Pith. sign in

REVIEW 3 major objections 4 minor 19 references

Power-Capping Metric Evaluation for Improving Energy Efficiency in HPC Applications

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Per-GPU-kernel power caps chosen by energy-runtime metrics can cut a scientific application's energy use more sharply than one chip-wide cap.

desk verdict Genuinely useful GH200 per-kernel power data, but the 151–200% savings headline is an arithmetic artifact of summing per-task percentages. read the letter →

arxiv 2505.21758 v2 pith:CMAJAYCY submitted 2025-05-27 cs.DC cs.CEcs.PFcs.SYeess.SY

Maria Patrou , Thomas Wang , Wael Elwasif , Markus Eisenbach , Ross Miller , William Godoy , Oscar Hernandez This is my paper · ORCID
classification cs.DCcs.CEcs.PFcs.SYeess.SY
keywords powercappingenergyefficiencyHPCGH200performancemetricsautomaticsteeringexascaleapplicationsLSMS
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

This paper tries to show that the way a power cap is chosen matters as much as the cap itself. Running the LSMS materials-science code on a GH200 superchip, the authors measure energy and runtime separately for each GPU kernel and for the phases when the GPU sits idle. They then use two decision rules—the speedup-energy-delay product and the Euclidean distance between normalized energy and runtime—to pick the best chip-level power limit for each task. Their central claim is that applying these per-task caps would save considerably more energy than a single application-wide cap, because compute-bound kernels need high power while memory-bound kernels and idle phases can be capped low at little runtime cost. They report roughly 151% energy savings at a 90% runtime increase for one metric and roughly 200% energy reduction at a 203% runtime increase for the other.

What carries the argument

The load-bearing machinery is a per-task scoring rule evaluated across a grid of power-cap settings. For each GPU task and each cap level, the paper computes either the speedup-energy-delay ratio $r_i = \frac{\mathrm{runtime}_1 \cdot \mathrm{energy}_1}{\mathrm{runtime}_i \cdot \mathrm{energy}_i}$ (higher is better) or the Euclidean distance $d_i = \sqrt{n_{\mathrm{energy},i}^2 + n_{\mathrm{runtime},i}^2}$, where each $n$ term is the min-max normalized energy or runtime of that task across all caps (lower is better). The recommended cap is the one that optimizes the chosen score, and the GH200's automatic power steering reallocates unused power between CPU and GPU, letting the chip-level cap act as a task-specific constraint.

What would settle it

A reader can falsify the headline totals by re-aggregating the paper's own Table 2 values: weight each task's energy reduction by its share of total baseline energy and each runtime increase by its share of baseline runtime. That calculation gives roughly 11% energy savings at about 6% runtime increase, far below the reported 151%/90% and 200%/203%; an experiment that applies the proposed per-task cap schedule on the same GH200 node and records superchip-level energy would show which aggregation matches reality.

Watch

Extended reading notes

Core claim

The paper's central finding is that power-capping decisions should be made per GPU task, not once for the whole application. Profiling the LSMS run on a GH200 node, the authors separate the dominant compute-bound dense matrix-multiply kernel, a memory-bound matrix-construction kernel, several LU-factorization and triangular-solve kernels, and the GPU compute idle phases between self-consistent-field iterations. At nine chip-level power caps from 200 W to the 1000 W default, the speedup-energy-delay rule peaks at 300–600 W for most tasks and at 900 W for the largest dense matrix-multiply kernel, while the Euclidean-distance rule settles most tasks at 400 W and the memory-bound kernel at 300 W. The authors conclude that kernel-level capping—for example, holding the GPU idle phase at 200–300 W—yields greater energy savings than application-wide tuning, with the exact trade-off depending on which metric the user trusts.

Load-bearing premise

The load-bearing premise is that per-GPU-task percentage changes in energy and runtime can be summed to estimate whole-application savings, and that switching each task to its optimal cap costs no additional time or energy.

