Pith. sign in

REVIEW 1 major objections 3 minor 64 references

Ground-Truthing AI Energy Consumption: Validating CodeCarbon Against External Measurements

T0 review · 1 major / 3 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read CodeCarbon and the ML Emissions Calculator misreport AI energy use, with dynamic estimates consistently low by 20–30% and static estimates off by up to 40%.

desk verdict A genuinely useful first-of-its-kind validation dataset for CodeCarbon and static energy estimators, but the headline error percentages rest on an uncalibrated plug-meter/OCR pipeline whose resolution is never reported. read the letter →

arxiv 2509.22092 v1 pith:AXVTCIUH submitted 2025-09-26 cs.AI

Raphael Fischer This is my paper · ORCID
classification cs.AI
keywords energyconsumptionestimationCodeCarbonMLEmissionsCalculatorsustainableAIfootprintaccountinginferencemeasurementground-truthvalidationresource-awaremachinelearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to answer a simple question: when CodeCarbon and the ML Emissions Calculator report the energy used by an AI model, how close are those numbers to what the hardware actually draws at the wall? By running 30 vision classifiers and 20 large language models on a single workstation, reading a plug-in energy meter with a camera and OCR as ground truth, and comparing those readings with the two estimation tools across hundreds of runs, the paper finds that both tools follow the general shape of energy demand but miss by large margins. Static estimates under- or overestimate consumption by up to 40 percent, while dynamic estimates consistently undercount by 20–30 percent. If correct, published carbon-footprint figures for local AI inference are mostly understated, and the error grows with model size, runtime, and CPU-only deployment.

What carries the argument

The load-bearing machinery is a three-way comparison loop run on the same inference jobs. Ground truth is obtained externally: a plug-in energy meter between wall and workstation, photographed by a camera, with the digits decoded by a custom random-forest OCR classifier. Static estimation uses the ML Impact Calculator formula $\text{Energy} = \text{Power} \times \text{Time}$ with fixed TDP values (300 W GPU, 125 W CPU); dynamic estimation uses CodeCarbon 3.0.1 to sample CPU/GPU power at intervals and sum $\text{Power}_t \cdot \Delta t$. The paper's numbers all ride on the difference between those two estimates and the meter readings.

What would settle it

Run the same 50-model battery on the same workstation while logging wall power with a calibrated power analyzer sampling at least once per second, and compare the analyzer's integrated kWh against the plug-in meter display readings; if the display readings differ from the analyzer by less than the claimed 20–40% error bands, the percentages lose their ground, and if CodeCarbon's dynamic estimates match the analyzer, the underestimation claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the energy numbers produced by the two most widely used estimation tools are not reliable as quantitative accounts even though they are reliable as qualitative trends. CodeCarbon's dynamic profiling captures CPU/GPU power but omits the power supply, cooling, and peripherals, so it consistently reports less than the meter sees; the ML Impact Calculator's static TDP assumption overshoots small models, undershoots large ones, and lands near the truth only for mid-sized models. The measured maximum draw, 534 W for a 24-billion-parameter model, is nearly double the static estimate's 300 W assumption. The paper frames this as a validation framework: any practitioner can reproduce the three-way comparison with an off-the-shelf meter, a camera, and OCR.

Load-bearing premise

The comparison assumes the plug-in energy meter plus camera/OCR pipeline yields true energy consumption, yet the paper does not state the meter's resolution or calibration or the OCR classifier's accuracy, and for short high-power runs the per-run energy total can sit near the meter's display step.

