Pith. sign in

REVIEW 2 major objections 6 minor 28 references

Pre-training Lucie 7B cost 21 tCO2eq once manufacturing of the whole HPC partition is amortised, nearly half the annual carbon on a clean French grid.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 14:23 UTC pith:ABLPO6JR

load-bearing objection Solid, transparent v1 LCA that finally gives measured H100-partition intensities with subsystem-level embodied carbon; the amortisation windows are the only real soft spot and the authors already flag them. the 2 major comments →

arxiv 2607.05408 v1 pith:ABLPO6JR submitted 2026-06-09 cs.CY cs.LG

Life Cycle Assessment of Pre-training the Lucie 7B Open-Source Large Language Model on the Jean Zay Supercomputer

Marc L\'eobet , Pierre-Fran\c{c}ois Lavall\'ee , Jean-Pierre Lorr\'e This is my paper
classification cs.CY cs.LG
keywords Life Cycle AssessmentLarge Language ModelsPre-trainingEmbodied carbonWater footprintHPCFrugal AIJean Zay
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Most published numbers for training large language models count only the electricity used while the GPUs run. This paper does a full life-cycle assessment of the open-source Lucie 7B model on the Jean Zay H100 supercomputer partition, folding in the manufacturing of the compute nodes, storage, power chain and cooling system as well as on-site water and waste-heat recovery. It finds that the partition itself emits 417.5 tonnes of CO2-equivalent per year, split almost evenly between embodied and operational carbon, which works out to 36.7 grams per H100 GPU-hour. Lucie 7B’s 574 564 GPU-hours therefore total 21 tonnes once the hardware share is included. Because French electricity is already low-carbon, further cuts will have to come from longer hardware life, re-use and lower-carbon manufacturing rather than greener power alone. The warm-water direct liquid cooling and urban heat recovery already built into the site also deliver a water-use effectiveness of 0.07 L/kWh and an energy-reuse factor of 0.37, numbers the authors argue should become standard in every LLM environmental report.

Core claim

A bottom-up life-cycle assessment of the Jean Zay H100 partition yields an annual footprint of 417.5 tCO2eq (46 % manufacturing, 54 % operations) and therefore an intensity of 36.7 gCO2eq per H100 GPU-hour. Applying that intensity to the documented 574 564 GPU-hours of Lucie 7B pre-training produces a campaign total of 21.1 tCO2eq that already amortises the hardware, together with roughly 76 m3 of on-site water and a heat-reuse factor of 0.37.

What carries the argument

The LCA-inclusive per-GPU-hour intensity: annual manufacturing emissions of the entire partition (amortised over 10 y compute, 9 y storage, 25 y power, 20 y cooling) plus measured operational electricity, divided by the partition’s effective yearly GPU-hour budget.

Load-bearing premise

The result rests on fixed hardware lifetimes chosen to match current operator practice; if GPUs and nodes are retired from intensive use much sooner, the embodied share and the reported intensity rise sharply.

What would settle it

Re-run the same bottom-up inventory with independently audited lifetimes for the H100 nodes and storage tiers that are three years shorter; if the resulting intensity exceeds roughly 50 gCO2eq per GPU-hour, the 21-tonne claim no longer holds under the paper’s own arithmetic.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper presents a life-cycle assessment of pre-training Lucie 7B (574 564 H100 GPU-hours) on the Jean Zay H100 partition, framed by AFNOR SPEC 2314 and Labos 1point5. It reports an annual partition footprint of 417.5 tCO2eq (194.2 manufacturing + 223.5 operational), an LCA-inclusive intensity of 36.7 gCO2eq per H100 GPU-hour, a campaign total of 21.1 tCO2eq, on-site water of ~76 m3 (WUE 0.07 L/kWh), and ERF 0.37 from waste-heat recovery. Embodied emissions are decomposed by subsystem (compute, storage, power, cooling) with explicit amortisation lifetimes; operational electricity is scaled from 273 days of monitored data. Scope covers data preparation through validation; inference is deferred.