Editorial extensions

If this is right

  • Applying the metric-selected caps task by task would give greater energy savings than any single application-wide cap, with the exact gain depending on the chosen metric.
  • GPU compute idle phases are the cheapest energy to reclaim: the measurements show roughly 40–47% energy reduction at 1–9% runtime increase.
  • Memory-bound kernels tolerate low caps well, saving about 23% energy for an 11% runtime increase, while compute-bound dense matrix-multiply kernels save little energy but keep runtime essentially unchanged.
  • The method needs only per-task energy and runtime measurements, so the same metric-based cap selection can be applied to other GPU-accelerated HPC applications without changing their source code.

Reading between the lines

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

  • A correct energy-weighted aggregation of the paper's own per-task numbers yields roughly 11% energy savings at about 6% runtime increase, so the reported 151% and 200% figures are best read as indicators of per-task headroom rather than physical savings.
  • The two metrics' divergent recommendations imply a natural 'policy dial': speedup-energy-delay favors preserving runtime, while Euclidean distance favors deeper energy cuts, and a production system could expose this choice per job.
  • The same offline analysis could be turned into an online controller that computes the Euclidean distance from rolling energy and runtime samples and adjusts the chip-level cap during kernel execution.
  • The approach should transfer to other CPU-GPU superchips and to idle periods caused by communication, where capping the accelerator during host-side work is nearly free.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper reports an empirical study of power-capping on an NVIDIA GH200 superchip for the LSMS exascale application. It profiles individual GPU kernels and GPU-idle phases under global superchip power caps from 200 W to 1000 W, evaluates two decision metrics (speedup-energy-delay and Euclidean distance of normalized energy/runtime), selects a per-task 'optimal' power cap for each metric, and reports percentage energy reductions and runtime increases relative to the 1000 W baseline. The central quantitative claim is that aggregating per-task percentage changes yields roughly 151% energy savings at about 90% runtime increase for speedup-energy-delay, and roughly 200% energy reduction at about 203% runtime increase for the Euclidean metric. The paper concludes that kernel-level power-capping can yield greater energy savings than application-wide tuning.

Significance. If the headline result were correct, this would be a practically important demonstration that fine-grained, per-GPU-task power capping on integrated CPU-GPU architectures can produce very large energy savings. The measurement methodology is a strength: the authors use Score-P, PAPI, and NVML to collect 5 ms power samples at GPU-task granularity, and the qualitative separation between compute-bound and memory-bound kernels is a useful characterization. However, the central quantitative claim is arithmetically invalid: the reported percentages are obtained by summing per-task percentage changes rather than weighting by each task's energy, and the resulting 'savings' exceed 100% of the total energy. Recomputing with Table 1 baseline energies gives roughly 11-14% energy savings on the profiled GPU-task energy. The contribution is therefore a preliminary, qualitatively informative case study, not the large efficiency gain claimed in the abstract and conclusions.

