Pith. sign in

REVIEW 3 major objections 5 minor 59 references

CEO-DC: Driving Decarbonization in HPC Data Centers with Actionable Insights

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Carbon-economy accounting of AI/HPC procurement shows platform improvements lag demand growth in 72% of cases, with four-year replacement cycles requiring about $512 per tonne of CO2-eq against a global average carbon price of $32.

desk verdict A genuinely useful framework paper whose headline numbers need sensitivity analysis before they should be treated as policy-grade; the 4% embodied-emissions trend is the main thing to scrutinize. read the letter →

arxiv 2507.08923 v2 pith:7YG6HHHT submitted 2025-07-11 cs.AR cs.CYcs.PF

Rubén Rodríguez Álvarez , Denisa-Andreea Constantinescu , Miguel Peón-Quirós , David Atienza This is my paper
classification cs.ARcs.CYcs.PF
keywords carbonefficiencypricedatacentersustainabilityAIhardwareprocurementembodiedemissionspricingdemandgrowthHPC
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 claims that data-center procurement decisions can be captured by two ratios, carbon efficiency (FLOP per tonne of CO2-equivalent) and price efficiency (FLOP per dollar), and that comparing those ratios across hardware generations exposes a systematic carbon-economy gap. Applying this accounting to the best cross-vendor AI training and HPC inference submissions from 2018 to 2024, the authors find that platform improvements lag compute-demand growth in 72% of benchmark-platform cases, and that only 28% of cases could absorb the projected baseline demand growth without increasing total emissions. The framework's headline policy numbers are that a four-year replacement cycle requires about 512 USD per tonne of CO2-eq in incentive, against a global average carbon price of 32 USD, and that in countries with medium-to-high carbon-intensity electricity, replacing four-year-old platforms cuts projected emissions by at least 75%. If these numbers are right, current carbon pricing and procurement practice are systematically misaligned with the sector's climate goals.

What carries the argument

The load-bearing machinery is a pair of efficiency ratios with a shared structure. Carbon Efficiency, $CE$, and Price Efficiency, $PE$, each combine an operational term, $CE_{OP} = \mathrm{Perf}/(\mathrm{Power} \times CI \times PUE)$ and $PE_{OP} = \mathrm{Perf}/(\mathrm{Power} \times EP \times PUE)$, with a capital term over the device lifetime, joined as a harmonic sum so that total efficiency is dominated by its weakest component. The carbon-economy gap is formalized by the incentive $w = (1/PE(d_B) - 1/PE(d_A))/(1/CE(d_A) - 1/CE(d_B))$, which converts the difference in price efficiency into the carbon price needed to make a sustainable upgrade also profitable. The companion identity $\eta_S \le CE(d_A)/CE_{OP}(d_B)$ states how much extra compute capacity an upgrade can buy without raising total emissions above the legacy system. Together these identities translate raw benchmark, price, and grid-carbon data into procurement thresholds, upgrade-cycle lengths, and required subsidies.

What would settle it

Use measured rack-level energy, not TDP, on the same training and HPC workloads for the 2025-2030 platform generations and refit the log-linear trends: the headline numbers would be falsified if the median doubling time for energy efficiency rises above roughly 23 months while compute-demand growth stays above 40% per year, or if measured energy makes operational emissions small enough that the required four-year incentive falls below existing carbon prices (e.g., below 167 USD per tonne of CO2-eq).

Watch

Extended reading notes

Core claim

CEO-DC is a decision framework built from six balances: OPEX savings versus CAPEX investment, sustainability versus economic viability, demand growth versus sustainability, operational versus capital carbon incentives, acceleration versus energy efficiency, and flexibility versus specialization. Its two central measures are Carbon Efficiency, $CE$ in FLOP per tonne of CO2-eq, and Price Efficiency, $PE$ in FLOP per dollar, each split into operational and capital components and combined through a harmonic sum. From these the paper derives the maximum sustainable growth $\eta_S \le CE(d_A)/CE_{OP}(d_B)$ under an iso-carbon policy, and the incentive $w = (1/PE(d_B) - 1/PE(d_A))/(1/CE(d_A) - 1/CE(d_B))$ that closes the carbon-economy gap. Applied to current AI/HPC trends, the framework reports that in 72% of benchmark-platform pairs the improvement rate does not justify four-year replacements on economic grounds, that only 28% of pairs can sustain the base projected capacity growth within an iso-carbon budget, and that the incentive needed for a four-year cycle, about 512 USD per tonne of CO2-eq, is an order of magnitude above the current global average carbon price. Regionally, the same model finds that medium-to-high carbon-intensity countries can cut projected emissions by 75% or more by replacing four-year-old platforms, while in very-low-carbon grids such as France and Switzerland, expansion under an iso-carbon cap is classified as non-feasible because embodied emissions alone exceed the target.

