Integrating Energy-Efficient Computing Research to Accelerate Energy Technology

Aaron Andersen; Charles Tripp; David Sickinger; Kristin Munch; Michael James Martin

REVIEW 3 major objections 5 minor 3 references

Integrating Energy-Efficient Computing Research to Accelerate Energy Technology

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper argues that pairing an energy-research supercomputer with a research program in energy-efficient computing creates unique opportunities to cut the environmental cost of computation.

desk verdict A clear institutional review of NREL's HPC portfolio and efficiency metrics, but the 'unique opportunities' claim is a programmatic assertion, not a demonstrated result. read the letter →

arxiv 2412.12355 v2 pith:VKCK5Z34 submitted 2024-12-16 cs.CY

Michael James Martin , Aaron Andersen , Charles Tripp , David Sickinger , Kristin Munch This is my paper · ORCID

classification cs.CY

keywords energy-efficientcomputinghigh-performancedatacenterefficiencypowerusageeffectivenessrenewableenergyresearchEEREportfoliolivinglaboratoryworkloadmeasurement

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 argues that a supercomputing center dedicated to energy research can double as a living laboratory for making computation itself more efficient, and that this combination produces opportunities siloed programs cannot. It documents ten years in which computing use by U.S. Department of Energy energy-efficiency and renewable-energy programs grew by a factor of 30, from 89 million to a projected 2.73 billion equivalent Eagle core-hours. It then shows how the same facility that runs energy simulations has achieved data-center efficiency far below industry average, using warm-water cooling and heat reuse, and is beginning to measure the energy cost of individual workloads. The payoff of the integrated model is that domain scientists, facility operators, and algorithm researchers work on the same machines, so lessons about energy-efficient computing flow directly back into the energy research that drives demand.

What carries the argument

The central object is the integrated 'living laboratory' coupling: an HPC data center whose design and instrumentation are themselves research subjects. The metrics that carry the quantitative argument are PUE (facility plus IT energy divided by IT energy), ERE (which credits waste-heat reuse), and WUE (water use per unit of IT energy); the paper uses decade-long PUE comparisons against industry surveys, water savings from a thermosyphon, and node-level energy measurements to close the loop between operating the facility and studying its efficiency.

What would settle it

Compare the actual FY24 and FY25 EERE usage records against the 2.73 billion equivalent Eagle core-hours projection; if realized usage is far lower, the central growth claim is contradicted. Alternatively, run a standard energy-research workload on a Skylake CPU node and on a Kestrel GPU node and check whether the core-hour equivalence assumed in the portfolio analysis is stable across workloads.

Watch

Extended reading notes

Core claim

The paper's central claim is that co-locating the largest HPC capability dedicated to energy research with an explicit research program in energy-efficient computing creates unique, mutually reinforcing opportunities: the facility's Power Usage Effectiveness, Energy Reuse Effectiveness, and Water Usage Effectiveness metrics show that data-center auxiliary energy can be cut by 90-95 percent relative to current industry practice, while node-level energy measurement lets researchers make energy-conscious choices in algorithms and job scheduling. On this view, efficiency is not a one-time design feature but an ongoing research capability, and the same experts who use HPC for energy science can contribute their domain knowledge to cutting computing's footprint.

Load-bearing premise

The growth and portfolio numbers depend on treating projected fiscal-year 2024 usage (2.73 billion equivalent Eagle core-hours) as real demand and on converting all systems, including Kestrel's GPUs, into equivalent Eagle core-hours on a Skylake CPU; if allocations shift or that conversion is not representative, the 30-fold growth claim loses its quantitative footing.

Editorial extensions

If this is right

If the integrated model is adopted broadly, data-center auxiliary energy can be reduced to roughly 5-10 percent of total energy, far below common industry practice.
Node-level energy measurement makes it possible to design and dispatch energy-intensive jobs with energy cost as an explicit criterion rather than an afterthought.
The same HPC systems used for energy research can train the AI/ML models that accelerate materials discovery, as in the redox-flow-battery candidate screening example.
EERE computing demand is projected to grow substantially again as more program offices adopt HPC, so efficiency gains must come from hardware, facility, and algorithm levels together.

