{"id":"c0225dc3-9598-404f-9dca-6fab899b9af5","arxiv_id":"2506.10531","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GPU-accelerated DQAOA with impact-factor based decomposition runs up to 10x faster than CPU simulations on Frontier, with better scaling up to 160 devices.","lead":"This paper benchmarks a distributed version of the quantum approximate optimization algorithm (DQAOA) on the Frontier supercomputer, testing new ways to split large optimization problems and comparing CPU and GPU simulations. It reports up to 10x speedups from GPUs, though the comparison treats a GPU as one device against one CPU core.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 'up to 10x GPU speedup' rests on an unequal resource comparison: one MI250X GPU is plotted as one CPU 'core' in Figs. 10-11. If the CPU baseline used a full 64-core node, the order-of-magnitude claim would likely shrink.","rationale":"The paper is a benchmarking and systems engineering contribution; its two quantitative claims are the 10x GPU speedup and the convergence improvement from IFD/BFS/PFS decomposition. The convergence claim has some internal tension, since IFD has a lower approximation ratio than random on sparse Max-Cut in Fig. 8(a) even though it reaches the time-to-solution threshold faster, but the claim the paper's abstract and conclusion foreground is the 10x speedup. The reader's weakest-assumption analysis identifies the right load-bearing point, and I agree with it. The x-axis labeling and the sentence that the number of CPU/GPU devices equals the number of sub-QUBOs make the unequal comparison explicit: 160 GPUs are not the same resource as 160 CPU cores. This is not an accusation of intentional unfairness; it may be a conventional 'one device per subproblem' choice, but the paper does not justify it, and the 10x number is not meaningful until the CPU baseline is defined at node level or by total compute budget. The QFw sections are honestly marked as in progress, and Figs. 17-18 show real scaling behavior, which is creditable. No machine-checked proof or public artifacts are supplied, but that is not itself a fatal issue for a systems paper. The concrete test above would settle whether the speedup survives a fair resource comparison; until then, the verdict should remain conditional rather than accept the 10x claim at face value. I therefore leave the reader's CONDITIONAL verdict unchanged.","tokens_in":17308,"tokens_out":7586,"duration_ms":89351,"concrete_test":"Re-run the Fig. 10 and Fig. 11 scaling tests with a matched resource definition: use one full 64-core AMD EPYC node as the CPU baseline and one full Frontier node with its MI250X GPUs as the GPU baseline at equal node counts, keeping sub-QUBO size and count fixed. Alternatively, in Fig. 9, repeat the QAOA(8)-QAOA(26) timings with Qiskit Aer CPU runs configured to use all 64 cores of a node (for example via OMP_NUM_THREADS=64 or 64 MPI ranks) instead of a single core. If the GPU/CPU time ratio at matched resources falls below 10x, the headline 'up to 10x' must be revised to a per-device speedup rather than a system-level speedup.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central 10x speedup claim is load-bearing and is derived from Figs. 9-11, where CPU and GPU curves are both labeled 'Number of Cores.' The text says the runs use '64-core AMD EPYC CPUs and AMD Instinct MI250X GPUs' and that the number of CPU/GPU devices used equals the number of sub-QUBOs. Thus a GPU point at x=160 uses 160 MI250X GPUs (40 Frontier nodes), while a CPU point at x=160 uses 160 EPYC cores (about 2.5 nodes). The comparison is therefore one GPU per sub-QUBO versus one CPU core per sub-QUBO, not one compute node versus one compute node. Because an MI250X contains far more parallelism and memory bandwidth than a single EPYC core, this baseline choice alone can explain an order-of-magnitude gap. Section II's claim that 'even a single GPU significantly outperforms a CPU-based implementation' does not specify whether the CPU baseline is one socket, one core, or one MPI rank, so the abstract's '10x over CPU-based simulations' is not tied to a reproducible resource definition. The IFD convergence claim is separate and may hold, but the headline speedup is the contribution most at risk.