Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

The paper claims that a 100-MW GPU data center participating in 2-second frequency regulation can reduce more grid-side carbon—via 'exogenous carbon' savings—than the operational carbon it adds by running and modulating best-effort workload

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 · deepseek-v4-flash

2026-08-03 06:32 UTC pith:NZIZM4BK

load-bearing objection Strong GPU regulation testbed and a useful new carbon metric, but the headline savings rest on an unreported MCE_resv from a same-group unit commitment model. the 3 major comments →

arxiv 2601.22487 v3 pith:NZIZM4BK submitted 2026-01-30 cs.DC

Coordinating GPU Data Centers and Power Grid Regulation Service for Exogenous Carbon Benefits

Ali Jahanshahi , Sara Rashidi Golrouye , Osten Anderson , Nanpeng Yu , Daniel Wong This is my paper
classification cs.DC
keywords frequency regulationdemand responsecarbon accountingGPU data centersexogenous carbonunit commitmentcarbon intensitygrid emissions
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.

This paper argues that GPU data centers can act as a carbon-reducing grid asset by providing 2-second frequency regulation. It introduces Exogenous Carbon, a metric that credits data centers for grid-side emissions avoided when their flexible power consumption displaces fossil-fueled regulation reserves. It also presents EcoCenter, a framework that coordinates GPU power caps, core allocation, and multi-GPU pause/resume to maximize regulation capacity. The paper reports that at 10–80% load, grid-side exogenous savings (34.2–433.2 kt CO2e/yr) outweigh the operational emissions increase (21.3–33 kt CO2e/yr) in every scenario, making the data center net carbon-negative at low loads.

Core claim

The paper's central claim is that a modern GPU data center, co-locating latency-critical and best-effort workloads, can provide frequency regulation at 2-second granularity with a performance score above 80% (average 91.2%), and that the resulting grid-side 'exogenous carbon' savings—up to 433.2 kt CO2e/year under a grid model at 50% load—outweigh the extra operational emissions from running and modulating the best-effort workloads. By quantifying the previously hidden emissions of fossil-fueled regulation reserves, the paper argues that data center regulation service reduces both the amount of fossil reserves dispatched and the efficiency penalty those plants incur, so the carbon benefit is

What carries the argument

Exogenous Carbon (C_exogenous = R_DC × MCE_resv) is the central metric: it credits a data center with the grid-side emissions avoided when its regulation capacity R_DC displaces fossil-fueled regulation reserves. MCE_resv, the marginal carbon emission of regulation reserves, is obtained by running a unit commitment model of the CAISO/Western Interconnection grid with and without data center regulation. EcoCenter is the accompanying framework that maximizes R_DC by coordinating GPU power capping, GPU core (CU) allocation, and multi-GPU pause/resume of best-effort workloads to follow the 2-second regulation signal.

Load-bearing premise

The whole carbon ledger rests on a simulated number—the marginal carbon emission of regulation reserves—that the paper does not measure or validate against real grid emissions; if that simulation overstates how much fossil backup is displaced, the claimed net savings disappear.

What would settle it

Meter the actual carbon emissions of a grid region with and without a data center providing frequency regulation for a month, and compare the measured difference to Eq. 1's prediction; if measured grid-side savings fall below the operational carbon increase at 50% load, the central claim fails.

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

If this is right

  • At low-to-medium utilization, a GPU data center can achieve net negative carbon emissions, because grid-side exogenous savings exceed both operational and embodied carbon.
  • Existing carbon-intensity metrics undercount the true emissions impact of data centers; adding the exogenous-carbon term changes which workloads and locations appear green.
  • GPU-based regulation provides roughly 3x more regulation capacity than CPU-based approaches, making modern AI/ML data centers the strongest candidates for this service.
  • Grid operators can use data-center regulation to reduce fossil-fueled reserve requirements, supporting higher renewable penetration without building new backup plants.
  • Even at 80% load, the paper reports that grid-side savings (34.2 kt CO2e/yr) outweigh the operational increase (21.3 kt CO2e/yr), yielding a net improvement of about 12 kt CO2e/yr.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: If MCE_resv is validated against real ISO emissions, the exogenous-carbon ledger could become a standard addition to data-center carbon accounting, changing procurement and workload-scheduling decisions.
  • Inference: The net-benefit result is strongest in the near term; as grids decarbonize and fossil regulation reserves become scarcer, the exogenous savings shrink, so the claim is time-sensitive.
  • Inference: The TCO model implies a practical dispatch rule—modulate low-cost, low-GPU-hour-price resources first—and a natural extension is joint optimization of carbon benefit and opportunity cost.
  • Inference: A testable extension is to apply the same exogenous-carbon metric to other large flexible loads, such as electrolyzers or EV charging fleets, to see whether the carbon benefit generalizes beyond GPU data centers.

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

3 major / 4 minor

Summary. The paper proposes a metric called Exogenous Carbon to quantify grid-side CO2 reductions when a GPU data center provides 2-second frequency regulation service, and an accompanying framework, EcoCenter, that modulates GPU power caps, core allocation, and multi-GPU coordination to maximize regulation provision. The authors evaluate EcoCenter on an 8-GPU AMD server with Facebook SWIM load traces and three regulation signals, reporting average performance scores above 91% and regulation provision 3-4x larger than CPU-only approaches. They then combine measured regulation provision with a marginal-carbon-emission factor for regulation reserves (MCE_resv) obtained from a unit commitment model of the CAISO/Western Interconnection grid, and claim that grid-side exogenous savings outweigh the operational carbon increase of running best-effort workloads in all scenarios, sometimes by a large margin.