Significance. Public, infrastructure-specific LCAs of LLM training that couple measured HPC operational data with subsystem-decomposed embodied carbon remain rare. The work supplies a transparent European (low-carbon-grid) benchmark, documents warm-water DLC and heat-reuse effects that hyperscale-only metrics miss, and aligns with AFNOR SPEC 2314. Strengths include clean arithmetic, explicit allocation on GPU-hours, and open reporting of water and ERF/ERE. If the figures hold under the stated conventions, the paper is a useful reference for frugal-AI design and for comparing future campaigns on similar European HPC systems.

major comments (2)
  1. Section 2.7 and Tables 1–2: manufacturing is annualised with fixed lifetimes (10 y compute, 9 y storage, 25 y power, 20 y cooling) that convert the 2169 t capital total into the 194.2 t/y share and therefore into the 36.7 g/GPU-h intensity. No sensitivity ranges are shown. Because the central claim is the LCA-inclusive intensity and the 21.1 t campaign total, a short sensitivity table (±2–3 y, or intensive-use-only lifetimes) is load-bearing; the paper itself flags this in §6. Without it the embodied/operational split (46/54) cannot be assessed for robustness.
  2. Section 4.1 and §6: the H100 GPU embodied figure (164 kgCO2eq) is taken from NVIDIA’s product carbon footprint, whose scope is acknowledged as partial. This value enters the per-node total (2056 kg) and the compute line of Table 1. A brief triangulation against ACT or foundry-level inventories (even as a one-sided bound) is needed so that the manufacturing share is not understated by an unquantified amount.
minor comments (6)
  1. Section 4.3: the effective annual GPU-hour budget is given as “≈11.39 million” (88 % of 12.75 M). State the exact utilisation figure and source so the 36.7 g intensity is fully reproducible.
  2. Section 4.2: operational electricity is scaled from 273 days of 2025 data. A one-sentence note on seasonal representativeness (or a simple uncertainty band) would strengthen the 223.5 t figure.
  3. Section 4.5: off-site EWIF (0.86 L/kWh) is documented but excluded from the headline water total. Either add it as a secondary line or move the dual-accounting discussion earlier so readers do not misread the 76 m3 figure as complete.
  4. Figure 1 is referenced but not described in sufficient detail for a text-only reader; ensure panel labels match the 194.2 / 223.5 split and the Lucie allocation call-out.
  5. Section 5.2: the LLaMA-2 comparison is appropriately caveated; a short table of scope differences (grid, GPU generation, tokens, embodied inclusion) would make the non-comparability even clearer.
  6. Typographical consistency: “21 t” in the abstract vs “21.1 t” in the body; “gCO2eq” spacing and “tCO2eq” notation vary slightly across sections.

Circularity Check

0 steps flagged

No circularity: transparent inventory arithmetic from measured electricity, published emission factors, and amortised capital totals; no fitted parameters re-presented as predictions.

full rationale

The derivation chain is a standard bottom-up LCA inventory under AFNOR SPEC 2314 and Labos 1point5 conventions. Manufacturing totals (Table 1: compute 748 t, storage 541 t, power 268 t, cooling 612 t) are taken from Labos 1point5 figures plus the NVIDIA H100 PCF, then annualised by explicit technical lifetimes chosen to match IDRIS practice (10 y compute, 9 y storage, 25 y power, 20 y cooling) plus yearly fugitive flows, yielding 194.2 tCO2eq/y (Table 2). Operational emissions (223.5 tCO2eq/y) are scaled from monitored electricity (10 299 MWh/y for the H100 partition) times the French grid factor 21.7 gCO2eq/kWh. Sum = 417.5 tCO2eq/y; division by the effective annual GPU-hour budget (~11.39 M h at 88 % utilisation) produces the intensity 36.7 gCO2eq per H100 GPU-hour; multiplication by the documented campaign consumption 574 564 GPU-hours produces 21.1 tCO2eq. Water (WUE 0.07 L/kWh, ~76 m3) and ERF 0.37 are likewise direct ratios of measured quantities. No parameter is fitted to a subset of the target data and then re-used as a prediction; lifetimes and allocation rules are declared modelling choices (Sections 2.6–2.7, 4.1), not self-definitions; citations supply external emission factors and frameworks, not uniqueness theorems by the present authors. The paper itself flags the point-estimate lifetimes as a limitation for v2 sensitivity analysis. The calculation is therefore self-contained and non-circular.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central numerical claims rest on a small set of domain modelling choices (lifetimes, allocation rule, emission factors, system boundaries) taken from Labos 1point5 / IDRIS practice and on one partial vendor PCF; no new physical entities are postulated.