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents GPU-accelerated Distributed Quantum Approximate Optimization Algorithm (DQAOA) for solving large-scale QUBO problems on the Frontier supercomputer. It introduces an impact-factor-directed (IFD) decomposition strategy, compares it with random, BFS, and PFS decomposition, and reports scaling experiments on up to 160 or 1000 CPU cores/GPUs. The authors claim an up-to-10x speedup for GPU-based DQAOA over CPU-based DQAOA, and they describe an integration with the Quantum Framework (QFw) for distributed execution of QAOA circuits. The paper also reports strong scaling for GHZ and QAOA workloads through QFw.","tokens_in":17609,"tokens_out":3374,"duration_ms":39892,"significance":"If the central speedup claim survives a fair resource comparison, this work provides a useful engineering benchmark for deploying hybrid quantum-classical optimization on an exascale system. The comparison of decomposition strategies on dense and sparse QUBO problems is also of interest, particularly the finding that IFD degrades approximate ratio on Max-Cut while improving time-to-solution. The QFw integration results, including 11.7x speedup for GHZ-30, are a positive contribution. However, the headline 10x GPU-over-CPU claim is currently supported only by an unequal device-count comparison, and the approximate-ratio baseline is a same-group heuristic without comparison to exact or classical baselines. The absence of artifacts further limits the reproducibility of the empirical claims.","major_comments":[{"comment":"The central claim of 'up to 10x speedup' rests on a resource-unfair comparison. The text states that the number of CPU/GPU devices used equals the number of sub-QUBOs, and both curves in Figures 10 and 11 are labeled with 'Number of Cores' on the x-axis. Since each AMD MI250X GPU contains far more parallelism, memory bandwidth, and compute throughput than a single AMD EPYC core, treating one GPU as equivalent to one CPU core in the scaling plots is not an apples-to-apples comparison of compute resources. The order-of-magnitude speedup could largely be an artifact of this accounting. Please re-run the comparison with matched resources (e.g., equal node counts, equal peak FLOPS, or equal memory bandwidth) or, at minimum, report the CPU core-to-GPU ratio and justify the comparison as an application-level metric. Without this, the abstract and Section VII's 'approximately 10 times faster' claim are not supported.","section":"Section VI-B1, Figures 10 and 11"},{"comment":"The approximate ratio is defined as the ratio between DQAOA energy and energy from hybrid quantum annealing (HQA), citing ref [46] by the same group. HQA is itself a heuristic, so the A.R. values do not establish absolute solution quality relative to optimal or classical exact baselines. Please add a comparison to exact results on small instances (e.g., brute force for N up to 20) or to a strong classical solver (e.g., Gurobi or simulated annealing) for larger instances, and report absolute energies or approximation ratios to those baselines. This is especially important because Sections VI-A and VII claim that IFD decomposition improves solution quality; the current baseline cannot validate that claim independently.","section":"Section VI-A, Figure 7 and Figure 8"},{"comment":"Figure 8(a) shows approximate ratios for random, IFD, BFS, and PFS decomposition, and the text says results come from 10 independent trials, yet no error bars or statistical significance tests are reported. Without measures of variance, the observed differences (e.g., BFS vs PFS, BFS vs random) may not be significant. Please add error bars for both A.R. and time-to-solution in Figure 8, or report confidence intervals and a statistical test.","section":"Section VI-A, Figure 8"},{"comment":"No artifacts, benchmark scripts, problem instances, or configuration details are provided. Since the paper's primary contribution is an empirical benchmark, the absence of released code and data makes it impossible for the community to reproduce the scaling and speedup results. Please provide the source code, the exact QUBO instances, the software versions (already partly given), the MPI configuration, and the CPU/GPU resource allocation details for each experiment.","section":"General (Reproducibility)"}],"minor_comments":[{"comment":"The