pith. machine review for the scientific record. sign in

arxiv: 2604.07345 · v1 · submitted 2026-04-08 · 📡 eess.SY · cs.DC· cs.LG· cs.SY

Recognition: unknown

Measurement of Generative AI Workload Power Profiles for Whole-Facility Data Center Infrastructure Planning

Roberto Vercellino (1) , Jared Willard (1) , Gustavo Campos (1) , Weslley da Silva Pereira (1) , Olivia Hull (1) , Matthew Selensky (1) , Juliane Mueller (1) ((1) National Laboratory of the Rockies (NLR) , Golden
show 2 more authors
CO USA)
Authors on Pith no claims yet

Pith reviewed 2026-05-10 17:18 UTC · model grok-4.3

classification 📡 eess.SY cs.DCcs.LGcs.SY
keywords generative AIpower consumptiondata centerenergy modelingworkload profilinginfrastructure planningGPU powerfacility energy
0
0 comments X

The pith

High-resolution power measurements of generative AI workloads scale to realistic whole-facility energy profiles via a bottom-up model.

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

The paper sets out a method to connect short-interval power readings from AI training, fine-tuning, and inference tasks directly to the total energy use of an entire data center. It takes fine-grained measurements of these workloads and feeds them into an event-driven model that builds up facility-level demand from the bottom. A reader would care because these profiles reveal how AI jobs and usage patterns create varying rather than constant loads, which matters for sizing electrical infrastructure, grid ties, and on-site power sources. The approach produces time-varying energy traces that reflect real workload and user-driven swings. The measurement data is released publicly to allow others to apply or extend the scaling method.

Core claim

The authors introduce a methodology that records power use of generative AI workloads at 0.1-second intervals, then applies a bottom-up, event-driven data center energy model to produce whole-facility demand profiles. These profiles incorporate the temporal variations that arise from different workload types and user behavior, supplying data suitable for planning grid connections, on-site generation, and microgrids.

What carries the argument

The bottom-up, event-driven data center energy model that aggregates individual workload power measurements into facility-wide demand while preserving realistic time fluctuations.

If this is right

  • Facility planners gain time-resolved demand traces that better match actual AI-driven loads for grid-connection sizing.
  • On-site generation and microgrid designs can be matched to fluctuating patterns instead of steady-state assumptions.
  • The public release of the high-resolution profiles supports repeated or comparative studies of energy impact.
  • Infrastructure decisions can differentiate effects from training versus inference workloads.

Where Pith is reading between the lines

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

  • The same measurement-and-scaling approach could be applied to other hardware generations to test whether the fluctuation patterns remain similar.
  • Combining these profiles with time-varying renewable supply data could show how AI facilities might shift workloads to improve grid stability.
  • Extending the model to include future changes in model size or utilization rates would allow forward-looking capacity estimates.

Load-bearing premise

The scaling model converts individual workload power readings into accurate facility-level energy use without large unaccounted losses, cooling overheads, or component interactions.

What would settle it

A side-by-side comparison of the model's predicted facility power trace against direct measurements of total data-center power during a known period of AI workloads would confirm or disprove the scaling accuracy.

Figures

Figures reproduced from arXiv: 2604.07345 by CO, Golden, Gustavo Campos (1), Jared Willard (1), Juliane Mueller (1) ((1) National Laboratory of the Rockies (NLR), Matthew Selensky (1), Olivia Hull (1), Roberto Vercellino (1), USA), Weslley da Silva Pereira (1).