Reading between the lines

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

The paper's own data imply that if AI/ML training energy continues doubling every four to six months, even a PUE near 1.0 will not keep data-center emissions in check; the binding constraints will become algorithmic efficiency and grid carbon intensity.
The Eagle-core-hour equivalence used to compare CPU and GPU systems could understate GPU-dominated workloads, since a Skylake core-hour is a fixed baseline; a workload-portable benchmark would test whether the 30-fold growth is real or an artifact of the conversion.
The living-laboratory model suggests a testable extension: publish per-workload energy and carbon cost alongside the science results, so energy-efficient computing becomes a visible output of energy research rather than an internal operational metric.

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.

Desk Editor's Note

read the letter

This is a descriptive case study, not a research paper. What it actually delivers is a useful snapshot of NREL's HPC usage growth (roughly 30x in ten years, per their core-hour conversion), a breakdown of how that compute is used across EERE programs, and a summary of their data-center efficiency work (PUE, ERE, WUE, thermosyphon savings). The FY24 use-model split (with AI/ML shares) is new descriptive data. The paper also does a fair job of citing external benchmarks like the Uptime Institute and Green500, and it is transparent that FY24 growth values are based on projected allocations.

The soft spots are real, but they need to be sized correctly. The main problem is the central claim in the abstract and conclusion: that combining an HPC center dedicated to energy research with an energy-efficient-computing research program creates unique opportunities. That claim is asserted, not demonstrated. There is no comparison group, no counterfactual, no pre/post analysis showing that the integration itself—rather than just running a well-funded, well-run data center—produced the reported benefits. The evidence in the 'Model of Energy Efficiency' section consists entirely of facility-level engineering achievements and program involvement, none of which requires the HPC system to be dedicated to energy research. So the uniqueness assertion is logically under-supported. This is a genuine weakness, and the stress-test note lands.

The growth numbers also rest on a conversion of all workloads, including GPU work, to equivalent Eagle core-hours on a Skylake CPU. That normalization is not justified in the text. And the PUE, ERE, and water-savings figures are self-reported facility data, which is normal for this kind of paper but should carry a clearer caveat. The projected FY24 usage is labeled as such, so that part is less problematic.

That said, the paper is not pretending to be a controlled study. It is a programmatic review, an institutional narrative. For that genre, it is coherent, honest about its own projections, and reasonably well written. The 'unique opportunities' claim should be softened to something like 'potential opportunities' and backed by a description of what would need to be measured to show integration matters.

Who gets value from this? People working on HPC operations, data-center efficiency, or DOE energy-technology programs. It is niche but informative. I would not call it a breakthrough, but it deserves a serious referee: a practice-oriented venue could publish it with minor revisions that hedge the integration claim and add a limitations paragraph. I'd accept it for peer review.

Referee Report

3 major / 5 minor

Summary. The paper, authored by NREL researchers, describes the Computational Sciences Center's role as the HPC provider for DOE's Office of Energy Efficiency and Renewable Energy (EERE). It reports a 30-fold growth in EERE HPC usage from FY14 to a projected FY24, analyzes the research portfolio by discipline (materials science, CFD, integrated energy systems, forecasting, manufacturing), and presents NREL's data-center efficiency record using PUE, ERE, and WUE metrics, including warm-water cooling and a thermosyphon retrofit. The paper then outlines a research agenda in energy-efficient computing and argues that combining an HPC system dedicated to energy research with a program in energy-efficient computing creates unique opportunities for accelerating data-center efficiency improvements that are 'not achievable with siloed initiatives' (Conclusion).