free parameters (5)
  • compute amortisation lifetime = 10 years
    Set to 10 years (6 intensive + 4 extended) by authors to match IDRIS practice; directly scales the 74.8 tCO2eq/y compute manufacturing term.
  • storage amortisation lifetime = 9 years
    Set to 9 years with 50 % re-conditioning credit; scales the 60.1 tCO2eq/y storage term.
  • power-chain amortisation lifetime = 25 years
    Set to 25 years; scales the 10.7 tCO2eq/y power term.
  • cooling amortisation lifetime = 20 years
    Set to 20 years; scales the 30.6 tCO2eq/y cooling term.
  • effective annual GPU-hour utilisation = ≈11.39 million GPU-h/y
    88 % of nominal 12.75 M h used as denominator for intensity; chosen from observed load.
axioms (5)
  • domain assumption GPU-hour proportional allocation of all partition overheads (storage, power, cooling, manufacturing) is a valid functional-unit basis.
    Stated in §2.6; standard in Labos 1point5 but not uniquely determined.
  • domain assumption French grid emission factor 21.7 gCO2eq/kWh (2024) applies to the entire operational electricity.
    Taken from RTE; used in §4.2 to convert 10 299 MWh into 223.5 tCO2eq.
  • domain assumption NVIDIA H100 product-carbon-footprint value of 164 kgCO2eq per GPU is an acceptable (if partial) manufacturing figure.
    Cited in §4.1 and Limitations; acknowledged as incomplete scope.
  • domain assumption Building construction, personnel, transport and end-of-life beyond amortisation are negligible or outside boundary.
    Explicit system-boundary choices in §2.3 following Arcep/Ademe and Labos 1point5.
  • domain assumption Waste-heat recovery is excluded from the GHG perimeter (Labos 1point5 convention) and reported only as ancillary ERF.
    §2.3 and §4.7; affects interpretation of net system benefit.

pith-pipeline@v1.1.0-grok45 · 15399 in / 3023 out tokens · 26950 ms · 2026-07-12T14:23:57.568614+00:00 · methodology

0 comments
read the original abstract

The environmental impact of training large language models (LLMs) is increasingly scrutinised, yet most published estimates focus on operational energy and disclose little about manufacturing (embodied) emissions, water consumption, or the underlying highperformance computing (HPC) infrastructure. We present a life cycle assessment (LCA) of the pre-training of Lucie 7B, an open-source multilingual Foundation Model developed by the OpenLLM-France consortium and trained on the NVIDIA H100 partition of the Jean Zay supercomputer operated by IDRIS (CNRS). The assessment is framed by the AFNOR SPEC 2314 "Frugal AI" reference and applies the Labos 1point5 methodology for greenhouse gas(GHG) accounting in computing. The study scope extends from data preparation to model validation, and integrates the full life cycle of the hardware infrastructure: manufacturing (including raw-material extraction), use (compute, temporary storage, system administration, cooling), and end-of-life. We report (i) an annual footprint of 417.5 tCO2eq for the Jean Zay H100 partition, split almost equally between manufacturing and operation; (ii) an effective intensity of 36.7 gCO2eq per H100 GPU-hour; (iii) a total training footprint of 21 tCO2eq for Lucie 7B (574 564 H100 GPU-hours), inclusive of amortised hardware manufacturing; (iv) on-site water consumption of approximately 76m3 for the training campaign and an annual Water Usage Effectiveness (WUE) of 0.07 L/kWh for IDRIS; (v) a heat-reuse factor (ERF) of 0.37 thanks to waste-heat recovery into the urban heating network. The study contributes one of the few publicly documented LCAs of an LLM training campaign that explicitly couples operational data with embodied emissions decomposed by subsystem (compute, storage, power chain, cooling), and discusses the implications for the design of frugal-by-construction AI systems in Europe.

Figures

Figures reproduced from arXiv: 2607.05408 by Jean-Pierre Lorr\'e, Marc L\'eobet, Pierre-Fran\c{c}ois Lavall\'ee.

Figure 1
Figure 1. Figure 1: Annual environmental footprint of the Jean Zay H100 partition [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 1 canonical work pages

  1. [1]

    Energy and policy considerations for deep learning in NLP

    Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations for deep learning in NLP. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pages 3645–3650, Florence, Italy, 2019. Association for Computational Linguistics. doi: 10.18653/v1/P19-1355

    doi:10.18653/v1/p19-1355 2019

  2. [2]

    Carbon emissions and large neural network training.arXiv preprint, arXiv:2104.10350, 2021

    David Patterson, Joseph Gonzalez, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, and Jeff Dean. Carbon emissions and large neural network training.arXiv preprint, arXiv:2104.10350, 2021. URLhttps://arxiv.org/abs/ 2104.10350

    Pith/arXiv arXiv 2021

  3. [3]

    The carbon footprint of machine learning training will plateau, then shrink.IEEE Computer, 55(7):18–28, 2022