Figure 1
Figure 1. Figure 1: Diagram illustrating the workflow and scaling from data center compute node to facility level [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Data Center Infrastructure Planning, Load Optimization for Energy E [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: DIPLOEE model details for inference data center use case. [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Llama-2 70B fine-tuning power profiles. Replicate experiments using the same node count are [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Llama-2 70B fine-tuning energy metrics as a function of number of nodes. Replicate experiments [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Stable Diffusion training power profiles. Replicate experiments using the same node count are shown as superimposed transparent curves with different colors. Each power profile aggregates power con￾sumption across all CPU and GPU devices utilized by the workload. A similar set of results was generated for the Stable Diffusion image-generation model. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 6
Figure 6. Figure 6: A few key differences can be promptly identified, compared to the LLM fine￾tuning results. First, execution times were longer, particularly at lower node counts, taking up to 4.2 hours to complete. Evaluation is also much more time-intensive for diffusion models due to the nature of diffusion sampling, which is iterative and expensive and re￾quires many denoising steps for each sample, compared to LLM eval… view at source ↗
Figure 7
Figure 7. Figure 7: Stable Diffusion training energy metrics as a function of number of nodes. Replicate experiments are shown as red dots, vertically distributed with a box-and-whisker plot for each node value. Fitted curves are also presented as dashed lines. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Llama-3 70B offline inference power profiles as a function of number of prompts and output token length. Median values are shown as lines, and 10-90th percentiles as shaded regions. power consumption profiles. A very fast ramp-up period was observed across all scenar￾ios, followed by either a sustained or decreasing trend. The overall execution time was in the order of seconds, ranging from ∼10 seconds for… view at source ↗
Figure 9
Figure 9. Figure 9: Llama-3 70B offline inference energy metrics as a function of number of prompts and output tokens. Median values across experiments are shown as dots, and 10-90th percentiles as a shaded regions. 3.5. Online/Serve Inference Workload Power Profiles In the previous section, we benchmarked LLM inference in offline/batch mode, i.e. submitting all prompts at once. Instead, here we present results from testing t… view at source ↗
Figure 10
Figure 10. Figure 10: Power series of Llama-3 70B online inference benchmark as a function of dataset, request rate [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Llama-3 70B offline inference performance metrics as a function of dataset, request-rate (in log￾scale) and output token length. Median values are shown as lines, and 10-90th percentiles as shaded regions. connected to user experience. As discussed in Section 2.3, in the whole-facility simulation we used latency as the metric correlating total incoming request rate at the inference data center to the numb… view at source ↗
Figure 12
Figure 12. Figure 12: Llama-3 70B online inference energy metrics as a function of dataset, request-rate (in log-scale) [PITH_FULL_IMAGE:figures/full_fig_p025_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Power series of Llama-3 70B online inference benchmark as a function of dataset, request rate [PITH_FULL_IMAGE:figures/full_fig_p026_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Facility-level power consumption profile distributions for a 10 MW colocation data center at [PITH_FULL_IMAGE:figures/full_fig_p029_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Facility-level power consumption profile distributions for a 1 MW inference data center at varying [PITH_FULL_IMAGE:figures/full_fig_p031_15.png] view at source ↗
read the original abstract

The rapid growth of generative artificial intelligence (AI) has introduced unprecedented computational demands, driving significant increases in the energy footprint of data centers. However, existing power consumption data is largely proprietary and reported at varying resolutions, creating challenges for estimating whole-facility energy use and planning infrastructure. In this work, we present a methodology that bridges this gap by linking high-resolution workload power measurements to whole-facility energy demand. Using NLR's high-performance computing data center equipped with NVIDIA H100 GPUs, we measure power consumption of AI workloads at 0.1-second resolution for AI training, fine-tuning and inference jobs. Workloads are characterized using MLCommons benchmarks for model training and fine-tuning, and vLLM benchmarks for inference, enabling reproducible and standardized workload profiling. The dataset of power consumption profiles is made publicly available. These power profiles are then scaled to the whole-facility-level using a bottom-up, event-driven, data center energy model. The resulting whole-facility energy profiles capture realistic temporal fluctuations driven by AI workloads and user-behavior, and can be used to inform infrastructure planning for grid connection, on-site energy generation, and distributed microgrids.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper presents high-resolution (0.1 s) power measurements of generative AI workloads (training, fine-tuning, inference) on NVIDIA H100 GPUs in NLR's data center, using MLCommons and vLLM benchmarks for reproducibility. It releases the resulting power-profile dataset publicly and scales the traces to whole-facility energy demand via a bottom-up, event-driven data-center model that incorporates workload events and user behavior, producing time-varying facility-level profiles intended for grid-connection, on-site generation, and microgrid planning.

