Omnisolver: An extensible interface to Ising spin-glass and QUBO solvers: adding a distributed GPU brute-force plugin

The gist: This software update extends Omnisolver with omnisolver-bruteforce, adding a first-class plugin for exact exhaustive search of QUBO and Ising instances on CUDA-enabled GPUs, which serves as a practical ground-truth oracle for state-of-the-art classical heuristics such as the Simulated Bifurcation Machine.

How it works

The update contributes three main components: packaging of the single-GPU brute-force kernel of [2] as a first-class Omnisolver plugin with a uniform Python and CLI interface, a Ray-based distributed framework that splits the search into 2k fixed-variable subproblems and dispatches them across multiple GPUs and hosts, with a controller that gathers and merges partial results, and numerical stabilization of the fast float32 ground-state path through compensated incremental updates, periodic exact energy re-anchoring, and best-buffer refresh

The architecture exposes two sampler classes: BruteforceGPUSampler (single-node, single-GPU) and DistributedBruteforceGPUSampler (multi-GPU/multi-node via Ray) The distributed sampler fixes num fixed vars = k variables and enumerates all 2k fixed assignments, each of which becomes a subproblem with N − k unfixed variables, solved by the GPU kernel on a worker Partial solutions are merged on the controller node

Purpose of an exhaustive solver in Omnisolver

An exhaustive backend provides certified reference, i.e. lowest-energy, solutions for QUBO/Ising instances Such ground-truth answers are essential for benchmarking approximate or heuristic methods, for verifying NISQ-oriented workflows, and for assessing claims of quantum runtime advantage In recent years, the latter point has become especially critical in the community, with many studies relying on metrics that explicitly depend on the ground-state energy, such as the optimality gap For example, in recent comparative studies of runtime advantage [5] and of the quantum-classical scaling gap [6], quantum resources were benchmarked against the Simulated Bifurcation Machine (SBM) [7, 8, 9, 10>, the current state-of-the-art classical heuristic for dense QUBO/Ising problems Such results rely on the ability to certify the true optimum on representative instances, which is precisely what an exhaustive GPU solver provides

Performance and Scaling

The performance evaluation was conducted on a local in-house cluster of 8× NVIDIA H100 (96 GB) GPUs The single-GPU sampler was run on one of the same H100s, the distributed sampler on all eight The O(2N) doubling per added spin is the dominant feature on both curves and is clean already at N = 40 on the singleGPU axis The distributed curve flattens out for N ≲ 44 because timings are dominated there by kernel-launch and Ray-scheduling overheads The distributed sampler crosses below the single-GPU sampler between N = 40 and N = 42, and reaches the full ≈8× speedup (matching the 8-GPU pool) from N ≳44 onwards The largest measured size, N = 60 in the distributed mode, takes ≈ 272,300 s (≈ 3.15 days), within 0.3% of the doubling-rule extrapolation from N = 58 (2×67,900 ≈ 271,600 s) The projected single-GPU wall-clock at N = 60 is ≈ 2.17 × 10 6 s (≈ 25 days), about 8.0× the measured 8-GPU time at the same size and consistent with linear strong scaling

Precision and Limitations

Because the GPU backend operates in float32 on the fast ground-state path, returned energies are subject to incremental-update roundoff We re-evaluate the energy of each returned spin configuration from scratch in float64 For the largest measured instances (N = 56, 58, 60), the absolute deviation between reported and recomputed energies stays below 4 × 10-6 for energies of magnitude O(10 2), roughly eight significant digits, consistent with the expected accuracy of the stabilized float32 accumulator Using the same recomputation, we cross-check an SBM run on the same instances: a simulated bifurcation solver running on a single H100 GPU with 2 12 = 4096 parallel samples and 3,000 integration steps, reporting the lowest-energy sample The recomputed energies of the brute-force and SBM solutions agree to within 6 × 10-6, and the returned spin configurations are bit-identical (Hamming distance 0) at N = 56, 58, and 60