Significance. If validated, the result would be significant: it identifies a new mechanism by which data centers with flexible GPU workloads can act as net carbon-reducing grid assets, and it quantifies hidden emissions from frequency regulation reserves that are absent from standard carbon-intensity metrics. The experimental work is a concrete strength: performance scores are directly measured at 2-second granularity, the LC-only baseline is explicitly described as conservative, and the comparison against CPU-only and UPS-only regulation services is useful. However, the headline carbon-savings claim rests almost entirely on MCE_resv, which is not measured, not reported, and not validated against empirical grid emissions. Since the margin at 80% load is only about 1.6x, the central conclusion is currently unsupported without additional evidence on MCE_resv.

major comments (3)
  1. [§3.1.2, Eq. (1)] MCE_resv is the single load-bearing multiplier in the Exogenous Carbon ledger, yet no numerical value for MCE_resv is reported anywhere in the paper, nor is a sensitivity analysis or validation against observed CAISO emissions provided. The model used to compute it is from the authors' own prior work [50]. At 80% load (Fig. 11D), the grid-model savings are 34.2 kt CO2eq vs. an operational increase of 21.3 kt CO2eq — a margin of only 1.6x — so an overestimate of MCE_resv by more than ~40% flips the “all scenarios” claim. The authors must report MCE_resv values (per scenario or per run), compare model outputs to historical ISO emission data, and provide sensitivity bounds.
  2. [§3.1.2] The unit commitment model's representation of regulation service is not described in enough detail to assess credibility. The paper simply cites Anderson et al. [50] and asserts a 770 MW regulation requirement (5.5% of 14 GW) without justification or sensitivity. Critical modeling choices — how 2-second AGC signals are represented in an hourly unit commitment, how part-load efficiency penalties are applied, and whether unit-commitment changes from data-center regulation are captured — are all absent. Without this detail, a reader cannot judge whether the grid-model MCE_resv is realistic or an artifact of the modeling assumptions.
  3. [§3.1.1, Eq. (1)] The exogenous-carbon formula C_exogenous = R_DC × MCE_resv assumes a constant linear multiplier, but MCE_resv is defined as a marginal emission and is computed in §3.1.2 as a difference between two unit commitment runs with and without data-center regulation. For a 100-MW data center providing regulation, R_DC is not infinitesimal; the marginal interpretation may not hold. The authors should either justify linearity, or compute C_exogenous directly from the UC model difference rather than through a single linearized coefficient. This is especially important because the 3x gap between grid-model savings and simple gas/battery bounds (433.2 vs. 132 kt at 50% load) is attributed entirely to H2/H3 effects that have not been independently verified.
minor comments (4)
  1. [§5.2, Table 2] The text states that at higher utilization traces performance scores observe “less than 80%”; however Table 2 reports scores of 84–86% for high-load traces, so the statement contradicts the table. Also, the claim that the Noisy signal yields the lowest score is not consistently supported by Table 2 (e.g., med-med Noisy 93.38 vs. HT 93.34).
  2. [§5.1.1 vs. §5.2 (footnote 4)] The baseline definition is inconsistent. §5.1.1 says the Baseline co-locates LC and BE workloads without regulation, while §5.2 and footnote 4 say the Baseline is LC-only. Clarify which baseline is used for the carbon computations; the current presentation makes it hard to verify the “conservative” claim.
  3. [Figure 11] Figure 11 uses a log scale but the axis labels and the exact values for each bar are difficult to read, especially for the small bars. A table of the numerical carbon components (operational, embodied, simple gas/battery, grid-model exogenous) would improve replicability and make the 34.2 vs. 21.3 comparison easier to verify.
  4. [Abstract and §5.2] Minor language issues: “oftentimes” in the abstract, “completely outweighed” in §5.2, and inconsistent hyphenation of “exogenous carbon.” These do not affect the technical content.

Circularity Check

1 steps flagged

Self-cited UC model provides the unvalidated MCE_resv multiplier; grid-model savings are the model's own output, though EcoCenter's measured R_DC is independent.

specific steps
  1. self citation load bearing [Section 3.1.2; Eq. (1); Section 5.2 / Fig. 11]
    "We formulate and solve the unit commitment problem as laid out in Anderson [50] ... The difference of both values provides the 'hidden' grid-side carbon emission impact of data center frequency regulation, which we use to obtain the grid-detailed marginal carbon emission, MCEresv, for use in Eq. 1."

    MCE_resv is the sole grid-side multiplier in Eq. (1): C_exogenous = R_DC x MCE_resv. The paper neither reports a numerical MCE_resv nor validates it against observed ISO emissions; instead it is taken from a unit-commitment model in Anderson et al. [50], whose co-authors (Anderson, Yu) are also co-authors of this paper. The model is run 'with' and 'without' data center regulation, and the difference is used to define MCE_resv. Consequently, the 'grid-model' Exogenous Carbon numbers in Fig. 11 (e.g., 433.2 kt vs 132 kt simple-gas at 50% load) are the self-cited model's own predicted emission difference, scaled by measured R_DC; the 3x amplification and the 'in all scenarios... completely outweighs' conclusion are outputs of that model, not independent empirical confirmations. The measured R

full rationale

The EcoCenter contribution is largely independent: R_DC, performance scores, throughput, and operational-carbon deltas are measured on real hardware. The circularity is confined to the grid-side 'Exogenous Carbon' numbers that carry the headline claim. Eq. (1) defines Exogenous Carbon as R_DC times MCE_resv, and Section 3.1.2 obtains MCE_resv from the difference of two runs of a unit-commitment model cited from the same research group [50]. The grid-model savings presented in Section 5.2 (e.g., 433.2 kt at 50% load) are therefore, by construction, the self-cited model's own output. No sensitivity analysis or ISO-emissions validation is provided, so the margin by which savings 'completely outweigh' operational increases (e.g., 34.2 kt vs 21.3 kt at 80% load) is an artifact of that model rather than an independent finding. This is a load-bearing self-citation, but because the paper also contains independent measurements, the circularity is partial rather than total.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 1 invented entities

