkçı, A.G.; Azizi, R.; Shahroodi, T.; Kanellopoulos, K.; Mutlu, O. EDEN: Enabling Energy-Efficient, High-Performance Deep Neural Network Inference Using Approximate DRAM. arXiv 2019, arXiv:1910.05340. [Google Scholar]
- Bao, Z.; Yang, S.; Huang, Z.; Zhou, M.; Chen, Y. A Lightweight Block with Information Flow Enhancement for Convolutional Neural Networks. IEEE Trans. Circuits Syst. Video Technol. 2023, 33, 3570–3584. [Google Scholar] [CrossRef]
- Wang, K.; Zheng, H.; Li, J.; Louri, A. Morph-GCNX: A Universal Architecture for High-Performance and Energy-Efficient Graph Convolutional Network Acceleration. IEEE Trans. Sustain. Comput. 2024, 9, 115–127. [Google Scholar] [CrossRef]
- Wen, X.; Zhou, M. Evolution and Role of Optimizers in Training Deep Learning Models. IEEE/CAA J. Autom. Sin. 2024, 11, 2039–2042. [Google Scholar] [CrossRef]
- Xu, J.; Zhou, W.; Fu, Z.; Zhou, H.; Li, L. A survey on green deep learning. arXiv 2021, arXiv:2111.05193. [Google Scholar]
- Kaack, L.H.; Donti, P.L.; Strubell, E.; Kamiya, G.; Creutzig, F.; Rolnick, D. Aligning artificial intelligence with climate change mitigation. Nat. Clim. Change 2022, 12, 518–527. [Google Scholar] [CrossRef]
- Patterson, D.; Gonzalez, J.; Le, Q.; Liang, C.; Munguia, L.M.; Rothchild, D.; So, D.; Texier, M.; Dean, J. Carbon emissions and large neural network training. arXiv 2021, arXiv:2104.10350. [Google Scholar]
- Deng, J.; Dong, W.; Socher, R.; Li, L.J.; Li, K.; Fei-Fei, L. ImageNet: A large-scale hierarchical image database. In Proceedings of the 2009 IEEE Conference on Computer Vision and Pattern Recognition, Miami, FL, USA, 20–25 June 2009; pp. 248–255. [Google Scholar] [CrossRef]
- Samad, S.A.; Gitanjali, J. Augmenting DenseNet: Leveraging Multi-Scale Skip Connections for Effective Early-Layer Information Incorporation. IEEE Access 2024, 12, 141344–141360. [Google Scholar] [CrossRef]
- Vaswani, A. Attention is all you need. In Proceedings of the Advances in Neural Information Processing Systems 30, Long Beach, CA, USA, 4–9 December 2017. [Google Scholar]
- Turner, R.E. An introduction to transformers. arXiv 2023, arXiv:2304.10557. [Google Scholar]
- Krizhevsky, A.; Sutskever, I.; Hinton, G.E. ImageNet Classification with Deep Convolutional Neural Networks. In Proceedings of the Advances in Neural Information Processing Systems, Lake Tahoe, NV, USA, 3–6 December 2012; Pereira, F., Burges, C., Bottou, L., Weinberger, K., Eds.; Curran Associates, Inc.: Red Hook, NY, USA, 2012; Volume 25. [Google Scholar]
- Simonyan, K.; Zisserman, A. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv 2014, arXiv:1409.1556. [Google Scholar]
- He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778. [Google Scholar]
- Tan, M.; Le, Q.V. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. arXiv 2020, arXiv:1905.11946. [Google Scholar]
- Liu, Z.; Mao, H.; Wu, C.Y.; Feichtenhofer, C.; Darrell, T.; Xie, S. A ConvNet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), New Orleans, LA, USA, 18–24 June 2022; pp. 11976–11986. [Google Scholar]
- PyTorch. Torchvision Models. 2024. Available online: https://pytorch.org/vision/main/models (accessed on 6 December 2024).
- Cohen, J. A coefficient of agreement for nominal scales. Educ. Psychol. Meas. 1960, 20, 37–46. [Google Scholar] [CrossRef]
- Bird, J. Electrical and Electronic Principles and Technology; Routledge: London, UK, 2017. [Google Scholar]
- Landis, J.R.; Koch, G.G. The measurement of observer agreement for categorical data. Biometrics 1977, 33, 159–174. [Google Scholar] [CrossRef]
- Courty, B.; Schmidt, V.; Luccioni, S.; Goyal-Kamal; Coutarel, M.; Feld, B.; Lecourt, J.; Connell, L.; Saboni, A.; Inimaz; et al. mlco2/codecarbon: v2.4.1. 2024. Available online: https://zenodo.org/records/11171501 (accessed on 17 June 2024).