Significance. If the scaling step is shown to be accurate, the public dataset and methodology would supply a rare non-proprietary, high-resolution resource for AI-driven data-center planning, where most existing data are aggregated or proprietary. The choice of standardized benchmarks and the explicit release of raw traces are clear strengths that could enable community reuse and further modeling.

major comments (2)
  1. [Scaling methodology] The scaling procedure (described after the measurement section) asserts that the bottom-up event-driven model produces 'realistic temporal fluctuations' at facility scale, yet no comparison of modeled totals against metered whole-facility power data, no calibration constants, and no sensitivity analysis on cooling or distribution losses are reported. This directly undermines the central claim that the profiles are suitable for infrastructure planning.
  2. [Results and discussion] No quantitative results—peak-to-average ratios, total energy per job class, or error statistics—are supplied for either the per-GPU traces or the scaled facility profiles, leaving the practical magnitude of the reported fluctuations unassessed.
minor comments (2)
  1. [Abstract] The abstract states the dataset is made publicly available but supplies neither a DOI, repository URL, nor license; this information should be added in the data-availability statement.
  2. [Model description] Clarify the precise definition of 'event-driven' in the model (e.g., how concurrent jobs, idle periods, and cooling-system response times are scheduled) so that readers can reproduce the aggregation step.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback and for recognizing the value of the high-resolution measurements and public dataset. We address each major comment below.

read point-by-point responses

  1. Referee: [Scaling methodology] The scaling procedure (described after the measurement section) asserts that the bottom-up event-driven model produces 'realistic temporal fluctuations' at facility scale, yet no comparison of modeled totals against metered whole-facility power data, no calibration constants, and no sensitivity analysis on cooling or distribution losses are reported. This directly undermines the central claim that the profiles are suitable for infrastructure planning.

    Authors: We agree that the absence of direct validation against metered facility-level data limits the strength of claims for immediate use in infrastructure planning. The model is presented as a generalizable bottom-up framework rather than a calibrated site-specific simulator. In revision we have added a limitations subsection that explicitly states the lack of calibration constants, discusses model assumptions on losses, and includes a basic sensitivity analysis (varying cooling and distribution loss factors by ±20% and reporting resulting changes in peak and average facility demand). We have also clarified in the text that the generated profiles illustrate workload-driven temporal structure and are intended to be combined with site-specific data by end users. revision: partial

  2. Referee: [Results and discussion] No quantitative results—peak-to-average ratios, total energy per job class, or error statistics—are supplied for either the per-GPU traces or the scaled facility profiles, leaving the practical magnitude of the reported fluctuations unassessed.

    Authors: We accept that the original manuscript relied too heavily on visual inspection of traces. The revised version adds a quantitative results subsection that reports: (i) peak-to-average power ratios for each workload class on single H100 GPUs, (ii) total energy per job (kWh) for representative training, fine-tuning, and inference runs, and (iii) run-to-run variability (standard deviation and coefficient of variation) computed from the repeated benchmark executions. These statistics are also computed for the scaled facility-level traces and presented in a new summary table. revision: yes

standing simulated objections not resolved
  • Direct comparison of modeled facility totals against metered whole-facility power data at matching resolution, because such proprietary meter data were not collected or released as part of this study.

Circularity Check

0 steps flagged

No circularity: direct measurements scaled by bottom-up model

full rationale

The paper's derivation chain consists of (1) direct 0.1 s power measurements on H100 GPUs using MLCommons and vLLM benchmarks and (2) scaling those traces via an event-driven bottom-up facility model. No step reduces a claimed prediction or result to the input data by construction, no fitted parameters are renamed as predictions, and no load-bearing claims rest on self-citations. The methodology is self-contained and externally falsifiable against metered facility data.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Insufficient information in the abstract to identify specific free parameters, axioms, or invented entities used in the bottom-up energy model.

pith-pipeline@v0.9.0 · 5563 in / 1149 out tokens · 51338 ms · 2026-05-10T17:18:09.000801+00:00 · methodology