The central numbers rest on a handful of model inputs: the marginal regulation-reserve emission rate MCE_resv, the assumed 770 MW regulation requirement, and linearity of Eq. (1). None is independently measured in this paper.

free parameters (2)
  • MCE_resv (marginal carbon emission of regulation reserves) = not reported
    Multiplies R_DC in Eq. (1) to produce every Exogenous Carbon number; the paper does not tabulate its value or provide independent validation, and the grid-model estimate is ~3x the simple gas/battery estimate.
  • Regulation requirement of 770 MW (5.5% of 14 GW) = 770 MW
    Assumed fixed regulation capacity in the unit commitment model; if actual CAISO regulation procurement differs, exogenous savings scale accordingly.
axioms (5)
  • domain assumption Frequency-regulation reserves are predominantly fossil-fueled and their emissions are omitted from grid-reported carbon intensity.
    Foundational to Exogenous Carbon; supported by CAISO docs [32,33], but for grids with hydro/battery reserves the benefit disappears.
  • domain assumption Unit commitment model of Anderson et al. [50] accurately replicates CAISO/Western Interconnection dispatch and emissions with/without data-center regulation.
    Used to obtain MCE_resv in §3.1.2; no code, data, or sensitivity analysis is provided in this paper.
  • domain assumption Marginal carbon emission of regulation reserves is constant, so C_exo = R_DC x MCE_resv (Eq. 1) holds linearly over the data center's operating range.
    The unit commitment model is discrete/nonlinear; the metric linearizes the displacement, so savings at different R_DC may not scale linearly.
  • domain assumption Predicted hourly average load P_avg and variance P_var are accurate enough that the optimizer's feasible (R_up, R_down) bounds are respected.
    EcoCenter Optimizer §4.6 relies on predictions; paper cites predictors [54-60] but does not evaluate prediction error.
  • domain assumption 2-second performance score measured on 8xAMD MI50 with Resnet152+GPT2 generalizes to other GPUs and workloads.
    §5.1.1 admits Albert/hipBLAS similar trends but omits details; NVIDIA GPUs and newer architectures are not demonstrated.
invented entities (1)
  • Exogenous Carbon (C_exogenous) no independent evidence
    purpose: New carbon-accounting term that credits data centers for grid-side emissions reduction from regulation service; Eq. 1.
    Its value depends on MCE_resv from the authors' own unit commitment model; no direct measurement of regulation-reserve emissions or market settlement validates it. Falsifiable only if ISO emissions data are released.

pith-pipeline@v1.3.0-alltime-deepseek · 22012 in / 16683 out tokens · 178407 ms · 2026-08-03T06:32:22.762519+00:00 · methodology

0 comments
read the original abstract

The rapid growth of AI/ML data centers has led to higher energy consumption and carbon emissions. The shift to renewable energy and growing data center energy demands can destabilize the power grid. Power grids rely on frequency regulation reserves, typically fossil-fueled power plants, to stabilize and balance the supply and demand of electricity. This paper sheds light on the hidden carbon emissions of frequency regulation service. Our work explores how modern GPU data centers can coordinate with power grids to reduce the need for fossil-fueled frequency regulation reserves. We first introduce a novel metric, Exogenous Carbon, to quantify grid-side carbon emission reductions resulting from data center participation in regulation service. We additionally introduce EcoCenter, a framework to maximize the amount of frequency regulation provision that GPU data centers can provide, and thus, reduce the amount of frequency regulation reserves necessary. We demonstrate that data center participation in frequency regulation can result in Exogenous carbon savings that can outweigh operational carbon emissions

Figures

Figures reproduced from arXiv: 2601.22487 by Ali Jahanshahi, Daniel Wong, Nanpeng Yu, Osten Anderson, Sara Rashidi Golrouye.

Figure 1
Figure 1. Figure 1: The growth of renewable solar and wind energy in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Various types of demand response (DR) services [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: The “hidden emissions” of grid regulation reserves [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: EcoCenter resource management. Dashed blue boxes show ROCm modifications to provide power capping and core allocations APIs for the Controller to reshape server power. its power based on the regulation signal. At the end of each hour, the power grid credits the data centers with a monetary reward based on the performance score, which is taken into account in electricity cost savings on the data center. We … view at source ↗
Figure 5
Figure 5. Figure 5: EcoCenter overview. Red boxes show the inputs to the framework and the green boxes show the output. 4.1 System Overview EcoCenter overcomes GPU’s power modulation limitations by coordinating GPU DVFS, compute unit scaling, and multi-GPU load assignment as novel power modulation knobs [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: (a) GPT2 model training power model (b) GPT2 [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: EcoCenter power reshaping policy with 8 GPUs. Workload Monitor monitors the LC resources, while Resource Manager (directly by Controller) reshapes the BE workload. to regulation provision. Unlike CPUs, which includes a plethora of low-power state knobs (DVFS, C-states, fine-grain power gating, preemption, etc.) to optimize non-peak utilization, GPU’s power knobs are limited to power capping and CU scaling,… view at source ↗
Figure 9
Figure 9. Figure 9: Regulation signals used in evaluation with varying [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of Exogenous carbon savings with data center carbon emissions at different operational loads. [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 14
Figure 14. Figure 14: Normalized TCO w.r.t. baseline data center shows [PITH_FULL_IMAGE:figures/full_fig_p011_14.png] view at source ↗