major comments (3)
  1. [Section 4, Table 2] The headline '~151% energy savings' and '~200% energy reduction' are unweighted sums of the per-task percentage changes in Table 2. For example, the SED energy column sums to 0.85+22.92+31.40+20.61+10.05+9.02+9.10+46.58 = 150.53, and the ED column sums to about 200.09. Because each percentage is relative to that task's own baseline energy, the percentages are not commensurable and cannot be added across tasks; total energy is the sum of task energies, so savings can never exceed 100% of that total. Using the baseline energies in Table 1, the correctly weighted savings are about 6.8 kJ out of 63.6 kJ of profiled GPU-task energy (~11%) with about a 6% weighted runtime increase for SED, and about 8.7 kJ (~14%) with about a 14% weighted runtime increase for ED; as fractions of the reported 93,574.7 J total, these are even smaller (~7-9%). The sentence 'These numbers are simple aggregations ... ideal scenarios with no overhead' acknowledges an idealization but does not address the weighting error, which persists at zero overhead.
  2. [Section 3.2 and Table 2] The 'optimal' power cap for each task is selected by minimizing or maximizing the chosen metric over the same measured power-cap sweep, and Table 2 then reports the energy/runtime difference between that selected setting and the 1000 W baseline within the same dataset. This is an in-sample selection of the optimum, so the reported percentage changes are best-case values for the measured data rather than expected savings under deployment. With only three runs per setting and no confidence intervals, cross-validation, or a holdout run at the selected caps, the claim in Section 4 that these results 'can be achieved' is overstated; the paper should either reframe the results as in-sample best-case observations or add validation to support predictive claims.
  3. [Section 4 and Section 5] The conclusion that 'kernel-level power-capping can yield greater energy savings than application-wide tuning' is not supported by the reported experiments. The paper does not compare against any application-wide power-capping tuning experiment, and the experiments apply a single global power cap for an entire run (Section 3 states 'We executed the entire application at every superchip power limit'), not dynamic per-task capping during a single execution. The claimed superiority of fine-grained capping is therefore an extrapolation, not a measured result. A direct comparison, or at minimum a clear statement that this is a hypothesis for future work, is needed.
minor comments (4)
  1. [References] Reference [1] contains a formatting error: 'In: 20Haidar,15 Sixth International' should be 'In: 2015 Sixth International Green and Sustainable Computing Conference'.
  2. [Section 3.2] The text says min-max normalization 'yielded values in the interval (0,1)', but the minimum and maximum values can be attained, so the interval should be [0,1].
  3. [Table 2] The column headers 'Energy (%)' and 'Runtime (%)' are ambiguous; the caption should state explicitly that the energy values are reductions (positive percentages) and the runtime values are increases.
  4. [Abstract and Section 1] The phrase 'dynamic power-cap adjustments' suggests that power caps were changed during execution, but the experiments apply a static cap for the whole run; consider rewording to 'potential dynamic adjustments' or describing the planned implementation more precisely.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the metric definitions, power-cap selections, and per-task measurements are not self-referential; the headline aggregation error is an arithmetic and interpretation problem, not a circular derivation.

full rationale

The paper's derivation chain is transparent: Section 3.2 defines the speedup-energy-delay and Euclidean-distance metrics; the optimal power cap for each GPU task is selected by minimizing or maximizing that metric over the measured 200–1,000 W sweep; and Table 2 reports the measured energy and runtime changes at those selected settings. Nothing in this chain fits a parameter to a subset of data and then predicts the same subset, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the choice. The self-citations [9,14,15] are methodological: the Euclidean-distance metric is explicitly adapted from prior CPU DVFS work, but its definition is given in full and does not presuppose any of the paper's conclusions. The 'optimal' caps are in-sample selections, so the associated savings are observed outcomes rather than out-of-sample predictions; that is a generalization limitation, not circularity. The central quantitative claim in Section 4 is indeed invalid: summing per-task percentage changes, each relative to that task's own baseline, cannot yield a whole-application energy saving, and a correct energy-weighted aggregation gives roughly 11% savings, not 151% or 200%. However, this is an arithmetic and interpretational error, not a reduction of a claimed result to its own inputs by construction. The paper even notes that the numbers are 'simple aggregations of per–GPU task energy and runtime values and represent ideal scenarios with no overhead,' which flags but does not repair the weighting problem. Because the circularity criteria are about self-definition, fitted-then-predicted values, and load-bearing self-citations, and none of those are present here, the appropriate finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities or fitted model constants. The only hand-chosen number list is the power-cap sweep, which bounds the reported optima. The central claim depends on hardware assumptions (GH200 power-steering behavior, measurement fidelity), profiling assumptions (per-kernel energy attribution), and the questionable assumption that per-task percentage savings can be summed into whole-application figures. The last assumption is ad hoc to this paper and breaks the headline result.