Significance. If the claims are accepted, the paper provides a useful, well-documented case study of a major HPC center dedicated to energy research, and its portfolio breakdown is a valuable reference for how HPC is used across EERE programs. The paper's strengths include a granular accounting of computing use by technology area, concrete illustrative use cases (redox-flow-battery ML screening, ExaWind CFD, PR100, golden-eagle modeling, kraft pulping), and the use of external benchmarks (Green500, Uptime Institute survey) to contextualize NREL's efficiency metrics. The three-strategy framing for reducing computing's environmental impact is clear and sensible. However, the paper's central claim about the uniqueness of the integrated model is asserted rather than demonstrated, and the headline 30-fold growth number depends on a core-hour normalization and projected allocations that are not fully justified. These issues are substantive for the paper's main argument, but they are addressable with clarifications and a more measured framing, so the contributions remain salvageable for a perspective or program-report venue.

major comments (3)

[Growth of computing capabilities and the EERE computing portfolio / Figure 1] The 30-fold growth claim in the Abstract and §2 rests on converting all HPC usage to 'equivalent Eagle core-hours' on an Intel Xeon Gold Skylake 6154 processor, but the paper gives no methodology for this conversion across radically different architectures (Swift's AMD EPYC and A100 GPUs, Kestrel's Sapphire Rapids and H100 GPUs). Without a justified normalization, the quantitative growth numbers—especially the projected 2.73 billion core-hours for FY24—are not reproducible or interpretable. The paper should state the conversion rule, cite a basis for it, or present separate CPU and GPU growth series.
[NREL Computing as a Model of Energy Efficiency / Conclusion] The central claim that the integrated model creates 'unique opportunities' and achieves outcomes 'not achievable with siloed initiatives' (Conclusion) is not supported by the evidence. The reported achievements—low PUE, warm-water cooling, thermosyphon water savings, COOLERCHIPS participation—are data-center engineering projects that do not obviously depend on the HPC system being dedicated to energy research. The paper provides no comparison group, counterfactual, or pre/post analysis that isolates the effect of the integration itself. Either present such evidence or reframe the claim as an argued opportunity rather than an established conclusion.
[Beyond Hardware: Energy-Efficient Computing Research at NREL] The 'Beyond Hardware' section describes the energy-efficient-computing research program entirely in future or enabling terms ('will enable', 'allows researchers to characterize', 'NREL is leveraging'), with no completed demonstrations or empirical results showing that the integration has produced new efficiency insights. Since this program is one of the two pillars of the claimed integration, the paper should either include preliminary results or explicitly label this as a research agenda and separate the agenda from the established facility-efficiency record.

minor comments (5)

[Figure 1 caption] The caption states that 'All values except FY25 are based on actual usage' and that 'FY25 values are projected', but the text (Section 2) reports a projected FY24 value of 2.73 billion core-hours. Please reconcile the caption with the text so the reader knows which year is projected.
[Table 1] There are typos in the hardware names: 'Xenon' should be 'Xeon', 'Sappphire' should be 'Sapphire', and 'PCle' should be 'PCIe'. These should be corrected for professionalism.
[NREL Computing as a Model of Energy Efficiency] The sentence 'This financial and carbon benefits of this approach can be quantified' contains a subject-verb agreement error ('This financial and carbon benefits' should be 'The financial and carbon benefits').
[Figure 7 and NREL facility metrics] The PUE and water-savings figures are self-reported by NREL with no independent audit or uncertainty quantification. Please add a caveat about the self-reported nature of these metrics and, if possible, provide error bars or confidence intervals for the PUE comparison.
[Beyond Hardware / Figure 8-9] Figures 8 and 9 would benefit from a data and code availability statement rather than only 'Data available upon request'. This is especially important because the AI training-energy estimates are derived from a specific methodology that is not fully described in the text.

Circularity Check

0 steps flagged · score 2.0 of 10

No reducing circularity: the growth and efficiency claims rest on standard metrics, internal usage accounting, and external benchmarks; the sole self-citation is minor and not load-bearing.