Review history (2 revisions) →

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Inter-Area Oscillation Damping in Data-Center-Integrated Power Systems

    eess.SY 2026-07 conditional novelty 5.0

    UPS-based demand response raises the critical inter-area mode's damping ratio from about 1.29% to 2.24% on the IEEE 39-bus system; the HVAC channel is too slow to contribute.

Reference graph

Works this paper leans on

90 extracted references · 4 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Powering intelligence: Analyzing artificial intelligence and data center energy consumption,

    J. Aljbour, T. Wilson, and P. Patel, “Powering intelligence: Analyzing artificial intelligence and data center energy consumption, ”EPRI White Paper, 2024

    2024

  2. [2]

    Stanley, 2023

    M. Stanley, 2023. [Online]. Available: https://www.datacenterdynamics.com/en/ news/morgan-stanley-data-center-industry-will-emit-25bn-tons-of-co2-by- 2030/

    2023

  3. [3]

    Environmental report,

    Google, “Environmental report, ” 2023. [Online]. Available: https://www.gstatic. com/gumdrop/sustainability/google-2023-environmental-report.pdf

    2023

  4. [4]

    Powering sustainable transformation,

    Microsoft, “Powering sustainable transformation, ” 2025. [On- line]. Available: https://datacenters.microsoft.com/globe/powering-sustainable- transformation/

    2025

  5. [5]

    Electricity data browser,

    U. E. I. A. (EIA), “Electricity data browser, ” https://www.eia.gov/electricity/data/ browser/, 2023

    2023

  6. [6]

    Powermorph: Qos-aware server power reshaping for data center regulation service,

    A. Jahanshahi, N. Yu, and D. Wong, “Powermorph: Qos-aware server power reshaping for data center regulation service, ”ACM Transactions on Architecture and Code Optimization (TACO), vol. 19, no. 3, pp. 1–27, 2022

    2022

  7. [7]

    Supporting power grids with demand response at google data centers,

    Google, “Supporting power grids with demand response at google data centers, ”

  8. [8]

    EnergyQARE: QoS-Aware Data Center Participation in Smart Grid Regulation Service Reserve Provision,

    H. Chen, Y. Zhang, M. C. Caramanis, and A. K. Coskun, “EnergyQARE: QoS-Aware Data Center Participation in Smart Grid Regulation Service Reserve Provision, ”ACM Trans. Model. Perform. Eval. Comput. Syst., vol. 4, no. 1, pp. 2:1–2:31, Jan. 2019. [Online]. Available: http://doi.acm.org/10.1145/3243172

    doi:10.1145/3243172 2019

  9. [9]

    Towards improved power management in cloud gpus,

    P. Patel, Z. Gong, S. Rizvi, E. Choukse, P. Misra, T. Anderson, and A. Srira- man, “Towards improved power management in cloud gpus, ”IEEE Computer Architecture Letters, vol. 22, no. 2, pp. 141–144, 2023

    2023

  10. [10]

    Clicking clean virginia the dirty energy powering data center alley,

    G. Cook, E. Jardim, and C. Craighill, “Clicking clean virginia the dirty energy powering data center alley, ”Greenpeace A vailable from: https://www. greenpeace. org/usa/reports/click-clean-virginia/[Accessed 21 Mar 2020], 2019

    2020

  11. [11]

    Ai is poised to drive 160% increase in data center power demand,

    “Ai is poised to drive 160% increase in data center power demand, ” 2024. [Online]. Available: https://www.goldmansachs.com/insights/articles/AI-poised- to-drive-160-increase-in-power-demand

    2024

  12. [12]

    Llama 2: Open foundation and fine-tuned chat models,

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosaleet al., “Llama 2: Open foundation and fine-tuned chat models, ”arXiv preprint arXiv:2307.09288, 2023

    Pith/arXiv arXiv 2023

  13. [13]

    Carbon explorer: A holistic framework for designing carbon aware datacenters,

    B. Acun, B. Lee, F. Kazhamiaka, K. Maeng, U. Gupta, M. Chakkaravarthy, D. Brooks, and C.-J. Wu, “Carbon explorer: A holistic framework for designing carbon aware datacenters, ” inProceedings of the 28th ACM International Confer- ence on Architectural Support for Programming Languages and Operating Systems, Volume 2, 2023, pp. 118–132

    2023

  14. [14]

    Carbon responder: Coordinating demand response for the datacenter fleet,

    J. Xing, B. Acun, A. Sundarrajan, D. Brooks, M. Chakkaravarthy, N. Avila, C.-J. Wu, and B. C. Lee, “Carbon responder: Coordinating demand response for the datacenter fleet, ” 2023. [Online]. Available: https://arxiv.org/abs/2311.08589

    Pith/arXiv arXiv 2023

  15. [15]