free parameters (1)
  • Power-cap sweep grid = 200,300,400,500,600,700,800,900,1000 W
    The set of tested power caps is chosen by hand. The reported optimal settings are restricted to this grid, so the magnitude of the best-case savings depends on this experimental choice.
assumptions (5)
  • domain assumption The GH200 automatic power-steering system reallocates power between CPU and GPU as described, and the hwmon/NVML measurements accurately reflect chip and GPU power.
    Section 2 and Section 3 rely on the documented power-steering behavior and on Score-P/PAPI sampling; no calibration against a reference power meter is provided.
  • domain assumption Measured energy can be attributed to individual GPU kernels through Score-P tracing with 5 ms sampling.
    Per-kernel energy values in Table 1 depend on associating power samples with kernel execution intervals; sampling granularity and attribution errors are not analyzed.
  • domain assumption The default 1,000 W power cap is the appropriate baseline for comparison.
    All savings percentages are computed relative to the default setting in Section 3.1; a different baseline would change all reported improvements.
  • domain assumption The two metrics (speedup-energy-delay and Euclidean distance of min-max normalized values) correctly identify the optimal power cap for each task.
    Section 3.2 assumes the Pareto-optimal interpretation of the Euclidean distance and the EDP-like behavior of speedup-energy-delay apply to per-kernel power-capping decisions.
  • ad hoc to paper Per-GPU-task percentage energy and runtime changes can be aggregated by summation to estimate whole-application impact with no overhead.
    Section 4 sums the percentage values from Table 2 to produce '~151%' and '~200%' savings; the paper provides no justification for linear summation, which is arithmetically invalid for percentages across tasks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Power-Capping Metric Evaluation for Improving Energy Efficiency in HPC Applications." pith.science (2026). https://pith.science/paper/CMAJAYCY

@misc{pith2026250521758,
  author       = {Pith},
  title        = {Pith review of: Power-Capping Metric Evaluation for Improving Energy Efficiency in HPC Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CMAJAYCY}},
  note         = {Machine review of arXiv:2505.21758}
}
read the original abstract

With high-performance computing systems now running at exascale, optimizing power-scaling management and resource utilization has become more critical than ever. This paper explores runtime power-capping optimizations that leverage integrated CPU-GPU power management on architectures like the NVIDIA GH200 superchip. We evaluate energy-performance metrics that account for simultaneous CPU and GPU power-capping effects by using two complementary approaches: speedup-energy-delay and a Euclidean distance-based multi-objective optimization method. By targeting a mostly compute-bound exascale science application, the Locally Self-Consistent Multiple Scattering (LSMS), we explore challenging scenarios to identify potential opportunities for energy savings in exascale applications, and we recognize that even modest reductions in energy consumption can have significant overall impacts. Our results highlight how GPU task-specific dynamic power-cap adjustments combined with integrated CPU-GPU power steering can improve the energy utilization of certain GPU tasks, thereby laying the groundwork for future adaptive optimization strategies.

Figures

Figures reproduced from arXiv: 2505.21758 by the authors.

Figure 1
Figure 1. Power and total energy consumed by LSMS on the GH200 superchip (SC), including CPU and GPU components, with the default maximum power setting of 1,000W (no power capping). We executed LSMS on the GH200 superchip by using the default power limit of 1,000W (no power capping) as a baseline scenario. We measured ker￾nel execution times along with power and total energy consumption for both CPU and GPU. The total runtime… view at source ↗
Figure 2
Figure 2. Speedup-energy-delay per GPU task and power-cap setting. Higher is better [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Euclidean distance of normalized energy/runtime per GPU task and power-cap setting. Lower is better [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 14 canonical work pages

  1. [1]

    In: 20Haidar,15 Sixth In- ternational Green and Sustainable Computing Conference (IGSC)

    Abdulsalam, S., Zong, Z., Gu, Q., Qiu, M.: Using the greenup, powerup, and speedup metrics to evaluate software energy efficiency. In: 20Haidar,15 Sixth In- ternational Green and Sustainable Computing Conference (IGSC). pp. 1–8 (2015). https://doi.org/10.1109/IGCC.2015.7393699

    work page arXiv 2015

  2. [2]