    David Patterson, Joseph Gonzalez, Urs Hölzle, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, and Jeff Dean. The carbon footprint of machine learning training will plateau, then shrink.IEEE Computer, 55(7):18–28, 2022. doi: 10.1109/ MC.2022.3148714

    arXiv 2022

  4. [4]

    LLaMA: Open and 10 efficient foundation language models.arXiv preprint, arXiv:2302.13971, 2023

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and 10 efficient foundation language models.arXiv preprint, arXiv:2302.13971, 2023. URLhttps: //arxiv.org/abs/2302.13971

    Pith/arXiv arXiv 2023

  5. [5]

    Power hungry processing: Watts driving the cost of AI deployment? InProceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT)

    Alexandra Sasha Luccioni, Yacine Jernite, and Emma Strubell. Power hungry processing: Watts driving the cost of AI deployment? InProceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT). ACM, 2024. doi: 10.1145/3630106. 3658542

    doi:10.1145/3630106 2024

  6. [6]

    Islam, and Shaolei Ren

    Pengfei Li, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren. Making AI less “thirsty”: Uncovering and addressing the secret water footprint of AI models.arXiv preprint, arXiv:2304.03271, 2023. URLhttps://arxiv.org/abs/2304.03271

    Pith/arXiv arXiv 2023

  7. [7]

    Islam, and Shaolei Ren

    Pengfei Li, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren. Making AI less thirsty. Communications of the ACM, 2025. doi: 10.1145/3702322

    work page doi:10.1145/3702322 2025

  8. [8]

    Lee, Gu-Yeon Wei, David Brooks, and Carole-Jean Wu

    Udit Gupta, Young Geun Kim, Sylvia Lee, Jordan Tse, Hsien-Hsin S. Lee, Gu-Yeon Wei, David Brooks, and Carole-Jean Wu. Chasing carbon: The elusive environmental footprint of computing. In2022 IEEE International Symposium on High-Performance Computer Architecture (HPCA), pages 854–867. IEEE, 2022. doi: 10.1109/HPCA53966.2022.00069

    doi:10.1109/hpca53966.2022.00069 2022

  9. [9]

    Unraveling the hidden environmental impacts of AI solutions for environment: Life cycle assessment of AI solutions.Sustainability, 14(9):5172, 2022

    Anne-Laure Ligozat, Julien Lefèvre, Aurélie Bugeau, and Jacques Combaz. Unraveling the hidden environmental impacts of AI solutions for environment: Life cycle assessment of AI solutions.Sustainability, 14(9):5172, 2022. doi: 10.3390/su14095172

    doi:10.3390/su14095172 2022

  10. [10]

    Sustainable AI: Environmental implications, challenges and opportunities

    Carole-Jean Wu, Ramya Raghavendra, Udit Gupta, Bilge Acun, Newsha Ardalani, Kiwan Maeng, Gloria Chang, Fiona Wambari Aga, Jinshi Huang, Charles Bai, et al. Sustainable AI: Environmental implications, challenges and opportunities. InProceedings of Machine Learning and Systems (MLSys), volume 4, pages 795–813, 2022

    2022

  11. [11]

    LLMCarbon: Modeling the end-to-end carbon footprint of large language models

    Ahmad Faiz, Sotaro Kaneda, Ruhan Wang, Rita Osi, Prateek Sharma, Fan Chen, and Lei Jiang. LLMCarbon: Modeling the end-to-end carbon footprint of large language models. In Proceedings of the 12th International Conference on Learning Representations (ICLR), 2024. URLhttps://openreview.net/forum?id=aIok3ZD9to

    2024

  12. [12]

    AFNOR SPEC 2314: Référentiel général pour l’IA frugale — mesure et réduction de l’impact environnemental de l’IA

    Association française de normalisation (AFNOR). AFNOR SPEC 2314: Référentiel général pour l’IA frugale — mesure et réduction de l’impact environnemental de l’IA. Technical report, AFNOR, June 2024. URL https://www.afnor.org/presse/2024/06/ publication-du-referentiel-ia-frugale/

    2024

  13. [13]

    Estimation de l’empreinte carbone d’une heure de calcul sur un cœur CPU ou sur un GPU

    Labos 1point5. Estimation de l’empreinte carbone d’une heure de calcul sur un cœur CPU ou sur un GPU. Technical report, Labos 1point5, January 2024. URLhttps://labos1point5. org/les-rapports/estimation-empreinte-calcul

    2024

  14. [14]

    Estimating the carbon footprint of BLOOM, a 176B parameter language model.Journal of Machine Learning Research, 24(253):1–15, 2023