    Carbon-aware electricity cost minimization for sustainable data centers,

    H. Dou, Y. Qi, W. Wei, and H. Song, “Carbon-aware electricity cost minimization for sustainable data centers, ”IEEE Transactions on Sustainable Computing, vol. 2, no. 2, pp. 211–223, 2017

    2017

  16. [16]

    Temporal load balancing with service delay guarantees for data center energy cost optimization,

    J. Luo, L. Rao, and X. Liu, “Temporal load balancing with service delay guarantees for data center energy cost optimization, ”IEEE Transactions on Parallel and Distributed Systems, vol. 25, no. 3, pp. 775–784, 2013

    2013

  17. [17]

    On the limitations of carbon-aware temporal and spatial workload shifting in the cloud,

    T. Sukprasert, A. Souza, N. Bashir, D. Irwin, and P. Shenoy, “On the limitations of carbon-aware temporal and spatial workload shifting in the cloud, ” in Proceedings of the Nineteenth European Conference on Computer Systems, ser. EuroSys ’24. New York, NY, USA: Association for Computing Machinery, 2024, p. 924–941. [Online]. Available: https://doi.org/10...

    arXiv 2024

  18. [18]

    Clover: Toward sustainable ai with carbon-aware machine learning inference service,

    B. Li, S. Samsi, V. Gadepally, and D. Tiwari, “Clover: Toward sustainable ai with carbon-aware machine learning inference service, ” inProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, 2023, pp. 1–15

    2023

  19. [19]

    In california, solar and wind boost the price of frequency regulation,

    GreenTechMedia, “In california, solar and wind boost the price of frequency regulation, ” 2022. [Online]. Available: http://www.greentechmedia.com/articles/ read/in-california-solar-and-wind-boosts-the-price-for-frequency-regulation

    2022

  20. [20]

    2025 california demand response potential study-charting california’s demand response future. final report on phase 2 results,

    P. Alstone, J. Potter, M. A. Piette, P. Schwartz, M. A. Berger, L. N. Dunn, S. J. Smith, M. D. Sohn, A. Aghajanzadeh, S. Stenssonet al., “2025 california demand response potential study-charting california’s demand response future. final report on phase 2 results, ” Lawrence Berkeley National Lab.(LBNL), Berkeley, CA (United States), Tech. Rep., 2017

    2025

  21. [21]

    Data center participation in demand response programs with quality-of-service guarantees,

    Y. Zhang, I. C. Paschalidis, and A. K. Coskun, “Data center participation in demand response programs with quality-of-service guarantees, ” inProceedings of the Tenth ACM International Conference on Future Energy Systems, 2019, pp. 285–302

    2019

  22. [22]

    Hpc data center participation in demand response: An adaptive policy with qos assurance,

    Y. Zhang, D. C. Wilson, I. C. Paschalidis, and A. K. Coskun, “Hpc data center participation in demand response: An adaptive policy with qos assurance, ”IEEE Transactions on Sustainable Computing, vol. 7, no. 1, pp. 157–171, 2022

    2022

  23. [23]

    Opportunities and challenges for data center demand response,

    A. Wierman, Z. Liu, I. Liu, and H. Mohsenian-Rad, “Opportunities and challenges for data center demand response, ” inInternational Green Computing Conference, 2014, pp. 1–10

    2014

  24. [24]

    Iswitch: Coordinating and optimizing renewable energy powered server clusters,

    C. Li, A. Qouneh, and T. Li, “Iswitch: Coordinating and optimizing renewable energy powered server clusters, ” inProceedings of the 39th Annual International Symposium on Computer Architecture, ser. ISCA ’12. USA: IEEE Computer Society, 2012, p. 512–523

    2012

  25. [25]

    Greenslot: scheduling energy consumption in green datacenters,

    Í. Goiri, K. Le, M. E. Haque, R. Beauchea, T. D. Nguyen, J. Guitart, J. Torres, and R. Bianchini, “Greenslot: scheduling energy consumption in green datacenters, ” inInternational Conference for High Performance Computing, Networking, Storage and Analysis, 2011

    2011

  26. [26]

    Solarcore: Solar energy driven multi-core architecture power management,

    C. Li, W. Zhang, C. Cho, and T. Li, “Solarcore: Solar energy driven multi-core architecture power management, ” in2011 IEEE 17th International Symposium on High Performance Computer Architecture, Feb 2011, pp. 205–216

    2011

  27. [27]

    Chameleon: Adapting throughput server to time-varying green power budget using online learning,

    C. Li, X. Li, R. Wang, T. Li, N. Goswami, and D. Qian, “Chameleon: Adapting throughput server to time-varying green power budget using online learning, ” inInternational Symposium on Low Power Electronics and Design (ISLPED), Sep. 2013, pp. 100–105

    2013

  28. [28]

    Geographical load balancing with renewables,

    Z. Liu, M. Lin, A. Wierman, S. H. Low, and L. L. Andrew, “Geographical load balancing with renewables, ”SIGMETRICS Perform. Eval. Rev., vol. 39, no. 3, p. 62–66, Dec. 2011. [Online]. Available: https://doi.org/10.1145/2160803.2160862

    arXiv 2011

  29. [29]

    Adapting datacenter capacity for greener datacenters and grid,

    L. Lin and A. A. Chien, “Adapting datacenter capacity for greener datacenters and grid, ” inProceedings of the 14th ACM International Conference on Future Energy Systems, 2023, pp. 200–213

    2023

  30. [30]

    Carbonscaler: Leveraging cloud workload elasticity for optimizing carbon-efficiency,