Load-bearing premise

The load-bearing premise is that the 2018-2024 exponential trends in platform performance, energy efficiency, and compute demand continue unchanged to 2035 and that thermal design power stands in for measured energy use, since bending any of those assumptions would materially change the 72%, 75%, and 512 USD per tonne of CO2-eq numbers.

Editorial extensions

If this is right

  • If the fitted trends hold, 72% of today's benchmark-platform pairs are not economically incentivized to replace hardware on a four-year cycle even though such replacements would lower emissions.
  • An iso-carbon cap makes sustainable growth depend on efficiency doubling roughly as fast as demand; with 40-66% annual demand growth, only 28% of current pairs can keep up.
  • At the current global average carbon price of 32 USD per tonne of CO2-eq, the roughly 512 USD per tonne needed for four-year cycles is out of reach, so pricing alone will not drive the transition.
  • In medium-to-high carbon-intensity grids, replacing platforms at four, five, and six years cuts projected emissions by roughly 75%, 80%, and 88%, making early replacement a near-term lever independent of grid decarbonization.
  • Meeting a 57% global emissions reduction by 2035 while preserving demand growth requires grid decarbonization of roughly 23-30% per year, and in very-low-carbon grids such as France and Switzerland, iso-carbon expansion is not feasible with current embodied-emission intensities.

Reading between the lines

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

  • The same ratio structure can be inverted: once a regulator fixes a carbon price, the model can read off the minimum energy-efficiency doubling time a region must hit to make a given upgrade cycle viable; the paper's annex tables already list such thresholds per country.
  • Because the paper uses thermal design power rather than measured energy, the headline figures are testable at the rack level; measuring actual power draw on the same workloads would either confirm the 72% and 75% numbers or move them.
  • The metric pair suggests a per-job or per-workload version for cloud scheduling, where heterogeneous jobs could be placed to maximize portfolio-level carbon and price efficiency; the paper only notes QoS-aware profit modeling as future work.
  • A policy experiment follows directly: in high-carbon grids, setting an operational carbon price near the model's required level should measurably shorten procurement cycles, and operators' observed upgrade behavior over the next few years would test that prediction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript proposes CEO-DC, a decision framework that couples embodied and operational carbon footprints with procurement economics for HPC/AI data centers. It introduces six 'balances' (Eqs. 1-9), carbon and price efficiency metrics, and applies them to MLPerf training/HPC results (2018-2024), country-level grid carbon intensities and electricity prices, and IEA/Top500 demand-growth scenarios. The headline quantitative claims are that platform improvements lag demand growth in 72% of benchmark cases, that replacing platforms older than four years could reduce emissions by at least 75% in high-carbon-intensity countries, and that aligning incentives to a four-year upgrade cycle would require roughly 512 USD/tCO2-eq, far above the observed global average carbon price of 32 USD/tCO2-eq. The framework's equations are explicit and the annex lists benchmark submission IDs and input tables for reproducibility.

Significance. If the empirical inputs withstand scrutiny, CEO-DC is a useful integrative tool for procurement planning, and the headline numbers are policy-relevant. The paper's strengths are its transparent derivations, the public input data and MLPerf submission IDs in the annex, the use of an external World Bank carbon-price benchmark rather than fitting conclusions to a target, and the explicit separation of economic viability from environmental sustainability. The framework also goes beyond carbon-only procurement studies by modeling demand growth and incentive mechanisms. However, the quantitative conclusions rest on several empirical extrapolations that are not yet supported by sensitivity analysis, most importantly the claimed 4%/year embodied-emissions growth rate, which is contradicted by the paper's own Table 4.