- Ahmed, M.S.; Giordano, S. Federated Transfer Learning for Energy Efficient Privacy-preserving Medical Image Classification. In Proceedings of the 2022 IEEE International Conference on E-health Networking, Application & Services (HealthCom), Genoa, Italy, 17–19 October 2022; pp. 240–245. [Google Scholar] [CrossRef]
- Douwes, C.; Esling, P.; Briot, J.P. Energy consumption of deep generative audio models. arXiv 2021, arXiv:2107.02621. [Google Scholar]
Classes | Train | Test | Total |
---|---|---|---|
Cassette player | 993 | 357 | 1350 |
Tench | 963 | 387 | 1350 |
Garbage truck | 961 | 389 | 1350 |
Parachute | 960 | 390 | 1350 |
French horn | 956 | 394 | 1350 |
English springer | 955 | 395 | 1350 |
Golf ball | 951 | 399 | 1350 |
Church | 941 | 409 | 1350 |
Gas pump | 931 | 419 | 1350 |
Chain saw | 858 | 386 | 1244 |
Total | 9469 | 3925 | 13,394 |
Architecture | Number of Parameters | Year of Creation |
---|---|---|
AlexNet [52] | 61,100,840 | 2012 |
VGG16 [53] | 138,357,544 | 2014 |
ResNet18 [54] | 11,689,512 | 2015 |
EfficientNet-B3 [55] | 12,233,232 | 2019 |
Swin-T [9] | 28,288,354 | 2021 |
ConvNeXt-T [56] | 28,589,128 | 2022 |
Kappa Statistic | Agreement Levels |
---|---|
No agreement | |
Slight | |
Fair | |
Moderate | |
Substantial | |
Almost perfect |
Feature | OpenZmeter | CodeCarbon | Carbontracker |
---|---|---|---|
Type | Hardware and software | Software | Software |
Energy Measurement | Reactive, active, apparent, RMS voltage/current, etc. | kWh during code execution | kWh during DL training |
Carbon Emission Tracking | No | Yes (based on hardware and location) | Yes |
Connectivity | Ethernet, Wi-Fi, 4G | API, Python libraries | Python libraries |
Integration | Web interface, API | API, Python libraries | API, Python libraries |
Focus | Energy and Power Quality Analysis | Carbon footprint reduction in software execution | Energy management in DL training |
Advanced Features | Real-time alerts, power quality analysis | Reports, geographical data center insights | Dynamic adjustment of training epochs |
Category | Specifications |
---|---|
Hardware | |
Architecture | x86_64 |
Processors | 2× Intel Xeon E5-2640 v4 @ 2.40 GHz, 10 cores each, 90 W TDP each |
RAM | 126 GB |
GPUs | 3× NVIDIA GTX 1080 Ti, 250 W TDP, 3584 CUDA cores, 11 GB each |
1× NVIDIA TITAN Xp, 250 W TDP, 3840 CUDA cores, 12 GB | |
Power Meters | OpenZmeter |
Software | |
Operating System | Ubuntu 24.04 LTS |
Python Version | 3.10.14 |
Pytorch Version | 2.5.1 |
CodeCarbon Version | 2.4.1 |
Carbontracker Version | 1.2.5 |
Category | Specifications |
---|---|
Number of epochs | 90 |
Batch size | 64 |
Optimizer | Stochastic Gradient Descent (SGD) |
Learning rate | 0.001 |