statement that 'even a single GPU significantly outperforms a CPU-based implementation' should specify whether the CPU baseline is one core, one socket, one MPI rank, or one node; otherwise the comparison is not interpretable.","section":"Section II-A, Figure 2 and Figure 9"},{"comment":"The x-axis is labeled 'Number of Cores' for both CPU and GPU curves, which is misleading for the GPU runs; please use 'Number of Devices' or 'Number of CPU cores / GPUs' and separately define the resource allocation in the caption.","section":"Figure 10 and Figure 11 captions"},{"comment":"The variable 'subQUBO_size' is used in the loop bounds but is not defined in the algorithm's input list; it presumably equals 'sizeSubQ'. Please clarify.","section":"Algorithm 1"},{"comment":"The complexity expressions O(N^3 + mk^2) and O(mkN^2) for decomposition and aggregation are stated without derivation; a brief justification or a reference would improve clarity.","section":"Section VI-B3, Figure 12"},{"comment":"The definition of time-to-solution as 'the change of the A.R. remains within ±0.1 for 5 consecutive cycles' is ambiguous; please specify whether the threshold applies to the absolute or relative change and how the final time is measured.","section":"Section VI-A, Figure 8(b)"},{"comment":"Ref [46] (the HQA benchmark) is an arXiv preprint by the same research group; the paper should explicitly disclose this self-reference and discuss the potential for bias.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is an empirical benchmark study, and the main claimed contribution (10x GPU speedup) is currently undercut by a resource-unfair comparison. The issue is potentially fixable by re-running with matched resources or by carefully reporting normalization, so I do not recommend rejection. However, the same-group heuristic baseline and missing artifacts raise reproducibility concerns that the editor may wish to weigh. I would ask for a revision that addresses the four major comments before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a real engineering benchmark paper, not a breakthrough. The contribution is new benchmark data on GPU-accelerated DQAOA at Frontier scale, comparing random, IFD, BFS, and PFS decomposition on dense and sparse QUBOs. That comparison is genuinely new and useful for practitioners.\n\nIt does several things right: the scaling tests up to 160 and 1000 devices are non-trivial, the complexity equations for decomposition and aggregation are stated, and the QFw integration is honestly labeled as in progress.\n\nThe main problem is the 'up to 10x speedup' claim. In Figures 10-11, the x-axis is 'Number of Cores' for both CPU and GPU, and the text says the number of devices equals the number of sub-QUBOs. That means each CPU point is one EPYC core and each GPU point is one MI250X. An MI250X has far more parallelism and memory bandwidth than a single core, so the comparison is not resource-fair. If you ran the CPU baseline on a full 64-core node, the gap would shrink drastically. The stress-test note holds up. The claim in Section II that 'even a single GPU significantly outperforms' a CPU-based implementation is also under-specified because it doesn't say whether the CPU baseline is one socket, one core, or one MPI rank.\n\nOther soft spots: the approximate-ratio baseline for Figure 7 is HQA from the same group, not an exact or classical method, so the 'A.R.' numbers are relative to a method they are also benchmarking elsewhere. Figure 8 has no error bars. No code or data is provided, so nothing here is independently reproducible yet.\n\nTo be clear, the decomposition convergence claims seem separate and more robust: IFD and BFS/PFS appear to beat random decomposition on dense and sparse instances respectively, and that is the more interesting result. The paper is worth a serious referee, but the CPU/GPU comparison needs to be redone on a per-node or per-cost basis, and the artifacts should be released.