major comments (3)
  1. [Section 4.1 and Table 4] The claimed 4%/year growth in embodied emissions per platform is not supported by the data in Table 4. Using only the IC and memory columns, the V100 SXM2 16GB (2017) has 57.9 kgCO2-eq/device and the B200 SXM 192GB (2024) has 387 kgCO2-eq/device, a factor of 6.7 increase, i.e., roughly 31%/year. Even including the fixed 1780 kgCO2-eq server module amortized over four slots, plus assembly, transport, and cooling, gives approximately 600 kgCO2-eq/device for the V100 versus roughly 1000 kgCO2-eq/device for the B200, i.e., about 7-8%/year, not 4%/year. Since CCFP enters Eq. (1), the capital-emissions term of Eq. (6), the maximum sustainable growth in Eq. (8), and the incentive formula in Eq. (9), the Figure 5 savings percentages and the 512 USD/tCO2-eq four-year incentive cannot be considered trustworthy until the embodied-emissions growth rate is either derived explicitly from the annex data or replaced by a range with sensitivity analysis.
  2. [Section 4.2.2, Figure 5, and Section 5] The 2035 results are obtained by extrapolating log-linear fits to best-per-year MLPerf submissions (Tables 7-8) over 7-11 years with no uncertainty propagation and no plateau scenario. The text in Section 5 acknowledges that a significant slowdown in hardware improvements would require extending platform lifetimes, but that is exactly the load-bearing assumption: the design incentives in Table 3 and the decarbonization rates in Figure 5 would change materially if the months-to-double values in Tables 7-8 varied within their reported standard deviations or decelerated after 2024. Please provide confidence intervals or scenario bounds for the headline percentages and incentives, not only for the central trend fits.
  3. [Section 4.1 and Tables 7-8] Using TDP as a proxy for energy consumption can bias the energy-efficiency improvement trends because TDP is a design envelope rather than a workload-dependent power measure, and the bias can differ across vendors and generations. Since the 72% and 75% claims are ratios of energy-efficiency improvements, the authors should either use measured or power-capped MLPerf energy submissions, which the paper already mentions, or state and test the direction of the bias. This matters especially because the fastest and most energy-efficient submissions are selected separately; if those submissions were not run under comparable power conditions, the fitted efficiency trend is not controlled.
minor comments (5)
  1. [Section 4.2.1] The text says 'a capacity increase of 6.63×/year derived from the IEA Base Case,' but Table 3 defines the column as 'Compute demand growth in 4 years,' so the correct expression is '6.63× over the four-year cycle.'
  2. [Figure 5] The multiple numbers attached to each country (e.g., 'US -75% -24%') need an explicit legend distinguishing emission-savings percentages, decarbonization-rate targets, and incentive levels; the current caption leaves this to inference.
  3. [Section 4.1 and Table 4] The text and GPU price list use 'B200 SXM 180GB' while Table 4 and Figure 3 use 'B200 SXM 192GB'; the product name should be harmonized.
  4. [Section 4.1] The cooling-system factor of 0.0788 kgCO2-eq/W is attributed to 'Gupta et al. [30],' but reference [30] is the Li et al. LLM life-cycle paper; the attribution should be corrected to the appropriate prior work.
  5. [Table 2] The row 'CAPEX-dominated CE(A)=CECA(A)=0' is ambiguous because it is not clear whether the zero denotes negligible operational carbon or a limiting-case assignment; the notation should be defined.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CEO-DC's quantitative claims are computed from explicit definitions and external data inputs, not from self-referential or fitted-as-prediction logic.

full rationale

The paper's central results are derived by applying algebraic definitions (Eqs. 1-9) to externally sourced data: MLPerf benchmark submissions for performance and energy trends, IEA and Top500 reports for demand growth, Boavizta, IMEC NetZero, and vendor documentation for embodied emissions, and World Bank data for carbon prices. The headline claims (72% of benchmarks lag demand growth, 75% emissions savings from four-year replacements, and the 512 USD/tCO2 incentive) are scenario-level extrapolations from these inputs, not restatements of the model's definitions. The maximum sustainable growth in Eq. 8 and the incentive in Eq. 9 are algebraic consequences of the cost and carbon accounting, and they are then compared against independent external benchmarks such as World Bank carbon prices, so the conclusions are not forced by construction. There are no load-bearing self-citations: the cited prior work (ACT, Boavizta, Gupta et al., Schneider et al.) provides parameter values or modeling context rather than an unverified uniqueness theorem that forbids alternatives. The paper's own Section 5 explicitly acknowledges that 'a significant slowdown in hardware improvements would require extending platform lifetimes,' indicating that the trend assumptions are treated as empirical inputs, not as derived conclusions. The apparent inconsistency between the stated 4%/year embodied-emissions growth rate and the device-level values in Table 4 is a data-consistency or correctness concern, not a circularity: the rate is fitted as an input and used in the model, rather than being defined in terms of the model's output. Accordingly, no circular step can be exhibited, and the appropriate finding is no significant circularity with score 0.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central results rest on fitted improvement and pricing trends, an assumed PUE and utilization level, and external LCA datasets. The framework equations themselves are derived transparently, but the headline numbers (72%, 75%, 512 USD/tCO2) are outputs of these fitted inputs rather than independent measurements.