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

54 extracted references · 28 canonical work pages · 5 internal anchors

  1. [1]

    2024 United States Data Center Energy Usage Report,

    A. Shehabi, S. J. Smith, A. Hubbard, A. Newkirk, N. Lei, M. A. Siddik, B. Holecek, J. G. Koomey, E. R. Masanet, and D. A. Sartor, “2024 United States Data Center Energy Usage Report,” tech. rep., Lawrence Berkeley National Laboratory (LBNL),

    2024

  2. [2]

    doi: 10.71468/P1WC7Q

    work page doi:10.71468/p1wc7q

  3. [3]

    Energy and AI,

    International Energy Agency, “Energy and AI,” tech. rep., International Energy Agency (IEA), Paris, France, 2025. url: https://www.iea.org/reports/energy-and- ai/executive-summary

    2025

  4. [4]

    DOE Releases New Report Eval- uating Increase in Electricity Demand from Data Centers,

    U.S. Department of Energy (DOE), “DOE Releases New Report Eval- uating Increase in Electricity Demand from Data Centers,” 2024. url: https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase- electricity-demand-data-centers, (Accessed: 2026-04-06)

    2024

  5. [5]

    Electricity demand and grid impacts of AI data centers: Challenges and prospects,

    X. Chen, X. Wang, A. Colacelli, M. Lee, and L. Xie, “Electricity demand and grid impacts of AI data centers: Challenges and prospects,”arXiv, 2025. doi: 10.48550/arXiv.2509.07218

    work page doi:10.48550/arxiv.2509.07218 2025

  6. [6]

    Preprint at https://arxiv.org/abs/ 2508.14318 (2025)

    E. Choukse, B. Warrier, S. Heath, L. Belmont, A. Zhao, H. A. Khan, B. Harry, M. Kappel, R. J. Hewett, K. Datta, Y . Pei, C. Lichtenberger, J. Siegler, D. Lukofsky, Z. Kahn, G. Sahota, A. Sullivan, C. Frederick, H. Thai, R. Naughton, D. Jurnove, J. Harp, R. Carper, N. Mahalingam, S. Varkala, A. G. Kumbhare, S. Desai, V . Rama- murthy, P. Gottumukkala, G. B...

    work page doi:10.48550/arxiv.2508.14318 2025

  7. [7]

    Incident re- view, considering simultaneous voltage-sensitive load reductions,

    North American Electric Reliability Corporation (NERC), “Incident re- view, considering simultaneous voltage-sensitive load reductions,” tech. rep., North American Electric Reliability Corporation (NERC), 2025. url: https://www.nerc.com/newsroom/nerc-publishes-incident-review-and-guidance- on-voltage-sensitive-large-load-integration

    2025

  8. [8]

    Machine learning-based cloud resource allocation al- gorithms: a comprehensive comparative review,

    D. Bodra and S. Khairnar, “Machine learning-based cloud resource allocation al- gorithms: a comprehensive comparative review,”Frontiers in Computer Science, vol. V olume 7 - 2025, 2025. doi: 10.3389/fcomp.2025.1678976

    work page doi:10.3389/fcomp.2025.1678976 2025

  9. [9]

    Data center power supply systems: From grid edge to point-of-load,

    Y . Chen, K. Shi, M. Chen, and D. Xu, “Data center power supply systems: From grid edge to point-of-load,”IEEE Journal of Emerging and Selected Topics in Power Elec- tronics, vol. 11, no. 3, pp. 2441–2456, 2023. doi: 10.1109/JESTPE.2022.3229063

    work page doi:10.1109/jestpe.2022.3229063 2023

  10. [10]

    Sources of data center energy estimates: A comprehensive review,

    D. Mytton and M. Ashtine, “Sources of data center energy estimates: A comprehensive review,”Joule, vol. 6, no. 9, pp. 2032–2056, 2022. doi: https://doi.org/10.1016/j.joule.2022.07.011

    work page doi:10.1016/j.joule.2022.07.011 2032

  11. [11]

