Pith. sign in

REVIEW 2 major objections 4 minor 17 references

Running GPU and CPU unit-commitment solvers together with aggressive presolve cuts max runtime on a 6049-bus grid from 42 minutes to under 6 minutes while keeping solution quality.

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 · grok-4.5

2026-07-12 01:04 UTC pith:TAOPHMTG

load-bearing objection Solid engineering speedups on hard GO Challenge UC instances; the real lever is aggressive Gurobi presolve, not the concurrent GPU–CPU framing. the 2 major comments →

arxiv 2607.03627 v1 pith:TAOPHMTG submitted 2026-07-03 math.OC

Rapid Concurrent GPU-CPU Solvers for Scalable Unit Commitment in Large Power Grids

Hussein Sharadga , Yuhan Du , Javad Mohammadi This is my paper
classification math.OC MSC 90C1190C0690C90
keywords unit commitmentGPU accelerationconcurrent solverspresolveoptimal power flowlarge-scale power gridsPDHGmixed-integer programming
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.

Large power grids must decide which generators to turn on and how hard to run them under tight physical and contingency constraints. The resulting unit-commitment problems contain millions of variables and are slow to solve. This paper shows that launching a GPU solver and a CPU solver on the same machine at the same time, then stopping the slower one the instant the faster one finishes, plus a deliberately aggressive presolve that doubles the number of reduction passes and forms the dual, reliably shrinks wall-clock time. On the hardest 6049-bus instances from the DOE Grid Optimization Challenge the method delivers speedups between roughly two- and five-and-a-half-fold, bringing the longest run from 42 minutes down to under 6 minutes across 45 cases, without harming solution quality. A sympathetic reader cares because the same formulation is already used for day-ahead and reliability operations; any method that makes it finish faster on existing hardware expands the set of grids and contingencies that can be cleared inside operational deadlines.

Core claim

On a 6049-bus system with millions of continuous and binary variables, concurrent execution of a GPU-based PDHG pipeline and a CPU-based barrier pipeline, terminated by the first to converge, together with doubled presolve passes and dual reductions, produces speedups of 2.14x to 5.61x versus the untuned baseline and reduces maximum runtime from 42.12 minutes to 5.77 minutes across 45 test cases while preserving objective values to machine precision.

What carries the argument

The concurrent GPU-CPU race: two full solver pipelines (presolve then PDHG-or-barrier, crossover, simplex clean-up, branch-and-bound) run on partitioned cores of one machine; the first to meet the optimality gap aborts the other. Aggressive presolve (extra passes plus dual formation) supplies the reduced model both pipelines actually solve.

Load-bearing premise

The claim rests on the premise that spending more time in an aggressive presolve will reliably shrink the mixed-integer model to a tractable continuous core without creating a numerically harder smaller problem.

What would settle it

Rerun the identical 45 instances with the concurrent race disabled (GPU-only and CPU-only) and with the default versus doubled-pass dual-enabled presolve; if the reported maximum and average speedups disappear or if solution quality falls below the claimed 0.9999999 ratio, the central claim fails.

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

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

2 major / 4 minor

Summary. The paper proposes a concurrent GPU–CPU solver framework for large-scale AC unit commitment (UC) problems drawn from the DOE ARPA-E GO Challenge 3 formulation. On a single machine it runs a GPU PDHG pipeline and a multi-threaded CPU barrier pipeline in parallel on the same relaxed MIP, terminating the slower solver upon first convergence to a shared optimality gap; both pipelines are followed by crossover/simplex and branch-and-bound. Convergence is further accelerated by an aggressive presolve (doubled passes plus dual reductions). On the 6,049-bus network (millions of variables/constraints) across 45 instances from Trials 2–3, the tuned concurrent framework reports wall-clock speed-ups of 2.14×–5.61× versus the untuned baseline, cutting maximum runtime from 42.12 min to 5.77 min while preserving solution quality (objective ratios ≥0.9999999, average ≈0.99999998).

Significance. If the empirical claims hold, the work supplies a practical, immediately usable acceleration recipe for industrial-scale UC that exploits complementary GPU/CPU strengths and modern MIP presolve features. Strengths include transparent reporting on externally defined ARPA-E instances, detailed stage-wise timing breakdowns (Figs. 2, Tables III–VII), explicit quality ratios, and open acknowledgment that smaller reduced models can sometimes be harder. The results are therefore falsifiable and useful to operators facing GO-Challenge-scale problems. Limitations of scope (single topology, Gurobi 13 beta–specific GPU PDHG path) do not erase the engineering value, but they do constrain claims of broad architectural generality.