free parameters (7)
  • Per-benchmark performance improvement rate (months to double) = 15 ± 9 months mean (range 6-43 per benchmark, Tables 7-8)
    Fitted via log-linear regression to selected MLPerf submissions; used to compute upgrade thresholds and sustainable growth rates in Fig. 4 and Table 3.
  • Per-benchmark energy efficiency improvement rate (months to double) = 23 ± 16 months mean
    Same regression; determines operational carbon savings from upgrades.
  • GPU price growth rate = ~24%/year
    Exponential fit to four market release prices (V100 $9500, A100 $18500, H100 NVL $33000, B200 $44250); drives CAPEX cost and PE_CA.
  • Embodied emissions growth rate = ~4%/year
    Exponential fit to the same four platforms; used for CAPEX emissions. Appears inconsistent with Table 4 (V100 IC+memory 57.9 kg vs B200 387 kg).
  • DC compute demand growth rate (eta) = 55%, 60%, 66%/yr (IEA Headwinds/Base/Lift-off); 40%, 49%/yr (Top500 #500, #1)
    Translated from IEA energy projections using eta = eavg x Etrend; central driver of the 72% lag claim.
  • Power usage effectiveness (PUE) = 1.25
    Assumed constant across all DCs; multiplies all operational power in CE_OP and PE_OP.
  • Gross income value GB = H100 cloud price 2.65 USD/h
    Used in profit equation (Eq. 3) to model incentives and feasibility.
assumptions (6)
  • domain assumption Exponential trends in hardware performance, energy efficiency, prices, and emissions fitted from 2018-2024 continue to 2035.
    Invoked in Section 4.1 and acknowledged as a limitation in Section 5; all headline results depend on it.
  • domain assumption Thermal design power (TDP) approximates actual energy consumption.
    Section 4.1: 'In the absence of measured energy data, we use TDP as a proxy for energy consumption.' Biases CE_OP and PE_OP if utilization is below maximum.
  • domain assumption Dedicated HPC DCs with predictable, monolithic workloads and full node utilization.
    Section 3 and 4: framework excludes cloud and mobile domains; authors state this limitation in Section 5.
  • domain assumption LCA data from IMEC NetZero, Boavizta, and Lenovo accurately represent embodied emissions.
    Section 4.1: all CCFP values derive from these external tools, including conservative 1.5 kgCO2-eq/GB for DRAM.
  • domain assumption Iso-carbon policy (total emissions <= legacy operational emissions) is the sustainability goal.
    Section 3.1: defines GCFP; this choice drives the maximum sustainable growth formula (Eq. 8).
  • domain assumption Static country-level carbon intensity and electricity prices for 2025 with constant annual decarbonization rate.
    Section 4.2.2 and Annex Table 5; regional results depend on these inputs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CEO-DC: Driving Decarbonization in HPC Data Centers with Actionable Insights." pith.science (2026). https://pith.science/paper/7YG6HHHT

@misc{pith2026250708923,
  author       = {Pith},
  title        = {Pith review of: CEO-DC: Driving Decarbonization in HPC Data Centers with Actionable Insights},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7YG6HHHT}},
  note         = {Machine review of arXiv:2507.08923}
}
read the original abstract

The rapid growth of data centers is increasing energy demand and widening the carbon gap in the ICT sector, as fossil fuels still dominate global energy production. Addressing this challenge requires collaboration across research, policy, and industry to rethink how computing infrastructures are designed and scaled sustainably. This work addresses central trade-offs in procurement decisions that affect carbon emissions, economic costs, and scaling of compute resources. We present these factors in a holistic decision-making framework for Carbon and Economy Optimization in Data Centers (CEO-DC). CEO-DC introduces new carbon and price metrics that enable DC managers, platform designers, and policymakers to make informed decisions. Applying CEO-DC to current trends in AI and HPC reveals that, in 72% of the cases, platform improvements lag behind demand growth. Moreover, prioritizing energy efficiency over latency can reduce the economic appeal of sustainable designs. Our analysis shows that in many countries with electricity with medium to high carbon intensity, replacing platforms older than four years could reduce their projected emissions by at least 75%. However, current carbon incentives worldwide remain insufficient to steer data center procurement strategies toward sustainable goals. In summary, our findings underscore the need for a shift in hardware design and faster grid decarbonization to ensure sustainability and technological viability.