    W. A. Hanafy, Q. Liang, N. Bashir, D. Irwin, and P. Shenoy, “Carbonscaler: Leveraging cloud workload elasticity for optimizing carbon-efficiency, ”Proc. ACM Meas. Anal. Comput. Syst., vol. 7, no. 3, Dec. 2023. [Online]. Available: https://doi.org/10.1145/3626788

    doi:10.1145/3626788 2023

  31. [31]

    The sunk carbon fallacy: Rethinking carbon footprint metrics for effective carbon-aware scheduling,

    N. Bashir, V. Gohil, A. B. Subramanya, M. Shahrad, D. Irwin, E. Olivetti, and C. Delimitrou, “The sunk carbon fallacy: Rethinking carbon footprint metrics for effective carbon-aware scheduling, ” inProceedings of the 2024 ACM Symposium on Cloud Computing, ser. SoCC ’24. New York, NY, USA: Association for Computing Machinery, 2024, p. 542–551. [Online]. Av...

    arXiv 2024

  32. [32]

    Greenhouse gas emission tracking methodology,

    California ISO, “Greenhouse gas emission tracking methodology, ” Tech. Rep., 2016

    2016

  33. [33]

    Greenhouse gas coordination stakeholder recommendations for policy development,

    ——, “Greenhouse gas coordination stakeholder recommendations for policy development, ” Tech. Rep., 2024

    2024

  34. [34]

    PJM Manual 12: Balancing Operations,

    PJM, “PJM Manual 12: Balancing Operations, ” https://www.pjm.com/-/media/ documents/manuals/m12.ashx, 2022

    2022

  35. [35]

    Performance accuracy scores in caiso and miso regulation markets: A comparison based on real data and mathematical analysis,

    A. Sadeghi-Mobarakeh and H. Mohsenian-Rad, “Performance accuracy scores in caiso and miso regulation markets: A comparison based on real data and mathematical analysis, ”IEEE Transactions on Power Systems, vol. 33, no. 3, pp. 3196–3198, 2018

    2018

  36. [36]

    The data center as a grid load stabilizer,

    H. Chen, M. C. Caramanis, and A. K. Coskun, “The data center as a grid load stabilizer, ” in19th Asia and South Pacific Design Automation Conference (ASP-DAC). Singapore: IEEE, Jan. 2014, pp. 105–112. [Online]. Available: http://ieeexplore.ieee.org/document/6742874/

    arXiv 2014

  37. [37]

    Dynamic Server Power Capping for Enabling Data Center Participation in Power Markets,

    H. Chen, C. Hankendi, M. C. Caramanis, and A. K. Coskun, “Dynamic Server Power Capping for Enabling Data Center Participation in Power Markets, ” in Proceedings of the International Conference on Computer-Aided Design, ser. ICCAD ’13. Piscataway, NJ, USA: IEEE Press, 2013, pp. 122–129

    2013

  38. [38]

    Reducing the data center electric- ity costs through participation in smart grid programs,

    H. Chen, M. C. Caramanis, and A. K. Coskun, “Reducing the data center electric- ity costs through participation in smart grid programs, ” inInternational Green Computing Conference, Nov. 2014, pp. 1–10

    2014

  39. [39]

    Characterizing power management opportunities for llms in the cloud,

    P. Patel, E. Choukse, C. Zhang, I. n. Goiri, B. Warrier, N. Mahalingam, and R. Bianchini, “Characterizing power management opportunities for llms in the cloud, ” inProceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 3, ser. ASPLOS ’24. New York, NY, USA: Association for Comp...

    arXiv 2024

  40. [40]

    Toward sustainable hpc: Carbon footprint estimation and environmental implications of hpc systems,

    B. Li, R. Basu Roy, D. Wang, S. Samsi, V. Gadepally, and D. Tiwari, “Toward sustainable hpc: Carbon footprint estimation and environmental implications of hpc systems, ” inProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, 2023, pp. 1–15

    2023

  41. [41]

    A guide to reducing carbon emissions through data center geographical load shifting,

    J. Lindberg, Y. Abdennadher, J. Chen, B. C. Lesieutre, and L. Roald, “A guide to reducing carbon emissions through data center geographical load shifting, ” in Proceedings of the Twelfth ACM International Conference on Future Energy Systems, ser. e-Energy ’21, 2021

    2021

  42. [42]

    Quantifying the carbon emissions of machine learning,

    A. Lacoste, A. Luccioni, V. Schmidt, and T. Dandres, “Quantifying the carbon emissions of machine learning, ” 2019. [Online]. Available: https: //arxiv.org/abs/1910.09700

    Pith/arXiv arXiv 2019

  43. [43]

    The leading api for granular electricity data reduce carbon emissions with actionable electricity data

    “The leading api for granular electricity data reduce carbon emissions with actionable electricity data. ” 2024. [Online]. Available: https://www. electricitymaps.com

    2024

  44. [44]

    Electricityemissions.jl: A framework for the comparison of carbon intensity signals,

    J. Gorka, N. Rhodes, and L. Roald, “Electricityemissions.jl: A framework for the comparison of carbon intensity signals, ” 2024. [Online]. Available: https://arxiv.org/abs/2411.06560

    Pith/arXiv arXiv 2024

  45. [45]

    On the limitations of carbon-aware temporal and spatial workload shifting in the cloud,