Momentum | 0.9 |
Weight decay | 0.0001 |
GPU | Energy Meter | Architecture | Execution Times (s) | Energy Consumed (kWh) | p-Values | ||
---|---|---|---|---|---|---|---|
Average | Std. Dev. | Average | Std. Dev. | ||||
TITAN Xp | OpenZmeter (reference meter) | AlexNet | 1198 | 4 | 0.1062 | 0.0005 | - |
VGG16 | 1197 | 3 | 0.1319 | 0.0010 | - | ||
ResNet18 | 1200 | 4 | 0.1332 | 0.0007 | - | ||
EfficientNet-B3 | 1183 | 5 | 0.1314 | 0.0018 | - | ||
Swin-T | 1201 | 4 | 0.1318 | 0.0020 | - | ||
ConvNeXt-T | 1259 | 2 | 0.1273 | 0.0014 | - | ||
CodeCarbon (evaluated meter) | AlexNet | 1199 | 4 | 0.0907 | 0.0003 | ||
VGG16 | 1197 | 3 | 0.1155 | 0.0005 | |||
ResNet18 | 1201 | 4 | 0.1154 | 0.0005 | |||
EfficientNet-B3 | 1183 | 5 | 0.1160 | 0.0003 | |||
Swin-T | 1201 | 5 | 0.1170 | 0.0002 | |||
ConvNeXt-T | 1259 | 1 | 0.1149 | 0.0002 | |||
Carbontracker (evaluated meter) | AlexNet | 1198 | 4 | 0.1218 | 0.0037 | ||
VGG16 | 1197 | 3 | 0.1537 | 0.0020 | |||
ResNet18 | 1201 | 4 | 0.1542 | 0.0025 | |||
EfficientNet-B3 | 1183 | 5 | 0.1545 | 0.0016 | |||
Swin-T | 1201 | 5 | 0.1551 | 0.0019 | |||
ConvNeXt-T | 1259 | 2 | 0.1509 | 0.0018 | |||
GTX 1080 Ti | OpenZmeter (reference meter) | AlexNet | 1200 | 3 | 0.1063 | 0.0011 | - |
VGG16 | 1191 | 3 | 0.1225 | 0.0013 | - | ||
ResNet18 | 1201 | 5 | 0.1247 | 0.0004 | - | ||
EfficientNet-B3 | 1199 | 7 | 0.1210 | 0.0018 | - | ||
Swin-T | 1206 | 3 | 0.1238 | 0.0033 | - | ||
ConvNeXt-T | 1384 | 2 | 0.1247 | 0.0021 | - | ||
CodeCarbon (evaluated meter) | AlexNet | 1201 | 4 | 0.0921 | 0.0007 | ||
VGG16 | 1192 | 3 | 0.1107 | 0.0003 | |||
ResNet18 | 1201 | 5 | 0.1110 | 0.0003 | |||
EfficientNet-B3 | 1199 | 7 | 0.1102 | 0.0003 | |||
Swin-T | 1206 | 3 | 0.1112 | 0.0001 | |||
ConvNeXt-T | 1385 | 2 | 0.1126 | 0.0010 | |||
Carbontracker (evaluated meter) | AlexNet | 1200 | 3 | 0.1241 | 0.0042 | ||
VGG16 | 1191 | 3 | 0.1454 | 0.0024 | |||
ResNet18 | 1201 | 5 | 0.1452 | 0.0042 | |||
EfficientNet-B3 | 1199 | 7 | 0.1450 | 0.0027 | |||
Swin-T | 1206 | 3 | 0.1469 | 0.0020 | |||
ConvNeXt-T | 1385 | 2 | 0.1492 | 0.0021 |
GPU | Energy Meter | Architecture | Kappa | Energy Consumed (kWh) | Kappa–Energy Index | p-Values | |||
---|---|---|---|---|---|---|---|---|---|
Average | Std. Dev. | Average | Std. Dev. | Average | Std. Dev. | ||||
TITAN Xp | OpenZmeter | AlexNet | 0.6427 | 0.0032 | 0.1062 | 0.0005 | 6.0518 | 0.0483 | - |
VGG16 | 0.5602 | 0.0000 | 0.1319 | 0.0010 | 4.2472 | 0.0339 | - | ||
ResNet18 | 0.7958 | 0.0023 | 0.1332 | 0.0007 | 5.9745 | 0.0281 | - | ||
EfficientNet-B3 | 0.7256 | 0.0000 | 0.1314 | 0.0018 | 5.5221 | 0.0772 | - | ||
Swin-T | 0.5444 | 0.0000 | 0.1318 | 0.0020 | 4.1305 | 0.0601 | - | ||