    Dynamic model and converter-based emulator of a data center power distribution system,

    J. Sun, S. Wang, J. Wang, and L. M. Tolbert, “Dynamic model and converter-based emulator of a data center power distribution system,”IEEE Transactions on Power Electronics, vol. 37, no. 7, pp. 8420–8432, 2022. doi: 10.1109/TPEL.2022.3146354

    work page doi:10.1109/tpel.2022.3146354 2022

  12. [12]

    Electromagnetic transient modeling of large data cen- ters for grid-level studies,

    B. A. Ross and J. D. Follum, “Electromagnetic transient modeling of large data cen- ters for grid-level studies,” Tech. Rep. PNNL–38817, Pacific Northwest National Laboratory (PNNL), Richland, W A (United States), 2026. doi: 10.2172/3013288

    work page doi:10.2172/3013288 2026

  13. [13]

    Short-term load forecasting for ai-data center,

    M. Mughees, Y . Li, Y . Chen, and Y . R. Li, “Short-term load forecasting for AI-data center,”arXiv, no. arXiv:2503.07756, 2025. doi: 10.48550/arXiv.2503.07756

    work page doi:10.48550/arxiv.2503.07756 2025

  14. [14]

    Empirical measurements of AI training power demand on a GPU - accelerated node,

    I. Latif, A. C. Newkirk, M. R. Carbone, A. Munir, Y . Lin, J. Koomey, X. Yu, and Z. Dong, “Empirical measurements of ai training power demand on a gpu-accelerated node,”arXiv, 2024. doi: 10.48550/arXiv.2412.08602

    work page doi:10.48550/arxiv.2412.08602 2024

  15. [15]

    ISBN 9798400703867

    P. Patel, E. Choukse, C. Zhang, I. n. Goiri, B. Warrier, N. Mahalingam, and R. Bian- chini, “Characterizing power management opportunities for llms in the cloud,” in Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 3, ASPLOS ’24, (New York, NY , USA), p. 207–222, Associat...

    work page doi:10.1145/3620666.3651329 2024

  16. [16]

    Splitwise: Efficient generative llm inference using phase splitting

    P. Patel, E. Choukse, C. Zhang, A. Shah, I. Goiri, S. Maleki, and R. Bianchini, “Split- wise: Efficient generative llm inference using phase splitting,” in2024 ACM/IEEE 51st Annual International Symposium on Computer Architecture (ISCA), pp. 118– 132, 2024. doi: 10.1109/ISCA59077.2024.00019

    work page doi:10.1109/isca59077.2024.00019 2024

  17. [17]

    Azure public dataset: Azure llm inference trace 2023,

    Microsoft, “Azure public dataset: Azure llm inference trace 2023,” 2023. url: https://github.com/Azure/AzurePublicDataset/blob/master/AzureLLMInferenceDataset2023.md, (Accessed: 2026-04-06)

    2023

  18. [18]

    Characterization of large lan- guage model development in the datacenter,

    Q. Hu, Z. Ye, Z. Wang, G. Wang, M. Zhang, Q. Chen, P. Sun, D. Lin, X. Wang, Y . Luo, Y . Wen, and T. Zhang, “Characterization of large lan- guage model development in the datacenter,” inProceedings of the 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI ’24), (Santa Clara, CA), pp. 709–729, USENIX Association, Apr. 2024. url: htt...

    2024

  19. [19]

    Acmetrace: Gpu workload traces from shanghai ai lab,

    InternLM, “Acmetrace: Gpu workload traces from shanghai ai lab,” 2023. url: https://github.com/InternLM/AcmeTrace, (Accessed: 2026-04-06)

    2023

  20. [20]

    MLaaS in the wild: Workload analysis and scheduling in Large-Scale het- erogeneous GPU clusters,

    Q. Weng, W. Xiao, Y . Yu, W. Wang, C. Wang, J. He, Y . Li, L. Zhang, W. Lin, and Y . Ding, “MLaaS in the wild: Workload analysis and scheduling in Large-Scale het- erogeneous GPU clusters,” in19th USENIX Symposium on Networked Systems De- sign and Implementation (NSDI 22), (Renton, W A), pp. 945–960, USENIX Associ- ation, Apr. 2022. url: https://www.useni...

