Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

This paper proposes a unified job-level water–carbon accounting framework for HPC runs and claims that increasing thread count generally lowers total footprint, though benefits diminish; water is mostly embodied while carbon is mostly opera

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

2026-08-01 13:16 UTC pith:S6MASW2L

load-bearing objection A useful job-level accounting framework whose headline thread-scaling result is governed by the node-hour embodied allocation; worth refereeing, but the central claim is not yet established. the 3 major comments →

arxiv 2607.19150 v1 pith:S6MASW2L submitted 2026-07-21 cs.PF

Job-level Carbon and Water Footprint Estimation for HPC: Bias Assessment from Runtime to Full Life Cycle

Xi Chen , Chris Broekema , Rob van Nieuwpoort This is my paper
classification cs.PF
keywords high performance computingcarbon footprintwater footprintembodied impactoperational impactthread scalingjob-level accountingsustainability benchmarking
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.

The paper attempts to establish that runtime configuration choices in HPC, such as thread count, can be compared fairly only when both operational and embodied water and carbon impacts are included. Using measurements from three standard benchmarks, it argues that more threads lower runtime and therefore lower total footprint, but the gain flattens at high thread counts. It also finds that water impact is dominated by manufacturing/embodied share, while carbon impact remains dominated by operational energy use. A sympathetic reader would care because it suggests operational-only sustainability metrics systematically misrank configurations, especially for water.

Core claim

The central claim is that a single job-level accounting boundary, combining operational and embodied water and carbon, changes how HPC runtime configurations should be interpreted. Across STREAM Triad, NPB CG.C, and NPB EP.C, raising the OpenMP thread count from 1 to 36 reduces both total water and total carbon footprint, chiefly because runtime falls; the reduction is strong in the low-to-mid thread range and marginal once performance saturates. The breakdown, however, differs: embodied water is the larger water component in every tested configuration, while operational carbon is the larger carbon component.

What carries the argument

The framework's core is a set of four component equations: total water and carbon each split into operational and embodied terms; operational impacts scale with job energy times intensity factors (water usage effectiveness and grid carbon intensity); embodied impacts are amortized at the node level into a per-node-hour rate, then multiplied by job runtime. This node-hour amortization (Eq. 5) is what connects configuration choices to embodied footprint.

Load-bearing premise

The load-bearing assumption is Eq. (5): each job is charged the full node's embodied water and carbon per hour of runtime, regardless of how many of the 36 cores it actually uses; if that embodied cost were instead allocated per core or per utilized fraction, the reported thread-scaling reductions could shrink or reverse.

What would settle it

Run the same three benchmarks in weak-scaling mode: double the thread count and double the work so runtime stays roughly constant; if total footprint still drops, the benefit is not just shorter runtime. Alternatively, recompute embodied impacts per utilized core instead of per node-hour; if the water footprint no longer falls with thread count, the main conclusion fails.

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

If this is right

  • Configuration tuning for sustainability should include embodied impacts; operational-only accounting underestimates water footprint more than carbon.
  • Thread scaling from 1 to roughly 18 threads yields most of the footprint reduction; beyond saturation, further thread increases give little sustainability benefit.
  • Water and carbon favour different lifecycle stages, so a single aggregate sustainability metric could mislead if it collapses the two.
  • The same framework can be applied to other HPC workloads and systems to identify low-footprint configurations.

Where Pith is reading between the lines

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

  • If embodied water dominates, scheduling decisions that improve node utilization (e.g., co-scheduling multiple jobs) may reduce per-job embodied water more than any energy optimization would, because embodied cost is amortized over more useful work.
  • The framing suggests a testable extension: for shared-node or GPU-equipped systems, the per-job allocation rule becomes decisive; the paper's own future-work note implies the ranking of configurations could change under per-core or co-location allocation.
  • The 'run-to-idle' effect means the measured benefit is partly an artifact of fixed-work benchmarks; for throughput-oriented workloads that keep nodes busy, the thread-count benefit may be much smaller.

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 / 5 minor

Summary. The manuscript proposes a job-level accounting framework for estimating the carbon and water footprints of HPC jobs, combining operational and embodied impacts. Operational terms are computed from measured job energy plus facility and grid intensity factors (Eqs. 2–3); embodied terms are obtained by amortizing node-level manufacturing impacts over hardware lifetime and utilization, then assigning them to jobs in proportion to wall-clock time (Eqs. 4–5). Experiments on DAS-6 with STREAM Triad and NPB CG.C/EP.C across OpenMP thread counts 1–36 are used to claim that higher thread counts generally reduce total footprint, that water is dominated by embodied impact while carbon is dominated by operational impact, and that operational-only accounting can bias configuration-level comparisons, especially for water.