    T. Sukprasert, A. Souza, N. Bashir, D. Irwin, and P. Shenoy, “On the limitations of carbon-aware temporal and spatial workload shifting in the cloud, ” inProceedings Coordinating Power Grid Frequency Regulation Service with Data Center Load Flexibility of the Nineteenth European Conference on Computer Systems, ser. EuroSys ’24, 2024

    2024

  46. [46]

    Process control strategies for reducing the minimum load of fossil- fired plants,

    S. Seachman, “Process control strategies for reducing the minimum load of fossil- fired plants, ” 2025. [Online]. Available: https://www.powermag.com/process- control-strategies-for-reducing-the-minimum-load-of-fossil-fired-plants/

    2025

  47. [47]

    Evaluation of market rules using a multi-agent system method,

    N.-P. Yu, C.-C. Liu, and J. Price, “Evaluation of market rules using a multi-agent system method, ”IEEE Transactions on Power Systems, vol. 25, no. 1, pp. 470–479, 2010

    2010

  48. [48]

    Feasibility study of combined cycle gas turbine (ccgt) power plant integration with adiabatic compressed air energy storage (acaes),

    J. D. Wojcik and J. Wang, “Feasibility study of combined cycle gas turbine (ccgt) power plant integration with adiabatic compressed air energy storage (acaes), ” Applied Energy, vol. 221, pp. 477–489, 2018

    2018

  49. [49]

    Ge gas turbine performance characteristics,

    F. J. Brooks, “Ge gas turbine performance characteristics, ” GE Power Systems, Tech. Rep., 2000

    2000

  50. [50]

    Optimizing deep decarbonization path- ways in California with power system planning using surrogate level-based Lagrangian relaxation,

    O. Anderson, M. A. Bragin, and N. Yu, “Optimizing deep decarbonization path- ways in California with power system planning using surrogate level-based Lagrangian relaxation, ”Applied Energy, vol. 377, 2025

    2025

  51. [51]

    Nvidia a100 tensor core gpu architecture,

    NVIDIA, “Nvidia a100 tensor core gpu architecture, ” https://images.nvidia. com/aem-dam/en-zz/Solutions/data-center/nvidia-ampere-architecture- whitepaper.pdf, 2020

    2020

  52. [52]

    Nvidia h100 tensor core gpu architecture,

    ——, “Nvidia h100 tensor core gpu architecture, ” https://www.advancedclustering. com/wp-content/uploads/2022/03/gtc22-whitepaper-hopper.pdf, 2022

    2022

  53. [53]

    Nvidia blackwell architecture technical brief,

    ——, “Nvidia blackwell architecture technical brief, ” https://cdn.prod.website- files.com/61dda201f29b7efc52c5fbaf/6602ea9d0ce8cb73fb6de87f_nvidia- blackwell-architecture-technical-brief.pdf, 2024

    2024

  54. [54]

    Workload prediction using arima model and its impact on cloud applications’ qos,

    R. N. Calheiros, E. Masoumi, R. Ranjan, and R. Buyya, “Workload prediction using arima model and its impact on cloud applications’ qos, ”IEEE Transactions on Cloud Computing, vol. 3, no. 4, pp. 449–458, 2015

    2015

  55. [55]

    Ame-wpc: Advanced model for efficient workload prediction in the cloud,

    K. Cetinski and M. B. Juric, “Ame-wpc: Advanced model for efficient workload prediction in the cloud, ”Journal of Network and Computer Applications, vol. 55, pp. 191–201, 2015

    2015

  56. [56]

    Host load prediction in a google compute cloud with a bayesian model,

    S. Di, D. Kondo, and W. Cirne, “Host load prediction in a google compute cloud with a bayesian model, ” inSC’12: Proceedings of the International Conference on High Performance Computing, Networking, Storage and Analysis. IEEE, 2012, pp. 1–11

    2012

  57. [57]

    Rpps: A novel resource prediction and provisioning scheme in cloud data center,

    W. Fang, Z. Lu, J. Wu, and Z. Cao, “Rpps: A novel resource prediction and provisioning scheme in cloud data center, ” in2012 IEEE Ninth International Conference on Services Computing, 2012, pp. 609–616

    2012

  58. [58]

    An adaptive prediction approach based on workload pattern discrimination in the cloud,

    C. Liu, C. Liu, Y. Shang, S. Chen, B. Cheng, and J. Chen, “An adaptive prediction approach based on workload pattern discrimination in the cloud, ”Journal of Network and Computer Applications, vol. 80, pp. 35–44, 2017

    2017

  59. [59]

    Cloud workload prediction and generation models,

    G. M. Wamba, Y. Li, A.-C. Orgerie, N. Beldiceanu, and J.-M. Menaud, “Cloud workload prediction and generation models, ” in2017 29th International Sympo- sium on Computer Architecture and High Performance Computing (SBAC-PAD), 2017, pp. 89–96

    2017

  60. [60]

    An efficient deep learning model to predict cloud workload for industry informatics,

    Q. Zhang, L. T. Yang, Z. Yan, Z. Chen, and P. Li, “An efficient deep learning model to predict cloud workload for industry informatics, ”IEEE Transactions on Industrial Informatics, vol. 14, no. 7, pp. 3170–3178, 2018

    2018

  61. [61]

    Wattwiser: Power & resource-efficient scheduling for multi-model multi-gpu inference servers,

    A. Jahanshahi, M. Rezvani, and D. Wong, “Wattwiser: Power & resource-efficient scheduling for multi-model multi-gpu inference servers, ” in2023 IEEE 14th Inter- national Green and Sustainable Computing Conference (IGSC), 2023