    Alexandra Sasha Luccioni, Sylvain Viguier, and Anne-Laure Ligozat. Estimating the carbon footprint of BLOOM, a 176B parameter language model.Journal of Machine Learning Research, 24(253):1–15, 2023. URLhttps://jmlr.org/papers/v24/23-0069.html

    2023

  15. [15]

    The Lucie-7B LLM and the Lucie training dataset: Open resources for multilingual language generation.arXiv preprint, arXiv:2503.12294, 2025

    OpenLLM-France Consortium. The Lucie-7B LLM and the Lucie training dataset: Open resources for multilingual language generation.arXiv preprint, arXiv:2503.12294, 2025. URL https://arxiv.org/abs/2503.12294. 11

    Pith/arXiv arXiv 2025

  16. [16]

    Étude de l’empreinte environnementale des centres nationaux de calcul intensif

    GENCI and Labos 1point5. Étude de l’empreinte environnementale des centres nationaux de calcul intensif. Technical report, GENCI / Labos 1point5, 2020

    2020

  17. [17]

    ISO 14040:2006 / ISO 14044:2006 — envi- ronmental management — life cycle assessment — principles and framework / requirements and guidelines, 2006

    International Organization for Standardization. ISO 14040:2006 / ISO 14044:2006 — envi- ronmental management — life cycle assessment — principles and framework / requirements and guidelines, 2006

    2006

  18. [18]

    The open source AI definition v1.0.https://opensource.org/ai/ open-source-ai-definition, 2024

    Open Source Initiative. The open source AI definition v1.0.https://opensource.org/ai/ open-source-ai-definition, 2024

    2024

  19. [19]

    Évaluation de l’empreinte environnementale du numérique en France et analyse prospective

    ARCEP and ADEME. Évaluation de l’empreinte environnementale du numérique en France et analyse prospective. Technical report, ARCEP / ADEME, January 2022. URL https://www.arcep.fr/uploads/tx_gspublication/ etude-empreinte-environnementale-numerique-france-note-methodologique_ janv2022.pdf

    2022

  20. [20]

    Regulation (EU) 2024/1689 of the European Parliament and of the council of 13 june 2024 (AI Act)

    European Parliament and Council. Regulation (EU) 2024/1689 of the European Parliament and of the council of 13 june 2024 (AI Act). Official Journal of the European Union, 12 July 2024, 2024

    2024

  21. [21]

    NVIDIA HGX H100 product carbon footprint summary.https:// images.nvidia.com/aem-dam/Solutions/documents/HGX-H100-PCF-Summary.pdf, 2025

    NVIDIA Corporation. NVIDIA HGX H100 product carbon footprint summary.https:// images.nvidia.com/aem-dam/Solutions/documents/HGX-H100-PCF-Summary.pdf, 2025

    2025

  22. [22]

    Commission delegated regulation (EU) 2021/2139 of 4 june 2021 supplementing regulation (EU) 2020/852 (EU Taxonomy)

    European Commission. Commission delegated regulation (EU) 2021/2139 of 4 june 2021 supplementing regulation (EU) 2020/852 (EU Taxonomy). Official Journal of the European Union, 2021

    2021

  23. [23]

    Bilan électrique 2024 — émissions

    Réseau de Transport d’Électricité (RTE). Bilan électrique 2024 — émissions. https: //analysesetdonnees.rte-france.com/bilan-electrique-2024/emissions, 2024

    2024

  24. [24]

    Enquête annuelle pour un numérique soutenable, édition 2025

    ARCEP. Enquête annuelle pour un numérique soutenable, édition 2025. Technical report, ARCEP, June 2025

    2025

  25. [25]

    Rapport extra-financier — intensité hydrique de la production électrique

    EDF. Rapport extra-financier — intensité hydrique de la production électrique. EDF Group Extra-Financial Report, 2024

    2024

  26. [26]

    Energy reuse effectiveness (ERE) and energy reuse factor (ERF)

    The Green Grid. Energy reuse effectiveness (ERE) and energy reuse factor (ERF). White Paper #29, 2010. URLhttps://www.thegreengrid.org

    2010

  27. [27]

    Commission delegated regulation (EU) 2024/1364 of 14 march 2024 on the first phase of the establishment of a common union rating scheme for data centres

    European Commission. Commission delegated regulation (EU) 2024/1364 of 14 march 2024 on the first phase of the establishment of a common union rating scheme for data centres. Official Journal of the European Union, 2024

    2024

  28. [28]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint, arXiv:2307.09288, 2023

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint, arXiv:2307.09288, 2023. URL https://arxiv.org/abs/2307.09288. 12

    Pith/arXiv arXiv 2023