    In: Proceed- ings of the International Conference for High Performance Computing, Net- working, Storage and Analysis

    Atchley, S., Zimmer, C., et al.: Frontier: Exploring exascale. In: Proceed- ings of the International Conference for High Performance Computing, Net- working, Storage and Analysis. SC ’23, Association for Computing Machin- ery, New York, NY, USA (2023). https://doi.org/10.1145/3581784.3607089, https://doi.org/10.1145/3581784.3607089

    arXiv 2023

  3. [3]

    In: 2018 Ninth International Green and Sustainable Computing Conference (IGSC)

    Azimi, R., Jing, C., Reda, S.: Powercoord: A coordinated power capping controller for multi-cpu/gpu servers. In: 2018 Ninth International Green and Sustainable Computing Conference (IGSC). pp. 1–9. IEEE (2018)

    work page 2018

  4. [4]

    https://docs.nvidia.com/grace-perf-tuning-guide/power-thermals.html (2024), ac- cessed: 2025-02-28

    Corporation, N.: Nvidia grace performance tuning guide - power and thermals. https://docs.nvidia.com/grace-perf-tuning-guide/power-thermals.html (2024), ac- cessed: 2025-02-28

    work page 2024

  5. [5]

    In: Kunkel, J.M., Yokota, R., Balaji, P., Keyes, D

    Eastep, J., Sylvester, S., Cantalupo, C., Geltz, B., Ardanaz, F., Al-Rawi, A., Liv- ingston, K., Keceli, F., Maiterth, M., Jana, S.: Global extensible open power man- ager: A vehicle for hpc community collaboration on co-designed energy manage- ment solutions. In: Kunkel, J.M., Yokota, R., Balaji, P., Keyes, D. (eds.) High Performance Computing. pp. 394–4...

    work page 2017

  6. [6]

    Computer Physics Commu- nications211, 2–7 (2017), high Performance Computing for Advanced Modeling and Simulation of Materials

    Eisenbach, M., Larkin, J., Lutjens, J., Rennich, S., Rogers, J.H.: Gpu acceleration of the locally selfconsistent multiple scattering code for first principles calculation of the ground state and statistical physics of materials. Computer Physics Commu- nications211, 2–7 (2017), high Performance Computing for Advanced Modeling and Simulation of Materials

    work page 2017

  7. [7]

    In: Proceed- ings of the ACM Workshop on Software Engineering Methods for Parallel and High Performance Applications

    Gerndt, M.: The readex project for dynamic energy efficiency tuning. In: Proceed- ings of the ACM Workshop on Software Engineering Methods for Parallel and High Performance Applications. pp. 11–12 (2016)

    work page 2016

  8. [8]

    Concurrency and Computation: Practice and Experience31(6), e4485 (2019)

    Haidar, A., Jagode, H., Vaccaro, P., YarKhan, A., Tomov, S., Dongarra, J.: Inves- tigating power capping toward energy-efficient scientific applications. Concurrency and Computation: Practice and Experience31(6), e4485 (2019)

    work page 2019

Show all 19 references
  1. [9]

    In: Proceedings of the 2025 International Conference on High Performance Computing in Asia-Pacific Region Workshops

    Hernandez, O., Wang, T., Elwasif, W., Spiga, F., Tartaglione, F., Eisenbach, M., Miller, R.: Preliminary study on fine-grained power and energy measurements on grace hopper gh200 with open-source performance tools. In: Proceedings of the 2025 International Conference on High P...

    2025

  2. [10]

    In: Brunst, H., Müller, M.S., Nagel, W.E., Resch, M.M

    Knüpfer, A., Rössel, C., Mey, D.a., Biersdorff, S., Diethelm, K., Eschweiler, D., Geimer, M., Gerndt, M., Lorenz, D., Malony, A., Nagel, W.E., Oleynik, Y., Philip- pen, P., Saviankou, P., Schmidl, D., Shende, S., Tschüter, R., Wagner, M., We- sarg, B., Wolf, F.: Score-p: A joi...