Editorial extensions

If this is right

  • Carbon-footprint reports that rely on CodeCarbon for local inference are likely too low by roughly a fifth to a third, not by a rounding error.
  • Static estimates like the ML Impact Calculator are not uniformly wrong: they are most accurate for mid-sized models and systematically off for small and large ones, so users should match the tool to model size or apply correction factors.
  • Estimation error scales with model size and with how long a model runs, making longer deployments the ones most in need of measurement.
  • CPU-only inference is where both tools fail worst; the paper's largest CPU dynamic error was hundreds of times the GPU error for the same model.
  • Hyperparameters that affect utilization, such as batch size and sampling temperature, change both the true energy draw and the size of estimator error, so energy reports should state those settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the dynamic underestimation is consistent across nearly all 50 models, a simple multiplicative correction of order 1.25–1.4 might bring CodeCarbon's local-inference numbers much closer to measured draw without changing its architecture.
  • The same meter-and-camera validation could be extended to training runs and to multi-GPU or data-center settings, where outlet-level power distribution units would replace the plug-in meter; the 20–30% and 40% figures would likely shift, but the direction of dynamic underestimation may persist.
  • A direct test of the paper's mechanism: if cooling and power-supply overhead are the cause, then a workstation with a more efficient PSU and better cooling should show smaller CodeCarbon underestimation; that is a prediction the paper does not test.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

1 major / 3 minor

Summary. The manuscript validates two widely used AI energy-estimation tools, the ML Impact Calculator (static TDP-based estimates) and CodeCarbon (dynamic hardware profiling), against external wall-plug measurements. The authors build a low-cost ground-truth pipeline consisting of a LogiLink plug-in energy meter, a Logitech camera, and a custom random-forest OCR classifier, and apply it to 30 vision classifiers and 20 language models under different batch sizes, temperatures, and CPU/GPU settings. They report that static estimates under- or overestimate by roughly -40% to +40%, while dynamic estimates consistently underestimate by 20-30%, and they derive six takeaways about how energy demand and estimation error scale with model size, hyperparameters, and processor choice.

Significance. If the quantitative error ranges are reliable, this is a useful and timely result for sustainable-AI practice: it would demonstrate that CodeCarbon-based carbon footprint reports for local inference are systematically understated in most cases. The paper has concrete strengths: it is reproducible in principle (code and logs are provided), the experiment spans 50 models across two domains, the comparison is not circular because no constants are fitted from tool outputs, and the general qualitative pattern (dynamic estimates track ground truth but with a downward bias) is plausible and consistent with known sources of overhead such as power supplies and cooling. However, the headline quantitative percentages rest entirely on the accuracy of the external meter/OCR pipeline, and that pipeline is not adequately characterized in the manuscript.

major comments (1)
  1. [§2.3, §2.4, §4 (Takeaways 3 and 4), Figure 1] This is a load-bearing issue: Figure 1 and Section 4, Takeaways 3 and 4, depend entirely on the external measurements from Sections 2.3-2.4.
minor comments (3)
  1. [§2.4] Minor presentation/clarity issue.
  2. [§2.4, Figure 4] Minor presentation/clarity issue.
  3. [§3, Figures 4 and 5] Minor presentation/clarity issue.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper compares tool outputs against independent external meter measurements, and none of the compared quantities are fitted or derived from the ground-truth data.

full rationale

The paper's central claim is that static and dynamic energy estimators deviate from measured ground-truth consumption by up to 40% (static) and 20–30% (dynamic). The derivation chain is a direct comparison: the static estimate uses the tools' own constants (e.g., a 300 W GPU TDP from the ML Impact Calculator and a 125 W CPU value from CodeCarbon's cpu_power.csv), the dynamic estimate uses CodeCarbon 3.0.1's profiling output, and the ground-truth is obtained from an external LogiLink energy meter read by a camera and OCR pipeline. None of these values are fitted to the measured data, and no equation in the paper defines the estimates in terms of the measurements. The paper does cite several works by the same author, but those citations concern terminology, framing, and prior context, not the load-bearing empirical comparison; the validation result stands on the external measurements independent of those citations. The main weakness identified by a skeptical reading—the unstated accuracy and display resolution of the plug-in meter and OCR classifier—is a measurement-validity concern, not a circularity concern. Even if the meter quantizes at 0.01 kWh, the comparison structure remains non-circular; the error percentages might be imprecise, but they are not forced by construction. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim is an empirical comparison, not a derivation. It introduces no fitted coefficients or new entities. The main postulates are the validity of the meter and OCR ground-truth pipeline and the representative tool usage. The static power values are not fitted by the paper but chosen from the tools' own defaults.