ConvNeXt-T | 0.1567 | 0.0000 | 0.1273 | 0.0014 | 1.2310 | 0.0133 | - | ||
CodeCarbon | AlexNet | 0.6427 | 0.0032 | 0.0907 | 0.0003 | 7.0860 | 0.0528 | ||
VGG16 | 0.5602 | 0.0000 | 0.1155 | 0.0005 | 4.8502 | 0.0207 | |||
ResNet18 | 0.7958 | 0.0023 | 0.1154 | 0.0005 | 6.8960 | 0.0273 | |||
EfficientNet-B3 | 0.7256 | 0.0000 | 0.1160 | 0.0003 | 6.2552 | 0.0149 | |||
Swin-T | 0.5444 | 0.0000 | 0.1170 | 0.0002 | 4.6530 | 0.0103 | |||
ConvNeXt-T | 0.1567 | 0.0000 | 0.1149 | 0.0002 | 1.3638 | 0.0031 | |||
GTX 1080 Ti | OpenZmeter | AlexNet | 0.6545 | 0.0012 | 0.1063 | 0.0011 | 6.1571 | 0.0694 | - |
VGG16 | 0.5515 | 0.0000 | 0.1225 | 0.0013 | 4.5020 | 0.0479 | - | ||
ResNet18 | 0.7946 | 0.0031 | 0.1247 | 0.0004 | 6.3721 | 0.0361 | - | ||
EfficientNet-B3 | 0.7411 | 0.0000 | 0.1210 | 0.0018 | 6.1248 | 0.0682 | - | ||
Swin-T | 0.4867 | 0.0000 | 0.1238 | 0.0033 | 3.9313 | 0.1005 | - | ||
ConvNeXt-T | 0.1600 | 0.0000 | 0.1247 | 0.0021 | 1.2831 | 0.0219 | - | ||
CodeCarbon | AlexNet | 0.6545 | 0.0012 | 0.0921 | 0.0007 | 7.1064 | 0.0595 | ||
VGG16 | 0.5515 | 0.0000 | 0.1107 | 0.0003 | 4.9819 | 0.0156 | |||
ResNet18 | 0.7937 | 0.0024 | 0.1110 | 0.0003 | 7.1505 | 0.0363 | |||
EfficientNet-B3 | 0.7411 | 0.0000 | 0.1102 | 0.0003 | 6.7250 | 0.0156 | |||
Swin-T | 0.4867 | 0.0000 | 0.1112 | 0.0001 | 4.3768 | 0.0042 | |||
ConvNeXt-T | 0.1600 | 0.0000 | 0.1126 | 0.0010 | 1.4210 | 0.0219 |
GPU | Comparison | AlexNet | VGG16 | ResNet18 | EfficientNet-B3 | Swin-T | ConvNeXt-T |
---|---|---|---|---|---|---|---|
TITAN Xp | AlexNet | - | - | ||||
VGG16 | - | - | - | ||||
ResNet18 | - | - | - | ||||
EfficientNet-B3 | - | - | - | ||||
Swin-T | - | - | - | ||||
ConvNeXt-T | - | - | |||||
GTX 1080 Ti | AlexNet | - | - | ||||
VGG16 | - | - | |||||
ResNet18 | - | ||||||
EfficientNet-B3 | - | - | |||||
Swin-T | - | - | - | ||||
ConvNeXt-T | - | - |
GPU | Architecture | Execution Times (s) | Kappa | Energy Consumed (kWh) | Kappa–Energy Index | |
---|---|---|---|---|---|---|
Average | Std. Dev. | |||||
TITAN Xp | AlexNet | 8.6000 | 0.6600 | 0.00060 | 1100.0000 | 117.4749 |
VGG16 | 11.8000 | 0.5589 | 0.00090 | 621.0000 | 53.3947 | |
ResNet18 | 8.4667 | 0.8089 | 0.00060 | 1348.1667 | 152.5554 | |
EfficientNet-B3 | 8.6667 | 0.7256 | 0.00070 | 1036.5714 | 87.6063 | |
Swin-T | 10.6000 | 0.5444 | 0.00080 | 680.5000 | 50.2012 | |
ConvNeXt-T | 10.7333 | 0.1567 | 0.00090 | 174.1111 | 14.2351 | |
GTX 1080 Ti | AlexNet | 8.0000 | 0.6744 | 0.00060 | 1124.0000 | 102.8994 |
VGG16 | 12.0667 | 0.5633 | 0.00100 | 563.3000 | 32.3208 | |
ResNet18 | 7.8000 | 0.8156 | 0.00050 | 1631.2000 | 140.3914 | |
EfficientNet-B3 | 9.1333 | 0.7411 | 0.00070 | 1058.7143 | 90.5074 | |