    2022

  21. [21]

    Alibaba cluster trace gpu 2020,

    Alibaba, “Alibaba cluster trace gpu 2020,” 2020. url: https://github.com/alibaba/clusterdata/tree/master/cluster-trace-gpu-v2020, (Accessed: 2026-04-06)

    2020

  22. [22]

    BurstGPT: A real-world workload dataset to optimize LLM serving systems,

    Y . Wang, Y . Chen, Z. Li, X. Kang, Y . Fang, Y . Zhou, Y . Zheng, Z. Tang, X. He, R. Guo, X. Wang, Q. Wang, A. C. Zhou, and X. Chu, “Burstgpt: A real-world workload dataset to optimize llm serving systems,” inProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’25), (Toronto, ON, Canada), ACM, 2025. doi: 10.1145/3711...

    work page doi:10.1145/3711896.3737413 2025

  23. [23]

    BurstGPT: A chatgpt (gpt-3.5) & gpt-4 workload trace to optimize llm serving systems,

    HPMLL, “BurstGPT: A chatgpt (gpt-3.5) & gpt-4 workload trace to optimize llm serving systems,” 2024. url: https://github.com/HPMLL/BurstGPT, (Accessed: 2026-04-06)

    2024

  24. [24]

    Awesome-CloudComputing-Datasets: A curated list of cloud computing datasets,

    ACAT-SCUT, “Awesome-CloudComputing-Datasets: A curated list of cloud computing datasets,” 2024. url: https://github.com/ACAT-SCUT/Awesome- CloudComputing-Datasets, (Accessed: 2026-04-06). 39

    2024

  25. [25]

    Dataset of generative ai workload power profiles,

    R. Vercellino, J. Willard, G. Campos, W. da Silva Pereira, O. Hull, M. Selen- sky, and J. Mueller, “Dataset of generative ai workload power profiles,” 01 2026. doi: 10.7799/3025227

    work page doi:10.7799/3025227 2026

  26. [26]

    At- lanta, GA: American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE), 5th ed., 2021

    ASHRAE TC 9.9,Thermal Guidelines for Data Processing Environments. At- lanta, GA: American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE), 5th ed., 2021. url: https://www.ashrae.org/technical- resources/bookstore/datacom-series

    2021

  27. [27]

    WattAMeter,

    W. da Silva Pereira, “WattAMeter,” 2025. url: https://github.com/NatLabRockies/WattAMeter, (Accessed: 2026-04-06)

    2025

  28. [28]

    NLR HPC Resources,

    NLR, “NLR HPC Resources,” 2025. url: https://natlabrockies.github.io/HPC, (Ac- cessed: 2026-04-06)

    2025

  29. [29]

    url: https://docs.nvidia.com/deploy/nvml-api

    Nvidia Corporation,NVML API Reference Guide, vR580 ed., September 2025. url: https://docs.nvidia.com/deploy/nvml-api

    2025

  30. [30]

    InProceedings of the International Conference for High Performance Computing, Networking, Storage, and Analysis(Atlanta, GA, USA)(SC ’24)

    Z. Yang, K. Adamek, and W. Armour, “Accurate and convenient energy mea- surements for gpus: A detailed study of nvidia gpu’s built-in power sen- sor,” inProceedings of the International Conference for High Performance Computing, Networking, Storage, and Analysis, SC ’24, IEEE Press, 2024. doi: 10.1109/SC41406.2024.00028

    work page doi:10.1109/sc41406.2024.00028 2024

  31. [31]

    Rapl in action: Experiences in using rapl for power measurements,

    K. N. Khan, M. Hirki, T. Niemi, J. K. Nurminen, and Z. Ou, “Rapl in action: Expe- riences in using rapl for power measurements,”ACM Trans. Model. Perform. Eval. Comput. Syst., vol. 3, Mar. 2018. doi: 10.1145/3177754

    work page doi:10.1145/3177754 2018

  32. [32]