free parameters (3)
  • Static power draw assumed for GPU (TDP) = 300 W
    Chosen from the ML Impact Calculator as the constant power for static estimation; the paper compares this against measured wall power.
  • Static power draw assumed for CPU = 125 W
    Taken from CodeCarbon's cpu_power.csv; used for CPU-only static estimates.
  • CO2 efficiency factor for Germany = 0.38 kg CO2/kWh
    Used to convert measured energy into emissions; taken from Zeppelin conversion factors, only for the study's own footprint estimate, not for the central validation.
assumptions (3)
  • domain assumption The external plug-in energy meter reading, captured via camera and OCR, is an accurate measurement of total system energy consumption.
    This is the ground-truth basis for every comparison in Sections 3 and 4; the paper does not report meter resolution, calibration, or OCR accuracy.
  • domain assumption CodeCarbon and the ML Impact Calculator were invoked in a way representative of ordinary practitioner use.
    The comparison intends to validate the tools as used; any configuration mistakes would shift the error magnitudes.
  • domain assumption The i9-13900K and RTX 4090 workstation is representative enough to support general claims about estimation errors.
    All experiments run on one environment; the paper acknowledges this limitation in the Discussion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ground-Truthing AI Energy Consumption: Validating CodeCarbon Against External Measurements." pith.science (2026). https://pith.science/paper/AXVTCIUH

@misc{pith2026250922092,
  author       = {Pith},
  title        = {Pith review of: Ground-Truthing AI Energy Consumption: Validating CodeCarbon Against External Measurements},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AXVTCIUH}},
  note         = {Machine review of arXiv:2509.22092}
}
read the original abstract

Although machine learning (ML) and artificial intelligence (AI) present fascinating opportunities for innovation, their rapid development is also significantly impacting our environment. In response to growing resource-awareness in the field, quantification tools such as the ML Emissions Calculator and CodeCarbon were developed to estimate the energy consumption and carbon emissions of running AI models. They are easy to incorporate into AI projects, however also make pragmatic assumptions and neglect important factors, raising the question of estimation accuracy. This study systematically evaluates the reliability of static and dynamic energy estimation approaches through comparisons with ground-truth measurements across hundreds of AI experiments. Based on the proposed validation framework, investigative insights into AI energy demand and estimation inaccuracies are provided. While generally following the patterns of AI energy consumption, the established estimation approaches are shown to consistently make errors of up to 40%. By providing empirical evidence on energy estimation quality and errors, this study establishes transparency and validates widely used tools for sustainable AI development. It moreover formulates guidelines for improving the state-of-the-art and offers code for extending the validation to other domains and tools, thus making important contributions to resource-aware ML and AI sustainability research.

Figures

Figures reproduced from arXiv: 2509.22092 by the authors.

Figure 1
Figure 1. Compared to ground-truth energy measurements, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Schematic visualization for validating the static and dynamic estimation of energy demand via external measurements. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Hardware setup for validating energy estimation with ground-truth data from an energy meter, tracked with a camera. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Power draw of the execution environment (first row), energy draw for performing inference with a fixed number of [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Impact of hyperparameter choice (x-axes) for AI energy demand. The higher batch size and temperature re￾sults in more efficient compute utilization (i.e., lower energy draw) for vision and language models, which also lead to lower estimation errors. truth energy demand…
Figure 6
Figure 6. Figure 6: Absolute energy estimation errors of vision models [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 29 canonical work pages

  1. [1]

    2024.DOI:10.48550/arXiv.2404.14219

    Marah Abdin et al.Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone. 2024.DOI:10.48550/arXiv.2404.14219

    arXiv doi:10.48550/arxiv.2404.14219 2024

  2. [2]