Swin-T | 10.8000 | 0.4867 | 0.00090 | 540.7778 | 30.9418 | |
ConvNeXt-T | 11.2000 | 0.1600 | 0.00090 | 177.7778 | 9.4729 |
GPU | Architecture | Batch Size | Kappa | Energy Consumed | Kappa-Energy | Energy Consumed | Energy-Cuda |
---|---|---|---|---|---|---|---|
Average | (Global, kWh) * | Index (Global) | (NVML, kWh) † | Core (NVML) † | |||
TITAN Xp | AlexNet | 32 | 0.7756 | 0.1100 | 7.0509 | 0.0600 | |
64 | 0.6427 | 0.1062 | 6.0518 | 0.0548 | |||
VGG16 | 32 | 0.6244 | 0.1349 | 4.6286 | 0.0862 | ||
64 | 0.5602 | 0.1319 | 4.2472 | 0.0836 | |||
ResNet18 | 32 | 0.8417 | 0.1335 | 6.3049 | 0.0818 | ||
64 | 0.7958 | 0.1332 | 5.9745 | 0.0795 | |||
EfficientNet-B3 | 32 | 0.7763 | 0.1319 | 5.8855 | 0.0851 | ||
64 | 0.7256 | 0.1314 | 5.5221 | 0.0840 | |||
Swin-T | 32 | 0.6411 | 0.1276 | 5.0243 | 0.0820 | ||
64 | 0.5444 | 0.1318 | 4.1305 | 0.0846 | |||
ConvNeXt-T | 32 | 0.1633 | 0.1314 | 1.2428 | 0.0825 | ||
64 | 0.1567 | 0.1273 | 1.2310 | 0.0793 | |||
GTX 1080 Ti | AlexNet | 32 | 0.7867 | 0.1064 | 7.3938 | 0.0580 | |
64 | 0.6545 | 0.1063 | 6.1571 | 0.0560 | |||
VGG16 | 32 | 0.6411 | 0.1242 | 5.1618 | 0.0813 | ||
64 | 0.5515 | 0.1225 | 4.5020 | 0.0789 | |||
ResNet18 | 32 | 0.8375 | 0.1172 | 7.1459 | 0.0706 | ||
64 | 0.7946 | 0.1247 | 6.3721 | 0.0755 | |||
EfficientNet-B3 | 32 | 0.7444 | 0.1180 | 6.3085 | 0.0756 | ||
64 | 0.7411 | 0.1210 | 6.1248 | 0.0776 | |||
Swin-T | 32 | 0.5882 | 0.1237 | 4.7551 | 0.0801 | ||
64 | 0.4867 | 0.1238 | 3.9313 | 0.0784 | |||
ConvNeXt-T | 32 | 0.1744 | 0.1168 | 1.4932 | 0.0711 | ||
64 | 0.1600 | 0.1247 | 1.2831 | 0.0752 |
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
Aquino-Brítez, S.; García-Sánchez, P.; Ortiz, A.; Aquino-Brítez, D. Towards an Energy Consumption Index for Deep Learning Models: A Comparative Analysis of Architectures, GPUs, and Measurement Tools. Sensors 2025, 25, 846. https://doi.org/10.3390/s25030846
Aquino-Brítez S, García-Sánchez P, Ortiz A, Aquino-Brítez D. Towards an Energy Consumption Index for Deep Learning Models: A Comparative Analysis of Architectures, GPUs, and Measurement Tools. Sensors. 2025; 25(3):846. https://doi.org/10.3390/s25030846
Chicago/Turabian StyleAquino-Brítez, Sergio, Pablo García-Sánchez, Andrés Ortiz, and Diego Aquino-Brítez. 2025. "Towards an Energy Consumption Index for Deep Learning Models: A Comparative Analysis of Architectures, GPUs, and Measurement Tools" Sensors 25, no. 3: 846. https://doi.org/10.3390/s25030846
APA StyleAquino-Brítez, S., García-Sánchez, P., Ortiz, A., & Aquino-Brítez, D. (2025). Towards an Energy Consumption Index for Deep Learning Models: A Comparative Analysis of Architectures, GPUs, and Measurement Tools. Sensors, 25(3), 846. https://doi.org/10.3390/s25030846