    2012

  3. [11]

    In: Wyrzykowski, R., Deelman, E., Dongarra, J., Karczewski, K

    Krzywaniak, A., Czarnul, P.: Performance/energy aware optimization of parallel applications on GPUs under power capping. In: Wyrzykowski, R., Deelman, E., Dongarra, J., Karczewski, K. (eds.) Parallel Processing and Applied Mathematics. Springer International Publishing (2020)

    2020

  4. [12]

    In: International Conference on Computational Science (ICCS)

    Krzywaniak, A., Czarnul, P., Proficz, J.: Gpu power capping for energy- performance trade-offs in training of deep cnns. In: International Conference on Computational Science (ICCS). pp. 123–133. Springer (2022)

    2022

  5. [13]

    https://developer.nvidia.com/blog/energy-efficiency- in-high-performance-computing-balancing-speed-and-sustainability/ (2023), ac- cessed: 2024-10-23

    NVIDIA: Energy efficiency in high performance computing: Balancing speed and sustainability. https://developer.nvidia.com/blog/energy-efficiency- in-high-performance-computing-balancing-speed-and-sustainability/ (2023), ac- cessed: 2024-10-23

    2023

  6. [14]

    In: 2021 IEEE In- ternational Conference on Cloud Engineering (IC2E)

    Patrou, M., Kent, K.B., Siu, J., Dawson, M.: Energy and runtime performance optimization of node.js web requests. In: 2021 IEEE In- ternational Conference on Cloud Engineering (IC2E). pp. 71–82 (2021). https://doi.org/10.1109/IC2E52221.2021.00021

    2021

  7. [15]

    In: 2022 IEEE 13th Interna- tional Green and Sustainable Computing Conference (IGSC)

    Patrou, M., Kent, K.B., Siu, J., Dawson, M.: Optimizing energy efficiency of node.js applications with cpu dvfs awareness. In: 2022 IEEE 13th Interna- tional Green and Sustainable Computing Conference (IGSC). pp. 1–8 (2022). https://doi.org/10.1109/IGSC55832.2022.9969367

    2022

  8. [16]

    In: Schulz, M., Trinitis, C., Papadopoulou, N., Pionteck, T

    Saba, I., Arima, E., Liu, D., Schulz, M.: Orchestrated co-scheduling, resource par- titioning, and power capping on cpu-gpu heterogeneous systems via machine learn- ing. In: Schulz, M., Trinitis, C., Papadopoulou, N., Pionteck, T. (eds.) Architecture of Computing Systems. pp. ...

    2022

  9. [17]

    In: Proceedings of the International Conference for High Performance Computing, Network- ing, Storage and Analysis

    Sourouri, M., Raknes, E.B., Reissmann, N., Langguth, J., Hackenberg, D., Schöne, R., Kjeldsberg, P.G.: Towards fine-grained dynamic tuning of hpc applications on modern multi-core architectures. In: Proceedings of the International Conference for High Performance Computing, Ne...

    2017

  10. [18]

    In: SC24-W: Workshops of the International Conference for High Per- formance Computing, Networking, Storage and Analysis

    Terboven, C., Liem, R., Gracia, J., Haldar, K., Engels, J., Giesselmann, P., Brayford, D., Wilde, T., Simmendinger, C., Marquardt, M., Eitzinger, J., Gruber, T.: Ee-hpc a framework for energy efficient hpc system manage- ment. In: SC24-W: Workshops of the International Confere...

    2024

  11. [19]

    In: Proceedings of the 2023 ACM Symposium on Cloud Computing

    Zhao, D., Samsi, S., McDonald, J., Li, B., Bestor, D., Jones, M., Tiwari, D., Gade- pally, V.: Sustainable supercomputing for AI: GPU power capping at HPC scale. In: Proceedings of the 2023 ACM Symposium on Cloud Computing. pp. 588–596 (2023)

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.