    2024.DOI: 10.48550/arXiv.2412.08905

    Marah Abdin et al.Phi-4 Technical Report. 2024.DOI: 10.48550/arXiv.2412.08905

    arXiv doi:10.48550/arxiv.2412.08905 2024

  3. [3]

    Yoshua Bengio et al.International AI safety report. Tech. rep. Accessed: 2025-09-12. AI Safety Institute, 2025.URL:https://coilink.org/20.500. 12592/30e27x2

    work page 2025

  4. [4]

    Tradi- tional Vs Smart Electricity Metering Systems: A Brief Overview

    Samuel Bimenyimana and Godwin Asemota. “Tradi- tional Vs Smart Electricity Metering Systems: A Brief Overview”. In:Journal of Marketing and Consumer Research46 (June 2018).URL:https : / / www . iiste . org / Journals / index . php / JMCR / article/view/42505

    work page 2018

  5. [5]

    The Values Encoded in Ma- chine Learning Research

    Abeba Birhane et al. “The Values Encoded in Ma- chine Learning Research”. In:5th Conference on Fair- ness, Accountability and Transparency (FAccT). 2022, pp. 173–184.ISBN: 978-1-4503-9352-2.DOI:10 . 1145/3531146.3533083

    arXiv 2022

  6. [6]

    The OpenCV Library

    G. Bradski. “The OpenCV Library”. In:Dr. Dobb’s Journal of Software Tools(2000).URL:https:// www . drdobbs . com / open - source / the - opencv-library/184404319

    work page arXiv 2000

  7. [7]

    Accessed: 2025-09-12

    Michael Chiang and Jeffrey Morgan.Ollama: Chat & build with open models. Accessed: 2025-09-12. 2023. URL:https://ollama.com/

    work page 2025

  8. [8]

    Accessed: 2025-09-12

    François Chollet et al.Keras. Accessed: 2025-09-12. 2015.URL:https://keras.io. 8

    work page 2025

Show all 64 references
  1. [9]

    Xception: Deep Learning with Depthwise Separable Convolutions

    François Chollet. “Xception: Deep Learning with Depthwise Separable Convolutions”. In:Conference on Computer Vision and Pattern Recognition (CVPR). 2017, pp. 1800–1807.DOI:10 . 1109 / CVPR . 2017.195

    2017

  2. [10]

    May 2024.DOI:10.5281/zenodo.11171501

    Benoit Courty et al.mlco2/codecarbon: v2.4.1. May 2024.DOI:10.5281/zenodo.11171501

    2024 doi

  3. [11]

    2025.DOI:10.48550/arXiv.2501.12948

    DeepSeek-AI.DeepSeek-R1: Incentivizing Reason- ing Capability in LLMs via Reinforcement Learning. 2025.DOI:10.48550/arXiv.2501.12948

    2025 doi arXiv

  4. [12]

    Imagenet: A Large-Scale Hierarchical Image Database

    Jia Deng et al. “Imagenet: A Large-Scale Hierarchical Image Database”. In:Conference on Computer Vision and Pattern Recognition (CVPR). 2009, pp. 248–255. DOI:10.1109/CVPR.2009.5206848

    2009

  5. [13]

    2025.DOI:10.48550/arXiv

    Sophia Falk et al.More than Carbon: Cradle-to- Grave environmental impacts of GenAI training on the Nvidia A100 GPU. 2025.DOI:10.48550/arXiv. 2509.00093

    2025 doi

  6. [14]

    Advancing the Sustainability of Machine Learning and Artificial Intelligence via La- beling and Meta-Learning

    Raphael Fischer. “Advancing the Sustainability of Machine Learning and Artificial Intelligence via La- beling and Meta-Learning”. PhD thesis. TU Dort- mund University, 2025.DOI:10.17877/DE290R- 25716

    2025 doi

  7. [15]

    Towards More Sustainable and Trustworthy Reporting in Machine Learning

    Raphael Fischer, Thomas Liebig, and Katharina Morik. “Towards More Sustainable and Trustworthy Reporting in Machine Learning”. In:Data Mining and Knowledge Discovery(2024).ISSN: 1573-756X.DOI: 10.1007/s10618-024-01020-3

    2024 doi

  8. [16]

    AutoXPCR: Automated Multi-Objective Model Selection for Time Series Forecasting

    Raphael Fischer and Amal Saadallah. “AutoXPCR: Automated Multi-Objective Model Selection for Time Series Forecasting”. In:Proceedings of the 30th International Conference on Knowledge Discovery and Data Mining (KDD). 2024, pp. 806–815.ISBN: 979-8-4007-0490-1.DOI:10 . 1145 / 36...