    A validation of dram rapl power mea- surements,

    S. Desrochers, C. Paradis, and V . M. Weaver, “A validation of dram rapl power mea- surements,” inProceedings of the Second International Symposium on Memory Sys- tems, MEMSYS ’16, (New York, NY , USA), p. 455–470, Association for Computing Machinery, 2016. doi: 10.1145/2989081.2989088

    work page doi:10.1145/2989081.2989088 2016

  33. [33]

    Mlperf training benchmark,

    P. Mattson, C. Cheng, C. Coleman, G. Diamos, P. Micikevicius, D. Patterson, H. Tang, G.-Y . Wei, P. Bailis, V . Bittorf, D. Brooks, D. Chen, D. Dutta, U. Gupta, K. Hazelwood, A. Hock, X. Huang, A. Ike, B. Jia, D. Kang, D. Kanter, N. Kumar, J. Liao, G. Ma, D. Narayanan, T. Oguntebi, G. Pekhimenko, L. Pentecost, V . J. Reddi, T. Robie, T. St. John, T. Tabar...

    work page doi:10.48550/arxiv.1910.01500 2019

  34. [34]

    Mlperf training benchmark v4.0,

    MLCommons, “Mlperf training benchmark v4.0,” 2024. url: https://mlcommons.org/2024/06/mlperf-training-v4-benchmark-results/, (Ac- cessed: 2026-04-06)

    2024

  35. [35]

    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. Bhosale, D. Bikel, L. Blecher, C. Canton Ferrer, M. Chen, G. Cucurull, D. Esiobu, J. Fernandes, J. Fu, W. Fu, B. Fuller, C. Gao, V . Goswami, N. Goyal, A. Hartshorn, S. Hosseini, R. Hou, H. Inan, M. Kardas, V . Kerkez, M. Khabsa, I. Kloumann, A. K...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2307.09288 2023

  36. [36]

    2022 , archiveprefix =

    U. Shaham, E. Segal, M. Ivgi, A. Efrat, O. Yoran, A. Haviv, A. Gupta, W. Xiong, M. Geva, J. Berant,et al., “Scrolls: Standardized comparison over long language sequences,”arXiv, 2022. doi: 10.48550/arXiv.2201.03533

    work page doi:10.48550/arxiv.2201.03533 2022

  37. [37]

    LoRA: Low-Rank Adaptation of Large Language Models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,”arXiv, no. arXiv:2106.09685, 2021. doi: 10.48550/arXiv.2106.09685

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2106.09685 2021

  38. [38]

    Zero-offload: Democratizing billion-scale model training,

    J. Ren, S. Rajbhandari, R. Y . Aminabadi, O. Ruwase, S. Yang, M. Zhang, D. Li, and Y . He, “Zero-offload: Democratizing billion-scale model training,” in2021 USENIX Annual Technical Conference (USENIX ATC 21), pp. 551–564, 2021. url: https://www.usenix.org/conference/atc21/presentation/ren-jie

    2021

  39. [39]

    LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs

    C. Schuhmann, R. Vencu, R. Beaumont, R. Kaczmarczyk, C. Mullis, A. Katta, T. Coombes, J. Jitsev, and A. Komatsuzaki, “Laion-400m: Open dataset of clip- filtered 400 million image-text pairs,”arXiv, 2021. doi: 10.48550/arXiv.2111.02114

    work page internal anchor Pith review doi:10.48550/arxiv.2111.02114 2021

  40. [40]

    Mlcommons training: Reference implementations of mlperf,

    MLCommons, “Mlcommons training: Reference implementations of mlperf,” 2025. url: https://github.com/mlcommons/training, (Accessed: 2026-04-06)

    2025

  41. [41]

    Efficient Memory Management for Large Language Model Serving with PagedAttention

    W. Kwon, Z. Li, S. Zhuang, Y . Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, and I. Stoica, “Efficient memory management for large language model serving with pagedattention,”arXiv, 2023. doi: 10.48550/arXiv.2309.06180. 41

    work page internal anchor Pith review doi:10.48550/arxiv.2309.06180 2023

  42. [42]