Significance. If established, the paper's central message would be valuable: configuration-level sustainability comparisons in HPC should include embodied impacts, and the choice of allocation rule can materially affect conclusions. The framework is simple, transferable, and grounds the operational side in direct runtime/energy measurements, which is a strength. However, the headline thread-scaling result is not yet established because it depends on an accounting choice in Eq. (5), and the manuscript does not provide the parameter values or sensitivity analysis needed to assess robustness. The paper is also candid in Sec. IV that different allocation rules are left to future work.

major comments (3)
  1. [Eq. (5), Sec. III-B, Sec. IV] The embodied allocation W_emb = r_W^node * t_job charges the full-node hourly embodied rate to every job regardless of how many of the node's 36 cores the job uses. Since water is reported to be dominated by embodied impact in Fig. 2, the monotonic decrease of total water with thread count in Fig. 1(a) is essentially the runtime decrease. Under a per-core or utilization-weighted allocation (e.g., W_emb = r_W^node * (n_threads/36) * t_job), the trend could shrink, flatten, or even reverse. The paper defers this issue to future work in Sec. IV, but this allocation rule is load-bearing for the central claim. A sensitivity analysis over plausible allocation rules is required before the thread-scaling conclusion can be accepted.
  2. [Sec. II, Eqs. (2)–(4)] No numerical values are reported for WUE(T_wb), PUE, EWF, CI_grid, W_emb,node, C_emb,node, hardware lifetime H, or long-term utilization u_active. Without these values, the figures cannot be reproduced and the reported dominance of embodied over operational water cannot be independently assessed. A table of all parameters, with sources and ranges, should be added, together with a one-way or multi-way sensitivity analysis for the main results.
  3. [Sec. III-A, Figs. 1–3] The experimental support is thin: 3 runs for STREAM Triad and 2 runs each for CG.C and EP.C, with no error bars, confidence intervals, or statistical comparison in the figures. Statements that the benefit 'diminishes' or 'becomes marginal' at higher thread counts are therefore not supported by uncertainty quantification. Per-point dispersion (std, min–max, or box plots) and, where possible, paired tests across thread counts are needed.
minor comments (5)
  1. [Abstract] Typo: 'We proposes' should be 'We propose'.
  2. [Sec. II-A / III-A] 'ASTRON'spmtmeasurement' should be spaced as 'ASTRON's pmt measurement'.
  3. [Fig. 1–3 captions] The figures are described only in captions; please add axis labels, units, and legends, and ensure error bars are visible if included.
  4. [Sec. III-A] The system has 72 logical CPUs but thread counts stop at 36. State explicitly whether hyperthreading was intentionally excluded and why, since this affects the interpretation of 'full utilization'.
  5. [Sec. III-B] The phrase 'run-to-idle effect' is used but not defined. Clarify whether this refers to fixed idle overhead during a job, or to energy consumed after the job completes while the node is idle.

Circularity Check

0 steps flagged

No significant circularity; thread-scaling results are a transparent accounting application of measured runtime and energy with external embodied-intensity factors.

full rationale

The paper's derivation chain is explicit and self-contained: operational water and carbon are linear functions of measured job energy (Eqs. 2–3), and embodied impacts are whole-node hourly rates (taken from external lifecycle models [4–6]) multiplied by measured job runtime (Eq. 5). The central result—higher thread counts reduce total footprint—follows directly from the empirically observed runtime reduction, as the paper itself states ('because the run times decrease'). No parameter is fitted to the reported trend, no output is used to define the inputs, and no load-bearing self-citation appears; references [4–6] are external prior work. The choice to allocate embodied impact on a whole-node hour basis is a modeling assumption, not a circular step, and the paper explicitly flags allocation-rule sensitivity as future work in Section IV. The operational component is independently measured, giving the framework external grounding. Therefore no circularity is present.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 0 invented entities

No invented physical entities. The central claims rest on (a) a specific embodied-allocation rule that charges full-node rate per wall-clock hour, (b) unstated numeric values for operational and embodied factors drawn from prior work, and (c) standard assumptions about lifetime and utilization. The paper would need to report these values and test alternative allocations to validate its conclusions.

free parameters (8)
  • Node embodied water (W_emb,node) = not reported (taken from refs [4,5,6])
    Amortized embodied water per node from component models; central to W_emb and the water-dominance result.
  • Node embodied carbon (C_emb,node) = not reported (taken from refs [4,5,6])
    Amortized embodied carbon per node; central to C_emb.
  • Hardware lifetime H = not reported
    In Eq. (4), divides embodied totals to get hourly rates; no value given.
  • Long-term utilization u_active = not reported
    In Eq. (4), scales hourly embodied rates; no value given.
  • Water usage effectiveness WUE(T_wb) = not reported
    Operational direct water in Eq. (2); function and value not given.
  • PUE = not reported
    Facility efficiency in Eqs. (2)-(3); no value given.
  • Energy water factor EWF = not reported
    Indirect water per kWh in Eq. (2); no value given.
  • Grid carbon intensity CI_grid = not reported
    Carbon per kWh in Eq. (3); no value given.
axioms (5)
  • ad hoc to paper A job's embodied impact equals the full-node embodied hourly rate multiplied by wall-clock time, independent of the fraction of the node's cores used.
    Eq. (5) applies the full node-hour rate to every thread count; this drives the thread-scaling result and is acknowledged as allocation-rule-sensitive in future work.
  • domain assumption Node-level embodied water/carbon from references [4,5,6] are valid for this hardware.
    Used without recalculation in Eq. (4).
  • domain assumption Operational factors (WUE, PUE, EWF, CI_grid) are known and constant for the facility/grid.
    Used in Eqs. (2)-(3); values not reported.
  • domain assumption Wall-clock time measured by pmt is the correct allocation basis for both energy and embodied impacts.
    All accounting uses t_job; shared-node effects ignored.
  • domain assumption The three benchmarks are representative of HPC jobs.
    STREAM and NPB CG/EP are standard, but conclusions about HPC generally may not generalize.