    2024

  9. [17]

    A Unified Framework for As- sessing Energy Efficiency of Machine Learning

    Raphael Fischer et al. “A Unified Framework for As- sessing Energy Efficiency of Machine Learning”. In: Machine Learning and Principles and Practice of Knowledge Discovery in Databases. 2022, pp. 39–54. DOI:10.1007/978-3-031-23618-1_3

    2022 doi

  10. [18]

    MetaQuRe: Meta-learning from Model Quality and Resource Consumption

    Raphael Fischer et al. “MetaQuRe: Meta-learning from Model Quality and Resource Consumption”. In:Machine Learning and Knowledge Discovery in Databases. 2024, pp. 209–226.ISBN: 978-3-031- 70368-3.DOI:10.1007/978- 3- 031- 70368- 3_13

    2024 doi

  11. [19]

    Estimation of Energy Con- sumption in Machine Learning

    Eva García-Martín et al. “Estimation of Energy Con- sumption in Machine Learning”. In:Journal of Par- allel and Distributed Computing134 (Dec. 2019), pp. 75–88.ISSN: 0743-7315.DOI:10 . 1016 / j . jpdc.2019.07.007

    2019

  12. [20]

    2025.DOI: 10.48550/arXiv.2503.19786

    Gemma Team.Gemma 3 Technical Report. 2025.DOI: 10.48550/arXiv.2503.19786

    2025 doi arXiv

  13. [21]

    2024.DOI:10.48550/arXiv.2407.21783

    Aaron Grattafiori et al.The Llama 3 Herd of Models. 2024.DOI:10.48550/arXiv.2407.21783

    2024 doi arXiv

  14. [22]

    Deep Residual Learning for Image Recognition

    Kaiming He et al. “Deep Residual Learning for Image Recognition”. In:Conference on Computer Vision and Pattern Recognition (CVPR). 2016, pp. 770–778.DOI: 10.1109/CVPR.2016.90

    2016 doi

  15. [23]

    Identity Mappings in Deep Resid- ual Networks

    Kaiming He et al. “Identity Mappings in Deep Resid- ual Networks”. In:Proceedings of the 14th European Conference on Computer Vision (ECCV). Ed. by Bas- tian Leibe et al. V ol. 9908. 2016, pp. 630–645.DOI: 10.1007/978-3-319-46493-0_38

    2016 doi

  16. [24]

    Towards the Systematic Re- porting of the Energy and Carbon Footprints of Ma- chine Learning

    Peter Henderson et al. “Towards the Systematic Re- porting of the Energy and Carbon Footprints of Ma- chine Learning”. In:Journal of Machine Learning Re- search21 (248) (2020), pp. 1–43.DOI:10.48550/ arXiv.2002.05651

    2020 doi arXiv

  17. [25]

    Towards Power Efficiency in Deep Learning on Data Center Hardware

    Miro Hodak, Masha Gorkovenko, and Ajay Dholakia. “Towards Power Efficiency in Deep Learning on Data Center Hardware”. In:International Conference on Big Data. 2019, pp. 1814–1820.DOI:10 . 1109 / BigData47090.2019.9005632

    2019

  18. [26]

    2017.DOI:10

    Andrew Howard et al.Mobilenets: Efficient Convo- lutional Neural Networks for Mobile Vision Appli- cations. 2017.DOI:10 . 48550 / arXiv . 1704 . 04861

    2017

  19. [27]