Figures

Figures reproduced from arXiv: 2507.08923 by the authors.

Figure 1
Figure 1. Key stakeholders for sustainable DC scaling. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Carbon-economy model. Key trade-offs are [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Embodied emissions for HPC platforms. 4.1 Methodology for Quantifying Emissions and Economic Factors We use the MLPerf AI Training and HPC Inference Benchmarks [34, 15] to analyze performance and energy consumption trends across hardware generations. These benchmarks provide standardized, cross-vendor metrics that are well-suited for evaluating carbon-aware upgrade strategies in DCs focused on AI. Our analysis inclu… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: AI platform trends from an Nvidia H100. We apply the carbon-economy model to address the trade-offs of the DC decision space. To this end, we plot the sustainable and viable upgrade thresholds from Eq. 6— green and red lines, respectively. All benchmarks fall above the…
Figure 5
Figure 5. Figure 5: Upgrade policies and decarbonization needs by country for current platform improvement trends and [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Minimum upgrade incentive and maximum sus [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 45 canonical work pages

  1. [1]

    Laksono Adhianto, Sinchan Banerjee, Mike Fagan, Mark Krentel, Gabriel Marin, John Mellor-Crummey, and Nathan R Tallent. 2010. HPCToolkit: Tools for performance analysis of optimized parallel pro- grams. Concurrency and Computation: Practice and Experience 22, 6 (2010), 685–701

    work page 2010

  2. [2]

    Advanced Micro Devices, Inc. 2024. AMD 2023-24 Corporate Responsibility Report . Technical Report. Advanced Micro Devices, Inc. https://www.amd.com/content/dam/ amd/en/documents/corporate/cr/corporate- responsibility-report.pdf

    work page 2024

  3. [3]

    International Energy Agency. 2024. Electricity 2024 Executive Summary. https://www.iea.org/ reports/electricity-2024/executive-summary

    work page 2024

  4. [4]

    Tuğana Aslan, Peter Holzapfel, Lutz Stobbe, Andreas Grimm, Nils F Nissen, and Matthias Finkbeiner. 2025. Toward climate neutral data centers: Greenhouse gas inventory, scenarios, and strategies.iScience 28, 1 (2025)

    work page 2025

  5. [5]

    Boavizta Working Group. 2025. Datavizta: En- vironmental footprint data visualisation frontend. https://datavizta.boavizta.org/

    work page 2025

  6. [6]

    Central Statistics Office, Ireland. 2023. Data Centres Metered Electricity Consumption 2022 - Key Findings. https://www.cso.ie/ en/releasesandpublications/ep/p-dcmec/ datacentresmeteredelectricityconsumption2022/ keyfindings/. Accessed: 2025-04-01

    work page 2023

  7. [7]

    Zeming Chen, Alejandro Hernández Cano, Angelika Romanou, Antoine Bonnet, Kyle Matoba, Francesco Salvi, Matteo Pagliardini, Simin Fan, Andreas Köpf, Amirkeivan Mohtashami, et al. 2023. Meditron-70b: Scaling medical pretraining for large language models. arXiv preprint arXiv:2311.16079(2023)

    arXiv 2023

  8. [8]

    CINECA. 2025. CINECA High Performance Comput- ing. https://www.cineca.it. Accessed: 2025-07-03

    work page 2025

Show all 59 references
  1. [9]

    Microsoft Corporation. 2024. Microsoft 2024 Envi- ronmental Sustainability Report.https://aka.ms/ SustainabilityReport2024. Accessed: 2025-05-01

    2024

  2. [10]

    Lieven Eeckhout. 2025. The sustainability gap for computing: Quo vadis?Commun. ACM68, 3 (2025), 70–79

    2025

  3. [11]

    Carloni, Asaf Cidon, Hubertus Franke, Martha A

    Tamar Eilam, Pradip Bose, Luca P. Carloni, Asaf Cidon, Hubertus Franke, Martha A. Kim, Eun K. Lee, Mahmoud Naghshineh, Pritish Parida, Clifford S. Stein, and Asser N. Tantawi. 2024. Reducing Data- center Compute Carbon Footprint by Harnessing the Power of Specialization: Princ...