\n\nRecommendation: send to peer review with major revision. The authors should fix the resource comparison, add error bars, and provide code/data. If they do, this becomes a useful benchmark for the hybrid quantum-HPC community.","headline":"Solid benchmark study of GPU-accelerated DQAOA on Frontier, but the headline '10x speedup' rests on an unfair one-GPU-vs-one-CPU-core comparison.","tokens_in":18148,"tokens_out":2182,"would_cite":false,"duration_ms":23986,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GPU-accelerated distributed QAOA on Frontier achieves up to a 10x speedup over CPU execution, and impact-factor decomposition converges faster than random decomposition on dense QUBO problems.","keywords":["distributed quantum computing","high performance computing","quantum optimization","GPU acceleration","QAOA","DQAOA","QUBO","Frontier"],"falsifier":"Run the same DQAOA benchmarks with the CPU baseline using all 64 cores of a Frontier node instead of a single core, and measure time-to-solution for QUBO sizes 300, 500, and 1000; if the GPU-to-CPU ratio falls close to 1, the 'up to 10x' claim collapses.","tokens_in":17093,"feed_emoji":"⚡","tokens_out":8552,"duration_ms":88990,"temperature":0.7,"pith_summary":"This paper tries to establish that distributed QAOA (DQAOA) can be made practically scalable for large QUBO optimization problems by running on GPU-accelerated HPC systems. It claims that on the Frontier supercomputer, replacing CPU-based quantum circuit simulation with GPU-based simulation cuts time-to-solution by roughly an order of magnitude for dense problems of 300 to 1000 variables. It also claims that a new impact-factor-directed (IFD) decomposition, which samples the variables whose flips change the objective most, converges faster than random decomposition on dense problems, while graph-traversal variants BFS and PFS do better on sparse Max-Cut problems. The work matters because it provides concrete benchmarks and a software path for running hybrid quantum-classical algorithms on exascale systems rather than small toy instances.","feed_headline":"GPUs make distributed QAOA 10x faster on Frontier","feed_subtitle":"Impact-factor decomposition also converges faster than random splitting on dense QUBO problems.","key_machinery":"The central mechanism is the DQAOA cycle: decompose a large QUBO matrix into smaller sub-QUBOs, solve each sub-QUBO with single-layer QAOA on a distributed simulator backend, then aggregate the sub-solutions into the global solution, accepting only updates that lower energy. QAOA is a variational quantum algorithm that prepares a parameterized quantum state to minimize the cost Hamiltonian of the QUBO. The decomposition machinery is the impact-factor-directed (IFD) method, which ranks decision variables by local sensitivity of the objective and packs the top-ranked variables into sub-QUBOs, plus BFS/PFS traversal variants that follow strong pairwise couplings. GPU-accelerated state-vector simulation is what carries the 10x speedup claim, while MPI-based distribution carries the scalability across nodes.","core_discovery":"The central discovery is that DQAOA on Frontier runs nearly 10x faster when quantum circuit simulation is done on GPUs instead of CPU cores, and that the choice of decomposition strategy controls practical performance at least as much as simulator speed. On fully connected metamaterial QUBO instances of size 300, 500, and 1000, IFD decomposition consistently reaches higher approximate ratios in fewer cycles than random decomposition. On sparse Max-Cut graphs, priority-first search (PFS) and breadth-first search (BFS) decompositions produce higher approximate ratios than random and IFD because they preserve local pairwise interactions. The paper also demonstrates strong scaling to thousands of sub-QUBOs on up to 1000 CPU/GPU devices, and it identifies aggregation as the dominant computational cost for large problem sizes.","pith_inferences":["Because the 10x speedup is measured entirely in simulation, a reader should not expect the same factor when QAOA runs on real quantum hardware; the decomposition improvements would transfer independently of simulation hardware.","The one-GPU-versus-one-CPU-core plotting convention probably understates CPU capability, so the headline speedup factor should be read as sensitive to how computational resources are counted.","Because decomposition cost scales as O(N^3) and aggregation as O(mkN^2), the paper's own runtime breakdown suggests that for very large N these classical steps, not the QAOA solves, will become the bottleneck; incremental or hierarchical aggregation would be a natural