    Searching for MobileNetV3

    Andrew Howard et al. “Searching for MobileNetV3”. In:International Conference on Computer Vision (ICCV). Oct. 2019.DOI:10.1109/ICCV.2019. 00140. 9

    2019 doi

  20. [28]

    Densely Connected Convolutional Networks

    Gao Huang et al. “Densely Connected Convolutional Networks”. In:Conference on Computer Vision and Pattern Recognition (CVPR). 2017, pp. 2261–2269. DOI:10.1109/CVPR.2017.243

    2017 doi

  21. [29]

    Accessed: 2025-09-08

    Intel Corporation.Intel ® 64 and IA-32 Architectures Software Developer’s Manual. Accessed: 2025-09-08. 2024.URL:https : / / intel . com / content / www / us / en / developer / articles / technical/intel-sdm.html

    2025

  22. [30]

    Jiang et al.Mistral 7B

    Albert Q. Jiang et al.Mistral 7B. 2023.DOI:10 . 48550/arXiv.2310.06825

    2023 doi arXiv

  23. [31]

    How Can Artificial Intelligence Impact Sustainability: A Systematic Literature Re- view

    Arpan Kumar Kar, Shweta Kumari Choudhary, and Vinay Kumar Singh. “How Can Artificial Intelligence Impact Sustainability: A Systematic Literature Re- view”. In:Journal of Cleaner Production(2022).DOI: 10.1016/j.jclepro.2022.134120

    2022

  24. [32]

    2025.DOI:10

    Imran Latif et al.Cooling Matters: Benchmarking Large Language Models and Vision-Language Mod- els on Liquid-Cooled Versus Air-Cooled H100 GPU Systems. 2025.DOI:10 . 48550 / arXiv . 2507 . 16781

    2025

  25. [33]

    Artificial Intelligence as a Ser- vice

    Sebastian Lins et al. “Artificial Intelligence as a Ser- vice”. In:Business & Information Systems Engineer- ing63 (4) (Aug. 2021), pp. 441–456.ISSN: 1867- 0202.DOI:10.1007/s12599-021-00708-w

    2021 doi

  26. [34]

    A ConvNet for the 2020s

    Zhuang Liu et al. “A ConvNet for the 2020s”. In:Con- ference on Computer Vision and Pattern Recognition (CVPR). 2022, pp. 11966–11976.DOI:10 . 1109 / CVPR52688.2022.01167

    2022

  27. [36]

    Power Hungry Processing: Watts Driving the Cost of AI Deployment?

    Sasha Luccioni, Yacine Jernite, and Emma Strubell. “Power Hungry Processing: Watts Driving the Cost of AI Deployment?” In:7th Conference on Fair- ness, Accountability and Transparency (FAccT). 2024, pp. 85–99.ISBN: 9798400704505.DOI:10.1145/ 3630106.3658542

    2024

  28. [37]

    Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model

    Sasha Luccioni, Sylvain Viguier, and Anne-Laure Ligozat. “Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model”. In: Journal of Machine Learning Research24 (253) (2023), pp. 1–15.URL:https : / / jmlr . org / papers/v24/23-0069.html

    2023

  29. [38]

    Quantifying the Carbon Emis- sions of Machine Learning

    Sasha Luccioni et al. “Quantifying the Carbon Emis- sions of Machine Learning”. In:NeurIPS Workshop on Tackling Climate Change with Machine Learning. 2019.URL:https : / / climatechange . ai / papers/neurips2019/22

    2019

  30. [39]

    2023.URL:https : / / huggingface

    Jai Suphavadeeprasit Luigi Daniele.Puffin dataset. 2023.URL:https : / / huggingface . co / datasets/LDJnr/Puffin

    2023

  31. [40]

    2025.DOI:10

    Mistral-AI.Magistral. 2025.DOI:10 . 48550 / arXiv.2506.10910

    2025 doi arXiv

  32. [41]

    Accessed: 2025-09-08

    NVIDIA Corporation.NVIDIA Management Library (NVML). Accessed: 2025-09-08. 2012.URL:https: / / developer . nvidia . com / management - library-nvml