full rationale

The paper contains no derived equations whose outputs are wired into their inputs. PUE, ERE, and WUE are defined from Green Grid sources [2], and the efficiency comparisons use external benchmarks (Uptime Institute [18], Green500 [17]). The 30-fold EERE usage growth is an accounting of used and allocated core-hours converted to Eagle-equivalent core-hours, not a fitted prediction; the FY24/FY25 figures are explicitly described as allocation-based projections. The central "unique opportunities" claim is a programmatic assertion rather than a result derived from a model, so it is unsupported or speculative but not circular. The only notable self-citation is [19], an NREL technical report authored by co-author David Sickinger, used to document thermosyphon water savings; this supports one facility engineering metric and is not load-bearing for the paper's central integration argument. No self-citation chain is used to forbid alternatives or to justify the main premise. Therefore the circularity burden is low; score 2 reflects the one minor non-load-bearing self-citation rather than any reducing derivation.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities. Its central narrative depends on accounting conventions (core-hour equivalence), trust in self-reported categories, and projected allocations. The metrics PUE, ERE, and WUE are standard definitions, not free parameters.

assumptions (4)

domain assumption Core-hour equivalence on an Intel Xeon Gold Skylake 6154 processor is a valid common measure for comparing HPC usage across heterogeneous systems including GPU nodes.
Used throughout Figure 1 and the text to compute growth factors; no justification for GPU conversion is provided.
domain assumption Self-identified use-model categories and AI/ML usage shares accurately reflect the research portfolio.
The FY24 percentages in Figure 2 and the AI/ML shares depend on users' self-identification, with no validation described.
domain assumption Projected FY24 and FY25 allocations reflect actual future usage.
The 30-fold growth claim relies on FY24 projections (expected to use), not measured usage, and on partial-year availability of Kestrel GPU nodes.
domain assumption PUE as defined in Eq. (1) is an appropriate primary metric for data center efficiency.
The paper's claim of world-leading efficiency rests on PUE, an industry standard, but one that the paper itself notes may be superseded by metrics like TUE.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Integrating Energy-Efficient Computing Research to Accelerate Energy Technology." pith.science (2026). https://pith.science/paper/VKCK5Z34

@misc{pith2026241212355,
  author       = {Pith},
  title        = {Pith review of: Integrating Energy-Efficient Computing Research to Accelerate Energy Technology},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VKCK5Z34}},
  note         = {Machine review of arXiv:2412.12355}
}

read the original abstract

NREL's computational sciences center hosts the largest high-performance computing (HPC) capabilities dedicated to energy research while functioning as a living laboratory for energy-efficient computing. NREL's HPC capabilities support the research needs of the Department of Energy's Office of Energy Efficiency and Renewable Energy (EERE). In ten years of operation, HPC use in EERE-sponsored research has grown by a factor of 30, including work in electricity generation, energy efficiency, transportation, and energy system modeling. This paper analyzes this research portfolio, providing examples of individual use cases. The paper documents NREL's history of operating one of the world's most energy-efficient data centers while examining pathways to reduce economic and environmental impact beyond reduction of Power Usage Efficiency (PUE). This paper concludes by examining the unique opportunities created for accelerating improvements in data center efficiency created by combining an HPC system dedicated to energy research and a research program in energy-efficient computing.

Figures

Figures reproduced from arXiv: 2412.12355 by the authors.

**Figure 1.** HPC Usage by EERE Technology Area/Office by Year. All values are given using the equivalent of one core hour on an Intel Xeon Gold Skylake 6154 3.0 GHz processor. All values except FY25 are based on actual usage. FY25 values are projected based on FY25 allocations that take into account the addition of expanded GPU capability on Kestrel. Use models for HPC in energy research The DOE’s 2014 Quadrennial Technology Rev… view at source ↗

**Figure 2.** Percentage of FY24 HPC Usage by Use Model. Chemistry and Materials Science is critical to the development and deployment of multiple technologies. The ability to perform large numbers of in silico experiments using validated computational methods becomes particularly valuable when combined with machine learning (ML) and artificial intelligence (AI) to accelerate discovery [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Combined computational science and machine learning work flow for identification of candidate materials for aqueous redox flow batteries. [4] Beginning with a set of nearly 100,000 HPC-enabled quantum chemistry simulations, researchers were able to train ML-based surrogate objective functions. These were used to identify molecules with the correct combination of electron transfer characteristics, stability, and synt… view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Vor#city field (lef) and drag versus #me (right) for a 6-row Concentrated Solar Power (CSP) parabolic solar collector array. [7] Integrated Energy Systems simulations allow the impact of changes in the physical systems that generate or use energy, to be quantified as p…