next step.","The BFS/PFS result on Max-Cut suggests a general recipe for sparse QUBOs: start from high-impact variables and extend along strong pairwise couplings; testing this recipe on other sparse problems such as TSP would show whether it transfers."],"forward_implications":["Dense QUBO problems up to at least 1000 variables become solvable by DQAOA in times that are roughly an order of magnitude shorter when a GPU-accelerated cluster is available.","Impact-factor decomposition gives a practical advantage for dense industrial QUBO instances by reducing the number of DQAOA cycles needed to converge.","For sparse problems like Max-Cut, graph-traversal decompositions should be preferred over pure impact-factor selection, since they preserve local structure.","Increasing the number of sub-QUBOs, when backed by enough parallel CPU or GPU devices, accelerates convergence and makes expensive HPC resources directly translate into faster time-to-solution.","Asynchronous task-based execution through the QFw software stack can replace the rigid one-process-per-subproblem MPI binding and improve resource utilization as the number of sub-QUBOs grows."],"supporting_citations":[{"why":"Defines DQAOA itself and the random-decomposition baseline that this work extends and compares against.","marker":"[4]"},{"why":"Supplies the impact-factor directed decomposition method used for sub-QUBO selection.","marker":"[40]"},{"why":"Introduces the Quantum Framework (QFw) used for asynchronous HPC-quantum task distribution.","marker":"[22]"},{"why":"Provides the simulator backend used in the CPU versus GPU benchmarks.","marker":"[5]"},{"why":"Provides the MPI-capable state-vector simulator used for distributed QAOA runs.","marker":"[6]"},{"why":"Supplies the metamaterial QUBO instances and the performance-analysis context for HPC/quantum optimization.","marker":"[3]"},{"why":"Supplies the hybrid quantum annealing baseline that defines the approximate ratio used to compare solution quality.","marker":"[46]"}],"fun_headline_variants":["GPU-accelerated QAOA runs 10x faster on Frontier supercomputer","Distributed QAOA achieves 10x GPU speedup on HPC systems","Frontier GPUs give distributed QAOA 10x speedup over CPUs","QAOA decomposition choice rivals GPU speed for HPC performance","Smart decomposition adds to 10x GPU speedup for QAOA on HPC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that one GPU and one CPU core are comparable units when counting 'cores' in the scaling plots; a CPU baseline using all 64 cores of a node, or matched peak compute, would make the claimed 10x speedup much smaller.","fun_headline_variants_meta":{"raw":{"variants":["GPU-accelerated QAOA runs 10x faster on Frontier supercomputer","Distributed QAOA achieves 10x GPU speedup on HPC systems","Frontier GPUs give distributed QAOA 10x speedup over CPUs","QAOA decomposition choice rivals GPU speed for HPC performance","Smart decomposition adds to 10x GPU speedup for QAOA on HPC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000801,"raw_usage":{"total_tokens":3475,"prompt_tokens":855,"completion_tokens":2620,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":2522}},"tokens_in":471,"tokens_out":2620,"duration_ms":20843,"temperature":1.0,"reasoning_tokens":2522,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:23:36.525392+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same DQAOA benchmarks with the CPU baseline using all 64 cores of a Frontier node instead of a single core, and measure time-to-solution for QUBO sizes 300, 500, and 1000; if the GPU-to-CPU ratio falls close to 1, the 'up to 10x' claim collapses.","supporting_citations":[{"cited_title":"Problem decomposition for quantum annealing,","cited_arxiv_id":null,"evidence_quote":"Supplies the impact-factor directed decomposition method used for sub-QUBO selection."},{"cited_title":"Nwqsim: Northwest quantum circuit simulation environment,","cited_arxiv_id":null,"evidence_quote":"Provides the MPI-capable state-vector simulator used for distributed QAOA runs."},{"cited_title":"Quantum Annealing for Combinatorial Optimization: A Benchmarking Study","cited_arxiv_id":"2504.06201","evidence_quote":"Supplies the hybrid quantum annealing baseline that defines the approximate ratio used to compare solution quality."}],"review_version":1}