    2025

  33. [42]

    2025.DOI:10.48550/arXiv.2508.10925

    OpenAI.gpt-oss-120b & gpt-oss-20b Model Card. 2025.DOI:10.48550/arXiv.2508.10925

    2025 doi arXiv

  34. [43]

    Scikit-learn: Machine Learn- ing in Python

    Fabian Pedregosa et al. “Scikit-learn: Machine Learn- ing in Python”. In:Journal of Machine Learning Re- search (JMLR)12 (2011), pp. 2825–2830.DOI:10. 5555/1953048.2078195

    2011

  35. [44]

    Accessed: 2025-08-13

    Plotly Technologies Inc.Plotly: Collaborative Data Science and Visualization. Accessed: 2025-08-13. 2024.URL:https://plotly.com

    2025

  36. [45]

    MobileNetV2: Inverted Residu- als and Linear Bottlenecks

    Mark Sandler et al. “MobileNetV2: Inverted Residu- als and Linear Bottlenecks”. In:Conference on Com- puter Vision and Pattern Recognition (CVPR). 2018, pp. 4510–4520.DOI:10 . 1109 / CVPR . 2018 . 00474

    2018

  37. [46]

    Green AI

    Roy Schwartz et al. “Green AI”. In:Communications of the ACM63 (12) (Nov. 2020), pp. 54–63.ISSN: 0001-0782.DOI:10.1145/3381831

    2020 doi

  38. [47]

    Compute Trends Across Three Eras of Machine Learning

    Jaime Sevilla et al. “Compute Trends Across Three Eras of Machine Learning”. In:International Joint Conference on Neural Networks (IJCNN). 2022, pp. 1–8.DOI:10 . 1109 / IJCNN55064 . 2022 . 9891914. 10

    2022

  39. [48]

    Very Deep Convolutional Networks for Large-Scale Im- age Recognition

    Karen Simonyan and Andrew Zisserman. “Very Deep Convolutional Networks for Large-Scale Im- age Recognition”. In:3rd International Conference on Learning Representations (ICLR). Ed. by Yoshua Bengio and Yann LeCun. 2015.URL:https : / / arxiv.org/abs/1409.1556

    2015 arXiv

  40. [49]

    Stress-Testing USB Accelerators for Efficient Edge Inference

    Alexander van der Staay, Raphael Fischer, and Sebas- tian Buschjäger. “Stress-Testing USB Accelerators for Efficient Edge Inference”. In:Proceedings of the 9th Symposium on Edge Computing (SEC). 2024, pp. 1– 14.DOI:10.1109/SEC62691.2024.00015

    2024

  41. [50]

    Energy and Policy Considerations for Modern Deep Learning Research

    Emma Strubell, Ananya Ganesh, and Andrew McCal- lum. “Energy and Policy Considerations for Modern Deep Learning Research”. In:AAAI Conference on Artificial Intelligence34 (09) (Apr. 2020), pp. 13693– 13696.DOI:10.1609/aaai.v34i09.7123

    2020 doi

  42. [51]

    Inception-v4, Inception- ResNet and the Impact of Residual Connections on Learning

    Christian Szegedy et al. “Inception-v4, Inception- ResNet and the Impact of Residual Connections on Learning”. In:31st AAAI Conference on Artificial In- telligence. 2017, pp. 4278–4284.DOI:10 . 1609 / AAAI.V31I1.11231

    2017

  43. [52]

    Rethinking the Inception Ar- chitecture for Computer Vision

    Christian Szegedy et al. “Rethinking the Inception Ar- chitecture for Computer Vision”. In:Conference on Computer Vision and Pattern Recognition (CVPR). 2016, pp. 2818–2826.DOI:10 . 1109 / CVPR . 2016.308

    2016

  44. [53]

    EfficientNet: Rethink- ing Model Scaling for Convolutional Neural Net- works

    Mingxing Tan and Quoc Le. “EfficientNet: Rethink- ing Model Scaling for Convolutional Neural Net- works”. In:36th International Conference on Ma- chine Learning (ICML). Ed. by Kamalika Chaudhuri and Ruslan Salakhutdinov. V ol. 97. Proceedings of Machine Learning Research. June...