major comments (2)
  1. [§III.D, Table VI, Tables III–IV, VII] §III.D and Table VI show that under the aggressive-presolve (“tuned”) settings the CPU barrier path finishes first in all 45 cases (GPU first = 0 %). Consequently the concurrent termination mechanism contributes nothing to the headline tuned runtimes reported in Tables III–IV and the abstract. Those speed-ups are produced by the ~99 % model collapse documented in Table VII (D2/D3 constraints 3.8 M → 38 k, binaries essentially eliminated). The title, abstract and contribution statement package concurrency as a primary ingredient; the evidence shows it is secondary once presolve has already reduced the problem to a compact continuous core. Either reframe the central claim around aggressive presolve (with concurrency as a robustness device for the untuned regime) or supply additional experiments in which concurrency still yields measurable gains after the same presolve.
  2. [§II.D, Table VII] The paper itself notes (§II.D) that “smaller optimization problems can be more difficult to solve ldots due to ldots numerical stability issues.” The entire speed-up narrative rests on the untested premise that doubling presolve passes and enabling dual reductions reliably produces a tractable continuous core rather than a numerically pathological one. No ablation isolates the dual-reduction flag, no multi-seed or multi-topology variation is reported, and all results are obtained with a single commercial solver (Gurobi 13 beta) whose GPU PDHG path is Linux/NVIDIA-only. A minimal sensitivity study (or at least a clear statement of the numerical conditions under which the reduction remains beneficial) is required before the 2–5× claims can be treated as robust.
minor comments (4)
  1. [Tables III–V, Fig. 1] Throughout the results section “Trail” is used for “Trial”; correct for consistency with the ARPA-E nomenclature.
  2. [§II.A, §II.D] §II.A heading contains a stray space (“F ormulation”); several other minor typographical slips appear (e.g., “problem size problem size” in §II.D).
  3. [Fig. 2] Figure 2 captions and axis labels would benefit from explicit units (seconds) and a clearer distinction between the three stacked stages; the current visual is readable but dense.
  4. [Table II, Algorithm 1] The hardware table lists “Threads in Gurobi 24 (default 32)” while Algorithm 1 and the text speak of 48 cores partitioned evenly; a single clarifying sentence would remove the apparent discrepancy.

Circularity Check

0 steps flagged

No circularity: empirical wall-clock speedups on external ARPA-E instances are measured, not derived by construction from self-cited baselines or fitted parameters.

full rationale

This is a computational engineering paper whose central claims are measured runtimes and objective ratios on 45 external GO Challenge 3 instances (6,049-bus system). The concurrent GPU–CPU pipeline (Algorithm 1) and aggressive presolve (doubled passes + dual reductions) are proposed methods; the reported 2.14×–5.61× speedups (Tables III–IV), model-size reductions (Table VII), and quality scores (Table V, average ≈0.99999998) are direct experimental outcomes against an untuned Gurobi baseline, not quantities defined in terms of the inputs or forced by self-citation. Prior works [11]–[14] supply the baseline UC model and earlier GPU/PDHG results, which is ordinary incremental research; they do not contain a uniqueness theorem, ansatz, or fitted parameter that is then re-presented as a prediction. No equation equates a claimed result to its own definition, no parameter is fit to a subset and then “predicted,” and the external ARPA-E data provide an independent benchmark. The observation that tuned presolve collapses the MIP so that the CPU path always finishes first (Table VI) is an empirical finding about the methods, not a circular reduction. Score 0 is therefore the correct, proportionate outcome.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim is an empirical systems result. It rests on the GO Challenge 3 problem definition, commercial Gurobi correctness, the authors’ prior baseline model, and several hand-chosen solver/presolve/thread settings. No new physical entities are postulated; free parameters are solver knobs, not fitted scientific constants.

free parameters (4)
  • PDHG iteration limit / feasibility / convergence tolerances
    Authors set iteration limit 100,000, absolute feasibility 1e-5, convergence tolerance 1e-4; these are hand-tuned and affect PDHG quality and crossover time.
  • Presolve pass count (doubled) and dual-reduction enablement
    Aggressive presolve is defined by doubling passes and allowing dual reductions; the exact commercial flags and pass count are chosen by the authors and drive most of the reported speedup.
  • Thread partition (24 threads per solver, 48 cores used)
    Even split of 48 of 56 cores is a design choice; authors note more threads did not help and sometimes hurt.
  • Optimality gap tolerance (identical for both solvers)
    Branch-and-bound termination depends on a specified gap; value is not numerically stated beyond ‘specified optimality gap’.
axioms (4)
  • domain assumption The DOE ARPA-E GO Challenge 3 AC multi-period unit-commitment formulation (53-page + supplements) correctly encodes the operational problem being solved.
    Section II.A adopts the full GO Challenge 3 specification as the problem definition.
  • domain assumption Gurobi 13.0.1 beta PDHG (GPU), barrier, crossover, simplex, and branch-and-bound implementations are numerically correct for these instances.
    All pipelines are Gurobi-based; results inherit commercial-solver correctness assumptions.
  • domain assumption The authors’ prior baseline model ([11]–[13]) with scaled score 0.98 is a valid starting formulation for the concurrent experiments.
    Section II.B adopts that baseline without re-deriving feasibility guarantees here.
  • ad hoc to paper Terminating the slower concurrent solver upon first convergence does not discard a meaningfully better feasible solution within the shared optimality gap.
    Algorithm 1 returns the first converged solution; quality is checked only via objective ratio to the untuned run, not via dual certificates across both paths.