**Figure 5.** Figure 5: shows a representative MD simulation result. These atomistic simulations provided molecular-level insight towards improving the pretreatment for maximum polysaccharide retention and thus increased yields [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Overlay of golden eagle rela#ve presence density in a 50 km x 50 km region in Wyoming during the northbound migra#on in spring with exis#ng wind turbine loca#ons.[13] An overview of current projects shows that HPC projects are increasingly integra-ve, bringing together…

**Figure 7.** Figure 7: NREL HPC Data center PUE compared to the average PUE reported from 2013-2024 in the Uptime Institutes annual survey . Typical survey sample size includes 500-600 respondents (Data available upon request) The next step is to optimize the ERE by utilize the heat for purp…

**Figure 8.** Figure 8: Estimated training energy of state-of-the-art published AI models over time using the best available hardware at the time of publication.(Data available upon request) Data center energy use is correlated with CO2 usage, which has led to concerns in the scientific commu…

**Figure 9.** Figure 9: Es#mated CO2 emissions from training state-of-the-art published AI/ML models over #me using the best available hardware and typical U.S. PUE, transmission losses, and CO2/kWH. (Data available upon request) NREL is leveraging its role as a data center operator to establ…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 2 canonical work pages

[2]

PUE™: A Comprehensive Examina-on of the Metric,

V. Avelar, et. al, “PUE™: A Comprehensive Examina-on of the Metric,” Green Grid, White Paper-49, 2012, hpps://www.thegreengrid.org/en/resources/library-and-tools/20-PUE%3A-A-Comprehensive-Examina-on-of-the-Metric [3] US Department of Energy, Quadrennial Technology Review: An Assessment of Energy Technologies and Research Opportuni#es. Washington, DC, USA:...

work page arXiv 2012
[15]

Red AI Era

Finally, the manufacturing of data centers also has an energy and environmental impact that may be reduced by careful design. [16] This creates a need to create effective approaches to reducing energy use in computing. Three strategies exist for avoiding large increases energy use even as computational demands increase: • increasing efficiency of computat...

work page 2013
[16]

Advancing Environmental Sustainability in Data Centers by Proposing Carbon Depreciation Models

S. Ji, et al., "Advancing Environmental Sustainability in Data Centers by Proposing Carbon Depreciation Models", arXiv:2403.04976. 2024. [17] Green 500 List, hpps://www.top500.org/lists/green500/, accessed December 12, 2023. [18] D. Donnellan, et al., "Uptime Institute Global Data Center Survey Results 2024", https://uptimeinstitute.com/resources/research...

arXiv 2024

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [2]

PUE™: A Comprehensive Examina-on of the Metric,

V. Avelar, et. al, “PUE™: A Comprehensive Examina-on of the Metric,” Green Grid, White Paper-49, 2012, hpps://www.thegreengrid.org/en/resources/library-and-tools/20-PUE%3A-A-Comprehensive-Examina-on-of-the-Metric [3] US Department of Energy, Quadrennial Technology Review: An Assessment of Energy Technologies and Research Opportuni#es. Washington, DC, USA:...

work page arXiv 2012

[2] [15]

Red AI Era

Finally, the manufacturing of data centers also has an energy and environmental impact that may be reduced by careful design. [16] This creates a need to create effective approaches to reducing energy use in computing. Three strategies exist for avoiding large increases energy use even as computational demands increase: • increasing efficiency of computat...

work page 2013

[3] [16]

Advancing Environmental Sustainability in Data Centers by Proposing Carbon Depreciation Models

S. Ji, et al., "Advancing Environmental Sustainability in Data Centers by Proposing Carbon Depreciation Models", arXiv:2403.04976. 2024. [17] Green 500 List, hpps://www.top500.org/lists/green500/, accessed December 12, 2023. [18] D. Donnellan, et al., "Uptime Institute Global Data Center Survey Results 2024", https://uptimeinstitute.com/resources/research...

arXiv 2024