    2019

  45. [54]

    EfficientNetV2: Smaller Models and Faster Training

    Mingxing Tan and Quoc V . Le. “EfficientNetV2: Smaller Models and Faster Training”. In:38th Inter- national Conference on Machine Learning (ICML). Ed. by Marina Meila and Tong Zhang. V ol. 139. 2021, pp. 10096–10106.URL:http://proceedings. mlr.press/v139/tan21a.html

    2021

  46. [55]

    2025.DOI:10

    Team OLMo.2 OLMo 2 Furious. 2025.DOI:10 . 48550/arXiv.2501.00656

    2025 doi arXiv

  47. [56]

    The pandas development team.pandas-dev/pandas: Pandas. Feb. 2020.DOI:10 . 5281 / zenodo . 3509134

    2020

  48. [57]

    Hype, Sustainability, and the Price of the Bigger-is-Better Paradigm in AI

    Gael Varoquaux, Sasha Luccioni, and Meredith Whit- taker. “Hype, Sustainability, and the Price of the Bigger-is-Better Paradigm in AI”. In:8th Confer- ence on Fairness, Accountability and Transparency (FAccT). 2025, pp. 61–75.ISBN: 9798400714825. DOI:10.1145/3715275.3732006

    2025

  49. [58]

    Carole-Jean Wu et al.Sustainable AI: Environmen- tal Implications, Challenges and Opportunities. 2022. DOI:10.48550/arXiv.2111.00364

    2022 doi arXiv

  50. [59]

    Sustainable AI: AI for Sus- tainability and the Sustainability of AI

    Aimee van Wynsberghe. “Sustainable AI: AI for Sus- tainability and the Sustainability of AI”. In:AI and Ethics1 (3) (Aug. 2021), pp. 213–218.ISSN: 2730- 5961.DOI:10.1007/s43681-021-00043-6

    2021 doi

  51. [60]

    2025.DOI: 10.48550/arXiv.2505.09388

    An Yang et al.Qwen3 Technical Report. 2025.DOI: 10.48550/arXiv.2505.09388

    2025 doi arXiv

  52. [61]

    Accelerating the Machine Learning Lifecycle with MLflow

    Matei A. Zaharia et al. “Accelerating the Machine Learning Lifecycle with MLflow”. In:Data Engineer- ing Bulletin41 (2018), pp. 39–45.URL:https : //api.semanticscholar.org/CorpusID: 83459546

    2018

  53. [62]

    Accessed: 2025-09-12

    Zeppelin GmbH.Conversion Factors for CO 2 Emis- sions. Accessed: 2025-09-12. 2024.URL:https:// sustainabilityreport . zeppelin . com / 2024 / appendix / conversion - factors - co2-emissions/

    2025

  54. [63]

    2024.DOI:10

    Peiyuan Zhang et al.TinyLlama: An Open-Source Small Language Model. 2024.DOI:10 . 48550 / arXiv.2401.02385

    2024 doi arXiv

  55. [64]

    Energy-efficient LLM Training in GPU datacenters with Immersion Cooling Systems

    Shuntao Zhu and Dan Wang. “Energy-efficient LLM Training in GPU datacenters with Immersion Cooling Systems”. In:16th International Conference on Fu- ture and Sustainable Energy Systems. 2025, pp. 407– 414.ISBN: 9798400711251.DOI:10 . 1145 / 3679240.3734609

    2025

  56. [65]

    Learning Transferable Architec- tures for Scalable Image Recognition

    Barret Zoph et al. “Learning Transferable Architec- tures for Scalable Image Recognition”. In:Conference on Computer Vision and Pattern Recognition (CVPR). 2018, pp. 8697–8710.DOI:10 . 1109 / CVPR . 2018.00907. 11

    2018

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.