pith-pipeline@v1.3.0-alltime-deepseek · 2818 in / 13562 out tokens · 130341 ms · 2026-08-01T13:16:59.407788+00:00 · methodology

0 comments
read the original abstract

High performance computing evaluation has traditionally focused on performance and energy, but these metrics alone cannot capture the sustainability cost of runtime configurations. We proposes a unified job-level water and carbon accounting framework with both operational and embodied impacts. Results show that higher thread counts generally reduce total footprint, but the benefit diminishes at higher thread counts. Water is mainly dominated by embodied impact, whereas carbon is mainly dominated by operational impact.

Figures

Figures reproduced from arXiv: 2607.19150 by Chris Broekema, Rob van Nieuwpoort, Xi Chen.

Figure 2
Figure 2. Figure 2: Operational, embodied, and total water footprint versus [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Operational, embodied, and total carbon footprint [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 1
Figure 1. Figure 1: Total water footprint (left) and total carbon foot [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

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. Enhancing SLMs for Sustainable Code Optimization in Radio-Astronomy

    cs.SE 2026-07 conditional novelty 4.0

    A 7-billion-parameter coding model with multi-sampling matches a 32-billion-parameter model on repository-level completion under the same time budget, and compiler-error feedback improves pass rates across all models tested.

Reference graph

Works this paper leans on

14 extracted references · cited by 1 Pith paper

  1. [1]

    Resources, Conservation and Recycling , volume=

    The water use of data center workloads: A review and assessment of key determinants , author=. Resources, Conservation and Recycling , volume=. 2025 , publisher=

    2025

  2. [2]

    Cleaner Environmental Systems , pages=

    Energy use and carbon emissions in high-performance computing: A case study for universities and reduction strategies , author=. Cleaner Environmental Systems , pages=. 2025 , publisher=

    2025

  3. [3]

    Proceedings of the international conference for high performance computing, networking, storage and analysis , pages=

    Toward sustainable hpc: Carbon footprint estimation and environmental implications of hpc systems , author=. Proceedings of the international conference for high performance computing, networking, storage and analysis , pages=

  4. [4]

    Proceedings of the SC'23 Workshops of the International Conference on High Performance Computing, Network, Storage, and Analysis , pages=

    Sustainability in hpc: Vision and opportunities , author=. Proceedings of the SC'23 Workshops of the International Conference on High Performance Computing, Network, Storage, and Analysis , pages=

  5. [5]

    Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis , pages=

    Core hours and carbon credits: Incentivizing sustainability in hpc , author=. Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis , pages=

  6. [6]

    ACM SIGENERGY Energy Informatics Review , volume=

    Understanding the operational carbon footprint of storage reliability and management , author=. ACM SIGENERGY Energy Informatics Review , volume=. 2024 , publisher=

    2024

  7. [7]

    Companion of the 15th ACM/SPEC International Conference on Performance Engineering , pages=

    Footprinter: Quantifying data center carbon footprint , author=. Companion of the 15th ACM/SPEC International Conference on Performance Engineering , pages=

  8. [8]

    Resources, Conservation and Recycling , volume=

    Climate-and technology-specific PUE and WUE estimations for US data centers using a hybrid statistical and thermodynamics-based approach , author=. Resources, Conservation and Recycling , volume=. 2022 , publisher=

    2022

  9. [9]

    Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis , pages=

    ThirstyFLOPS: Water Footprint Modeling and Analysis Toward Sustainable HPC Systems , author=. Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis , pages=

  10. [10]

    Applied Energy , volume=

    Facility-level energy-driven water footprint and scarcity implications of Chinese data centers: a bottom-up analysis and scenario-based projection , author=. Applied Energy , volume=. 2025 , publisher=

    2025

  11. [11]

    Environmental Research Letters , volume=

    The environmental footprint of data centers in the United States , author=. Environmental Research Letters , volume=. 2021 , publisher=

    2021

  12. [12]

    Procedia CIRP , volume=

    Life Cycle Assessment of Artificial Intelligence Applications: Research Gaps and Opportunities , author=. Procedia CIRP , volume=. 2025 , publisher=

    2025

  13. [13]

    arXiv preprint arXiv:2509.00093 , year=

    More than Carbon: Cradle-to-Grave environmental impacts of GenAI training on the Nvidia A100 GPU , author=. arXiv preprint arXiv:2509.00093 , year=

    arXiv

  14. [14]

    Global Environmental Change Advances , volume=

    The green paradox: The climate, environmental, and sustainability implications of artificial intelligence , author=. Global Environmental Change Advances , volume=. 2026 , publisher=

    2026