    What are tokens and how to count them?,

    OpenAI, “What are tokens and how to count them?,” 2025. url: https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to- count-them, (Accessed: 2026-04-06)

    work page arXiv 2025

  43. [43]

    What is a context window?,

    D. Bergmann, “What is a context window?,” 2025. url: https://www.ibm.com/think/topics/context-window, (Accessed: 2026-04-06)

    2025

  44. [44]

    AI Load Dynamics-A Power Electronics Perspec- tive

    Y . Li and Y . Li, “AI load dynamics–a power electronics perspective,”arXiv, no. arXiv:2502.01647, 2025. doi: 10.48550/arXiv.2502.01647

    work page doi:10.48550/arxiv.2502.01647 2025

  45. [45]

    In: Proceedings of the 29th Symposium on Operating Systems Principles

    W. Kwon, Z. Li, S. Zhuang, Y . Sheng, L. Zheng, C. H. Yu, J. Gonzalez, H. Zhang, and I. Stoica, “Efficient memory management for large language model serving with pagedattention,” inProceedings of the 29th Symposium on Operating Systems Prin- ciples, SOSP ’23, (New York, NY , USA), p. 611–626, Association for Computing Machinery, 2023. doi: 10.1145/360000...

    work page doi:10.1145/3600006.3613165 2023

  46. [46]

    Scherfke and O

    S. Scherfke and O. Lünsdorf, “SimPy,” 2023. url: https://simpy.readthedocs.io/en/latest/index.html

    2023

  47. [47]

    Nvidia h100 tensor core gpu - datasheet,

    NVIDIA, “Nvidia h100 tensor core gpu - datasheet,” N.D. url: https://www.nvidia.com/en-us/data-center/h100/, (Accessed: 2026-04-06)

    2026

  48. [48]

    4th generation amd epyc™processors,

    AMD, “4th generation amd epyc™processors,” N.D. url: https://www.amd.com/en/products/processors/server/epyc/4th-generation- 9004-and-8004-series.html, (Accessed: 2026-04-06)

    2026

  49. [49]

    Zero: Memory optimizations toward training trillion parameter models,

    S. Rajbhandari, J. Rasley, O. Ruwase, and Y . He, “Zero: Memory optimizations toward training trillion parameter models,” inSC20: international conference for high performance computing, networking, storage and analysis, pp. 1–16, IEEE,

  50. [50]

    doi: 10.1109/SC41405.2020.00024

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1109/sc41405.2020.00024 2020

  51. [51]

    Instructcoder: Instruction tuning large language models for code editing,

    K. Li, Q. Hu, J. Zhao, H. Chen, Y . Xie, T. Liu, M. Shieh, and J. He, “Instructcoder: Instruction tuning large language models for code editing,” in Proceedings of the 62nd Annual Meeting of the Association for Computa- tional Linguistics (Volume 4: Student Research Workshop), pp. 50–70, 2024. url: https://huggingface.co/datasets/likaixin/InstructCoder

    2024

  52. [52]

    mlperf-inference-llama2-data

    mgoin, “mlperf-inference-llama2-data.” url: https://huggingface.co/datasets/mgoin/mlperf- inference-llama2-data, (Accessed: 2026-04-06)

    2026

  53. [53]

    gpu-burn

    V . Timonen, “gpu-burn.” url: https://github.com/wilicc/gpu-burn, (Accessed: 2026- 04-06). 42

    2026

  54. [54]

    Insight Gained from Migrating a Machine Learning Model to Intelligence Processing Units,

    M. Bachstein, C. Gropp, V . Hazlewood, and G. Peterson, “Lessons from bench- marking the ai tennessee initiative resources,” inPractice and Experience in Advanced Research Computing 2024: Human Powered Computing, PEARC ’24, (New York, NY , USA), Association for Computing Machinery, 2024. doi: 10.1145/3626203.3670524. 43

    work page doi:10.1145/3626203.3670524 2024