pith-pipeline@v1.1.0-grok45 · 13666 in / 2939 out tokens · 29045 ms · 2026-07-12T01:04:57.270773+00:00 · methodology

0 comments
read the original abstract

This paper presents an accelerated solver for the unit commitment problem in large-scale power systems. The approach is based on the concurrent execution of GPU- and CPU-based optimization solvers on a single machine, with the solver that converges first terminating the other to minimize overall runtime. This strategy effectively harnesses the complementary strengths of different solvers. Convergence is further accelerated through a systematic and aggressive presolve approach. Numerical experiments on a 6,049-bus system with millions of decision variables and constraints demonstrate speedups ranging from 2.14x to 5.61x, reducing the maximum runtime from 42.12 minutes to 5.77 minutes across 45 test cases. These results highlight the scalability and computational efficiency of the proposed GPU-CPU concurrent solver framework.

Figures

Figures reproduced from arXiv: 2607.03627 by Hussein Sharadga, Javad Mohammadi, Yuhan Du.

Figure 1
Figure 1. Figure 1: Computation time distribution for the 6049-bus network. The tuned [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of computation times between tuned and default solvers across two random scenarios. Each bar represents the total runtime, decomposed [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

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

17 extracted references · 9 canonical work pages

  1. [1]

    Monitoring and optimization for power grids: a signal processing perspective,

    G. B. Giannakis, V . Kekatos, N. Gatsis, S. J. Kim, H. Zhu, and B. F. Wollenberg, “Monitoring and optimization for power grids: a signal processing perspective,”IEEE Signal Process. Mag., vol. 30, no. 5, pp. 107–128, 2013, doi: 10.1109/MSP.2013.2245726

    work page doi:10.1109/msp.2013.2245726 2013

  2. [2]

    Supporting ARPA-E Power Grid Optimization: Final Scien- tific/Technical Report,

    S. T. Elbert, J. T. Holzer, A. Veeramany, C. DeMarco, H. Mittelmann, and T. Overbye, “Supporting ARPA-E Power Grid Optimization: Final Scien- tific/Technical Report,” Pacific Northwest National Laboratory (PNNL), Richland, W A, PNNL-36158, 2024. Available: https://www.osti.gov/ser vlets/purl/2404530; last accessed Mar 15, 2026

    arXiv 2024

  3. [3]

    A Physics-Guided Graph Convolu- tion Neural Network for Optimal Power Flow,

    M. Gao, J. Yu, Z. Yang, and J. Zhao, “A Physics-Guided Graph Convolu- tion Neural Network for Optimal Power Flow,”IEEE Trans. Power Syst., vol. 39, no. 1, pp. 380–390, 2024, doi: 10.1109/TPWRS.2023.3238377

    doi:10.1109/tpwrs.2023.3238377 2024

  4. [4]

    Agent-based distributed security constrained optimal power flow,

    J. Mohammadi, G. Hug, and S. Kar, “Agent-based distributed security constrained optimal power flow,”IEEE Trans. Smart Grid, vol. 9, no. 2, pp. 1118–1130, 2018, doi: 10.1109/TSG.2016.2577684

    work page doi:10.1109/tsg.2016.2577684 2018

  5. [5]

    State-of-the-art, challenges, and future trends in security constrained optimal power flow,

    F. Capitanescuet al., “State-of-the-art, challenges, and future trends in security constrained optimal power flow,”Electr . Power Syst. Res., vol. 81, no. 8, pp. 1731–1741, 2011, doi: 10.1016/j.epsr.2011.04.003

    work page doi:10.1016/j.epsr.2011.04.003 2011

  6. [6]

    Operations research in optimal power flow: A guide to recent and emerging methodologies and appli- cations,

    J. K. Skolfield and A. R. Escobedo, “Operations research in optimal power flow: A guide to recent and emerging methodologies and appli- cations,”Eur . J. Oper . Res., vol. 300, no. 2, pp. 387–404, 2022, doi: 10.1016/j.ejor.2021.10.003

    work page doi:10.1016/j.ejor.2021.10.003 2022

  7. [7]