    2024

  4. [12]

    Electricity Maps ApS. 2025. Electricity Maps – Real-time electricity carbon intensity.https://app. electricitymaps.com/map/72h/hourly. Accessed: 2025-04-01

    2025

  5. [13]

    Yearly Electricity Data Europe

    Ember, Energy Institute, and Our World in Data. 2025. Carbon Intensity of Electricity Generation – Ember [dataset]. Our World in Data. https://archive.ourworldindata.org/ 20250717-120101/grapher/carbon-intensity- electricity.html Based on Ember, "Yearly Electricity Data Europe...

    2025

  6. [14]

    Ahmad Faiz, Sotaro Kaneda, Ruhan Wang, Rita Osi, Prateek Sharma, Fan Chen, and Lei Jiang

  7. [15]

    Steven Farrell, Murali Emani, Jacob Balma, Lukas Drescher, Aleksandr Drozd, Andreas Fink, Geoffrey Fox, David Kanter, Thorsten Kurth, Peter Mattson, et al. 2021. MLPerf HPC: A holistic benchmark suite for scientific machine learning on HPC systems. In 2021 IEEE/ACM Workshop on...

    2021

  8. [16]

    Kairos Fellowship. 2025. Google’s Eco-Failures: A Critical Analysis. Technical Report. No Climate Re- sults Found

    2025

  9. [17]

    Enea Figini and Mario Paolone. 2025. Achieving Dis- patchabilityinDataCenters: CarbonandCost-Aware Sizing of Energy Storage and Local Photovoltaic Gen- eration. Sustainable Energy, Grids and Networks (2025), 101920

    2025

  10. [18]

    Matthias Finkbeiner, Atsushi Inaba, Reginald Tan, Kim Christiansen, and Hans-Jürgen Klüppel. 2006. The new international standards for life cycle assess- ment: ISO 14040 and ISO 14044.The international journal of life cycle assessment11 (2006), 80–85

    2006

  11. [19]

    Blair, and Adrian Friday

    Charlotte Freitag, Mike Berners-Lee, Kelly Widdicks, Bran Knowles, Gordon S. Blair, and Adrian Friday

  12. [20]

    Anshul Gandhi, Dongyoon Lee, Zhenhua Liu, Shuai Mu, Erez Zadok, Kanad Ghose, Kartik Gopalan, Yu David Liu, Syed Rafiul Hussain, and Patrick Mc- Daniel. 2023. Metrics for Sustainability in Data Cen- ters. SIGENERGY Energy Inform. Rev.3, 3 (Oct. 2023), 40–46. doi:10.1145/3630614.3630622

    2023

  13. [21]

    Lee, David Brooks, and Carole- Jean Wu

    Udit Gupta, Mariam Elgamal, Gage Hills, Gu-Yeon Wei, Hsien-Hsin S. Lee, David Brooks, and Carole- Jean Wu. 2023. Architectural CO2 Footprint Tool: Designing Sustainable Computer Systems With an Architectural Carbon Modeling Tool.IEEE Micro (2023). doi:10.1109/mm.2023.3275139

    2023

  14. [22]

    Udit Gupta, Young Geun Kim, Sylvia Lee, Jor- dan Tse, Hsien-Hsin S Lee, Gu-Yeon Wei, David Brooks, and Carole-Jean Wu. 2021. Chasing carbon: The elusive environmental footprint of computing. In 2021 IEEE International Symposium on High- Performance Computer Architecture (HPCA)...

    2021

  15. [23]

    imec. 2023. NetZero: Towards Semiconductor Sus- tainability. https://netzero.imec-int.com/. Ac- cessed: 2025-05-01

    2023

  16. [24]

    Intel Corporation. 2024. Intel 2023-24 Corpo- rate Responsibility Report . Technical Report. Intel Corporation. https://csrreportbuilder. intel.com/pdfbuilder/pdfs/CSR-2023-24-Full- Report.pdf

    2024

  17. [25]

    Shixin Ji, Zhuoping Yang, Xingzhen Chen, Stephen Cahoon, Jingtong Hu, Yiyu Shi, Alex K Jones, and Peipei Zhou. 2024. SCARIF: Towards Carbon Mod- eling of Cloud Servers with Accelerators. In2024 IEEE Computer Society Annual Symposium on VLSI (ISVLSI). IEEE, 496–501