    2023

  62. [62]

    Green contexts - cuda driver api,

    NVIDIA, “Green contexts - cuda driver api, ” 2025, accessed: 2025-10-

    2025

  63. [63]

    Hardware compute partitioning on nvidia gpus for composable systems,

    J. Bakita and J. H. Anderson, “Hardware compute partitioning on nvidia gpus for composable systems, ” in37th Euromicro Conference on Real-Time Systems (ECRTS 2025), 2025. [Online]. Available: https://drops.dagstuhl.de/entities/document/10. 4230/LIPIcs.ECRTS.2025.21

    2025

  64. [64]

    Available: https://docs.nvidia.com/cuda/cuda-driver-api/group_ _CUDA__GREEN__CONTEXTS.html

    [Online]. Available: https://docs.nvidia.com/cuda/cuda-driver-api/group_ _CUDA__GREEN__CONTEXTS.html

  65. [65]

    Z3: An efficient smt solver,

    L. De Moura and N. Bjørner, “Z3: An efficient smt solver, ” inInternational con- ference on Tools and Algorithms for the Construction and Analysis of Systems. Springer, 2008, pp. 337–340

    2008

  66. [66]

    Beware of fragmentation: Scheduling{GPU-Sharing} workloads with fragmentation gradient descent,

    Q. Weng, L. Yang, Y. Yu, W. Wang, X. Tang, G. Yang, and L. Zhang, “Beware of fragmentation: Scheduling{GPU-Sharing} workloads with fragmentation gradient descent, ” in2023 USENIX Annual Technical Conference (USENIX ATC 23), 2023, pp. 995–1008

    2023

  67. [67]

    Scaleserve: A scalable multi-gpu machine learning inference system and benchmarking suite,

    A. Jahanshahi, M. Chow, and D. Wong, “Scaleserve: A scalable multi-gpu machine learning inference system and benchmarking suite, ” inProceedings of the 14th Workshop on General Purpose Processing Using GPU, 2022, pp. 1–2

    2022

  68. [68]

    Rocm system management interface (rocm smi) library,

    AMD, “Rocm system management interface (rocm smi) library, ” https://github. com/ROCm/ROC-smi, 2024

    2024

  69. [69]

    Elfen scheduling: Fine-grain principled borrowing from latency-critical workloads using simultaneous multi- threading,

    X. Yang, S. M. Blackburn, and K. S. McKinley, “Elfen scheduling: Fine-grain principled borrowing from latency-critical workloads using simultaneous multi- threading, ” in2016 USENIX Annual Technical Conference (USENIX ATC 16), Jun. 2016

    2016

  70. [70]

    𝜇DPM: Dynamic power management for the microsecond era,

    C. Chou, L. N. Bhuyan, and D. Wong, “𝜇DPM: Dynamic power management for the microsecond era, ” in2019 IEEE International Symposium on High Performance Computer Architecture (HPCA), Feb. 2019, pp. 120–132

    2019

  71. [71]

    Swim project,

    SWIMProjectUCB, “Swim project, ” 2013. [Online]. Available: https://github.com/ SWIMProjectUCB/SWIM/wiki

    2013

  72. [72]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskeveret al., “Language models are unsupervised multitask learners, ”OpenAI blog, vol. 1, no. 8, p. 9, 2019

    2019

  73. [73]

    Overall data center costs,

    J. Hamilton, “Overall data center costs, ” 2025. [Online]. Available: https: //perspectives.mvdirona.com/2010/09/overall-data-center-costs/

    2025

  74. [74]

    Field Experience from Li-Ion BESS Delivering Primary Frequency Regulation in the Danish Energy Market,

    M. Świerczyński, D. I. Stroe, R. Lærke, A. I. Stan, P. C. Kjær, R. Teodorescu, and S. K. Kær, “Field Experience from Li-Ion BESS Delivering Primary Frequency Regulation in the Danish Energy Market, ”ECS Transactions, vol. 61, no. 37, Sep. 2014

    2014

  75. [75]

    Google data center pue performance,

    Google, “Google data center pue performance, ” 2025. [Online]. Available: https://datacenters.google/efficiency/

    2025

  76. [76]

    L. A. Barroso, J. Clidaras, and U. Hölzle,The Datacenter as a Computer: An Introduction to the Design of Warehouse-Scale Machines, Second Edition, 2013

    2013

  77. [77]

    Market performance report, april 2024, ifm procurement and prices,

    CAISO, “Market performance report, april 2024, ifm procurement and prices, ”

    2024

  78. [78]

    Amd instinct mi300 series systems,

    E. Corporation, “Amd instinct mi300 series systems, ” 2025. [Online]. Available: https://www.exxactcorp.com/category/AMD-Radeon-Instinct-Solutions

    2025

  79. [79]

    Managing green datacenters pow- ered by hybrid renewable energy systems,

    C. Li, R. Wang, T. Li, D. Qian, and J. Yuan, “Managing green datacenters pow- ered by hybrid renewable energy systems, ” in11th International Conference on Autonomic Computing (ICAC 14), Jun. 2014, pp. 261–272

    2014

  80. [80]

    Benefits and limitations of tapping into stored energy for datacenters,

    S. Govindan, A. Sivasubramaniam, and B. Urgaonkar, “Benefits and limitations of tapping into stored energy for datacenters, ” in2011 38th Annual International Symposium on Computer Architecture (ISCA), June 2011, pp. 341–351

    2011

Showing first 80 references.