    GO Competition Challenge 3: Problem, Solvers, and Solution Analysis,

    J. T. Holzer, S. Elbert, H. Mittelmann, R. O’Neill, and H. Oh, “GO Competition Challenge 3: Problem, Solvers, and Solution Analysis,” Energy Systems, 2024, doi: 10.1007/s12667-024-00708-1

    work page doi:10.1007/s12667-024-00708-1 2024

  8. [8]

    Accurate and Warm-Startable Linear Cutting-Plane Relaxations for ACOPF,

    D. Bienstock and M. Villagra, “Accurate and Warm-Startable Linear Cutting-Plane Relaxations for ACOPF,”2024 IEEE 63rd Conference on Decision and Control (CDC), 2024, doi: 10.1109/CDC56724.2024.10886304

    doi:10.1109/cdc56724.2024.10886304 2024

  9. [9]

    A Parallelized, Adam-Based Solver for Reserve and Secu- rity Constrained AC Unit Commitment,

    S. Chevalier, “A Parallelized, Adam-Based Solver for Reserve and Secu- rity Constrained AC Unit Commitment,”Electric Power Syst. Res., vol. 235, no. 110685, 2024, doi: 10.1016/j.epsr.2024.110685

    work page doi:10.1016/j.epsr.2024.110685 2024

  10. [10]

    Managing Power Balance and Reserve Feasibility in the AC Unit Commitment Problem,

    R. Parker and C. Coffrin, “Managing Power Balance and Reserve Feasibility in the AC Unit Commitment Problem,”Electric Power Syst. Res., vol. 234, no. 110670, 2024, doi:10.1016/j.epsr.2024.110670

    work page doi:10.1016/j.epsr.2024.110670 2024

  11. [11]

    Optimizing Multi-Timestep Security-Constrained Optimal Power Flow for Large Power Grids,

    H. Sharadga, J. Mohammadi, C. Crozier, and K. Baker, “Optimizing Multi-Timestep Security-Constrained Optimal Power Flow for Large Power Grids,”2024 IEEE Texas Power Energy Conf., 2024, doi: 10.1109/TPEC60005.2024.10472229

    doi:10.1109/tpec60005.2024.10472229 2024

  12. [12]

    Scalable Solutions for Security-Constrained Optimal Power Flow with Multiple Time Steps,

    H. Sharadga, J. Mohammadi, C. Crozier and K. Baker, “Scalable Solutions for Security-Constrained Optimal Power Flow with Multiple Time Steps,”IEEE Transactions on Industry Applications, vol. 61, no. 3, pp. 4812-4821, 2025, doi: 10.1109/TIA.2025.3532927

    work page doi:10.1109/tia.2025.3532927 2025

  13. [13]

    Scalable Unit Commitment for Large Power Grids: Relaxation, Tightening, and GPU-Based Solvers Evaluation,

    H. Sharadga and J. Mohammadi, “Scalable Unit Commitment for Large Power Grids: Relaxation, Tightening, and GPU-Based Solvers Evaluation,”2025 57th North American Power Symposium (NAPS), 2025, doi: 10.1109/NAPS66256.2025.11272365

    doi:10.1109/naps66256.2025.11272365 2025

  14. [14]

    GPU-Accelerated Optimiza- tion Solver for Unit Commitment in Large-Scale Power Grids,

    H. Sharadga and J. Mohammadi, “GPU-Accelerated Optimiza- tion Solver for Unit Commitment in Large-Scale Power Grids,” 2026 IEEE Texas Power Energy Conf., 2026. [Online]. Available: https://doi.org/10.48550/arXiv.2512.06715

    work page doi:10.48550/arxiv.2512.06715 2026

  15. [15]

    Grid optimization challenge 3 problem formulation

    J. Holzer et al., “Grid optimization challenge 3 problem formulation.” https://www.pnnl.gov/publications/grid-optimization-competition-challen ge-3-problem-formulation (accessed Mar 15, 2026)

    2026

  16. [16]

    Data: ARPA-E Grid Optimization (GO) Competition Chal- lenge 3,

    ARPA-E, “Data: ARPA-E Grid Optimization (GO) Competition Chal- lenge 3,” https://catalog.data.gov/dataset/arpa-e-grid-optimization-go-c ompetition-challenge-3; accessed Mar 15, 2026

    2026

  17. [17]

    Avail- able online: https://docs.gurobi.com/13.0/; accessed Mar 15, 2026

    Gurobi Optimization.Gurobi 13.0.1 (Beta) Documentation, 2026. Avail- able online: https://docs.gurobi.com/13.0/; accessed Mar 15, 2026

    2026