    2024

  18. [26]

    Nicola Jones et al. 2018. How to stop data cen- tres from gobbling up the world’s electricity.Na- ture 561, 7722 (2018), 163–166. doi:10.1038/d41586- 018-06610-y

    2018 doi

  19. [27]

    Jülich Supercomputing Centre. 2025. Jülich Super- computing Centre (JSC).https://www.fz-juelich. de/ias/jsc/. Accessed: 2025-07-03

    2025

  20. [28]

    Loïc Lannelongue, Jason Grealey, and Michael Inouye

  21. [29]

    Lenovo. 2025. Lenovo Product Carbon Footprint and Eco Declarations. https://lenovopress.lenovo. com/lp1855-lenovo-product-carbon-footprint- and-eco-declarations. Accessed: 2025-07-03

    2025

  22. [30]

    Yueying Lisa Li, Omer Graif, and Udit Gupta. 2024. Towards Carbon-efficient LLM Life Cycle. InProceed- ings of the 3rd Workshop on Sustainable Computer Systems (HotCarbon)

    2024

  23. [31]

    doi:10.1002/advs.202100707

    Green Algorithms: Quantifying the Carbon Footprint of Computation.Advanced Science(2021). doi:10.1002/advs.202100707

    2021 doi

  24. [32]

    Matteo Manganelli, Alessandro Soldati, Luigi Mar- tirano, and Seeram Ramakrishna. 2021. Strategies for improving the sustainability of data centers via energy mix, energy conservation, and circular energy. Sustainability (2021). doi:10.3390/su13116114

    2021 doi

  25. [33]

    Eric Masanet, Arman Shehabi, Nuoa Lei, Sarah Smith, and Jonathan Koomey. 2020. Recalibrat- ing global data center energy-use estimates.Science 367, 6481 (2020), 984–986. doi: 10.1126/science. aba3758

    2020 doi

  26. [34]

    Alexandra Sasha Luccioni, Emma Strubell, and Kate Crawford. 2025. From Efficiency Gains to Rebound Effects: The Problem of Jevons’ Paradox in AI’s Polarized Environmental Debate. arXiv preprint arXiv:2501.16548 (2025)

    2025 arXiv

  27. [35]

    MIT PAIA. 2025. Product Attributes to Impact Al- gorithm. https://paia.mit.edu/ Accessed: 2025- 06-10

    2025

  28. [36]

    NERSC. 2025. National Energy Research Scientific Computing Center (NERSC).https://www.nersc. gov. Accessed: 2025-07-03

    2025

  29. [37]

    MLPerfTrainingBenchmark

    Peter Mattson, Christine Cheng, Gregory Diamos, Cody Coleman, Paulius Micikevicius, David Patter- son, Hanlin Tang, Gu-Yeon Wei, Peter Bailis, Victor Bittorf, etal.2020. MLPerfTrainingBenchmark. Pro- ceedings of Machine Learning and Systems2 (2020), 336–349

    2020

  30. [38]

    Sirui Qi, Dejan Milojicic, Cullen Bash, and Sudeep Pasricha. 2023. MOSAIC: A Multi-Objective Opti- mization Framework for Sustainable Datacenter Man- agement. In2023 IEEE 30th International Confer- ence on High Performance Computing, Data, and Analytics (HiPC). 51–60. doi:10.1...

    2023

  31. [39]

    Ana Radovanovic, Bokan Chen, Saurav Talukdar, Binz Roy, Alexandre Duarte, and Mahya Shah- bazi. 2021. Power modeling for effective datacenter planning and compute management. IEEE Trans- actions on Smart Grid 13, 2 (2021), 1611–1621. doi:10.1109/tsg.2021.3125275

    2021

  32. [40]

    David Patterson, Joseph Gonzalez, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, and Jeff Dean. 2021. Carbon 13 emissions and large neural network training.arXiv preprint arXiv:2104.10350 (2021)

    2021 arXiv

  33. [41]

    Victor Schmidt, Kamal Goyal, Aditya Joshi, Boris Feld, Liam Conell, Nikolas Laskaris, Doug Blank, Jonathan Wilson, Sorelle Friedler, and Sasha Luc- cioni. 2021. CodeCarbon: estimate and track car- bon emissions from machine learning computing. https://codecarbon.io/

    2021

  34. [42]

    Ian Schneider, Hui Xu, Stephan Benecke, David Pat- terson, Keguo Huang, Parthasarathy Ranganathan, and Cooper Elsworth. 2025. Life-Cycle Emissions of AI Hardware: A Cradle-To-Grave Approach and Gen- erational Trends. arXiv preprint arXiv:2502.01671 (2025)

    2025 arXiv

  35. [43]

    Katherine Richardson, Will Steffen, Wolfgang Lucht, Jørgen Bendtsen, Sarah E Cornell, Jonathan F Donges, Markus Drüke, Ingo Fetzer, Govindasamy Bala, Werner Von Bloh, et al. 2023. Earth beyond six of nine planetary boundaries.Science advances9, 37 (2023), eadh2458

    2023

  36. [44]

    Chetan Choppali Sudarshan, Aman Arora, and V. A. Chhabria. 2023. GreenFPGA: Evaluating FPGAs as Environmentally Sustainable Computing Solutions. Design Automation Conference (2023). doi:10.48550/arxiv.2311.12396

    2023 doi arXiv

  37. [45]

    Chetan Choppali Sudarshan, Nikhil Matkar, Sarma Vrudhula, Sachin S Sapatnekar, and Vidya A Chhabria. 2024. ECO-CHIP: Estimation of carbon footprint of chiplet-based architectures for sustain- ableVLSI.In 2024 IEEE International Symposium on High-Performance Computer Architectu...

    2024

  38. [46]

    Statista. 2024. Leading countries by number of data centers (as of March 2024). Accessed March 2024

    2024

  39. [47]

    Taiwan Semiconductor Manufacturing Company

  40. [48]

    TechPowerUp. 2025. TechPowerUp GPU Database. http://www.techpowerup.com/gpu-specs. Ac- cessed: 2025-07-03

    2025

  41. [49]

    Swiss National Supercomputing Centre (CSCS). 2025. Alps: a general-purpose compute and data research infrastructure. https://www.cscs.ch/computers/ alps. Accessed: 2025-07-08

    2025

  42. [50]

    TOP500.org. 2025. TOP500 Statistics Performance Development 1993-2025. https://top500.org/ statistics/perfdevel/. Accessed: 2025-07-09

    2025

  43. [51]

    United Nations Environment Programme (UNEP)

  44. [52]

    World Bank. 2025. Carbon Pricing Dashboard – Com- pliance Prices. https://carbonpricingdashboard. worldbank.org/compliance/price Accessed: 2025- 04-14

    2025

  45. [53]

    Bill Tomlinson, Rebecca W Black, Donald J Patter- son, and Andrew W Torrance. 2024. The carbon emissions of writing and illustrating are lower for AI than for humans. Scientific Reports 14, 1 (2024), 3732

    2024

  46. [54]

    Shiqiang Zhu, Ting Yu, Tao Xu, Hongyang Chen, Schahram Dustdar, Sylvain Gigan, Deniz Gunduz, Ekram Hossain, Yaochu Jin, Feng Lin, et al. 2023. Intelligent computing: the latest advances, challenges, and future. Intelligent Computing 2 (2023), 0006. ANNEX: Data for Reproducibil...

    2023

  47. [56]

    please! UN Environment Pro- gramme

    Emissions Gap Report 2024: No more hot air... please! UN Environment Pro- gramme. https://www.unep.org/resources/ emissions-gap-report-2024 15th edition

    2024

  48. [58]

    Carole-Jean Wu, Bilge Acun, Ramya Raghavendra, and Kim Hazelwood. 2024. Beyond efficiency: Scaling AI sustainably.IEEE Micro(2024)

    2024

  49. [2021]

    Patterns (2021)

    The real climate and transformative impact of ICT: A critique of estimates, trends, and regula- tions. Patterns (2021). doi:10.1016/j.patter.2021. 100340

    2021 doi

  50. [2023]

    Techni- cal Report

    TSMC 2023 Sustainability Report. Techni- cal Report. Taiwan Semiconductor Manufacturing Company. https://esg.tsmc.com/file/public/ e-all_2023.pdf

    2023

  51. [2024]

    arXiv:2309.14393 [cs.CL] https://arxiv.org/abs/ 2309.14393

    LLMCarbon: Modeling the end-to-end Carbon Footprint of Large Language Models. arXiv:2309.14393 [cs.CL] https://arxiv.org/abs/ 2309.14393

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.