Exhaustive enumeration scales as O(2N) irrespective of GPU count or numerical precision: the update lowers the constant factor and removes a float32 drift pitfall, but it cannot change the underlying complexity The plugin is therefore intended as a verification and certification backend for QUBO/Ising research pipelines—providing ground-truth optima against which heuristics, annealers, and NISQ-era quantum approaches can be calibrated—rather than as a routine optimizer for application-scale problems, where heuristic solvers such as the SBM remain the appropriate primary tool Concrete hard limits on the current implementation are: (i) the spin index is held in a 64-bit word, capping the supported problem size at N ≤ 64; (ii) the per-worker suffix size (the number of variables enumerated inside one CUDA kernel launch) is bounded by the GPU’s L2/working-set budget and is set to 23 on H100 / 96 GB in our runs, so larger N must be reached by raising num fixed vars, not suffix size (ii) the projected runtimes for N > 60 are back-of-the-envelope extrapolations under the empirical tN+1 = 2 tN rule and may drift by a small constant factor on different stacks Reported runtimes additionally depend on the hardware and software stack (GPU model, CUDA toolkit, Ray version, interconnect), so the absolute timings in Fig. 1 should be read as a hardware-specific operating point rather than a universal guarantee The natural next steps are: (i) exposing the re-anchoring cadence and compensated-update mode as user-tunable knobs for instance families where the default policy is suboptimal; (ii) replacing the controller-side merge by a hierarchical reduction so the algorithm scales cleanly to 2k workers with k approaching the system size; and (iii) publishing a versioned reproducibility capsule (fixed seeds, hardware descriptors, command presets) so that runtime and numerical claims can be re-checked on heterogeneous clusters The plugin is distributed as omnisolver-bruteforce on PyPI (pip install omnisolver-bruteforce) and is usable either stand-alone or through the Omnisolver CLI as the bruteforce-gpu solver Complete documentation and the benchmark script used to produce Fig. 1 are available in the GitHub repository The minimal distributed-sampler invocation shows how to use it with specific parameters like num fixed vars = 3 and suffix size = 23

The paper is published on arXiv as 2610.10542v1 [physics.comp-ph] on June 1, 2026 The corresponding author is Łukasz Pawela with the email lpawela@iitis.pl The work was supported by the National Science Center (NCN), Poland, via project Sonata Bis 10 and Sonata Bis 15 Quantumz.io Sp. z o.o., acknowledges support received from the Polish Agency for Enterprise Development (PARP) under Project No. FENG.01.01-IP.02-0625/23 The paper is structured across nine pages, detailing the technical contributions and performance metrics of the distributed brute-force plugin The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper

The summary is complete. I will now output it.

1. textbfGeneralization to Larger Instances via Distributed Search Architecture: The improved system can certify exhaustive ground states for QUBO/Ising instances up to N=60 on 8x H100 GPUs, with a projected scaling outlook suggesting that allocating an allocation within reach of leadership-class systems could reduce N=60 time from ≈3.15 days to ≈33 s, consistent with the linear-in-GPU-count outlook. This allows for verifying the optimality gap against classical heuristics like the Simulated Bifurcation Machine on larger problem sizes than previously feasible on a single device, as the distributed sampler crosses below the single-GPU sampler between N = 40 and N = 42 and reaches full ≈8× speedup from N ≳44 onwards.

2. textbfNumerical Stability Enhancement: The system incorporates stabilization mechanisms to mitigate floating-point drift in fast ground-state paths, specifically through compensated incremental updates, periodic exact energy re-anchoring, and best-buffer refresh. This ensures that returned energies are accurate for large instances (N=56, 58, 60), maintaining an absolute deviation below 4 × 10−6 for energies of magnitude O(10 2), which is crucial for reliable comparisons in quantum runtime advantage studies.

3. textbfContextual Certification Oracle: The plugin serves as a practical ground-truth oracle for state-of-the-art classical heuristics such as the Simulated Bifurcation Machine. This allows downstream AI research pipelines to verify the true optimum on representative instances, directly addressing the need to assess claims of quantum runtime advantage by providing certified reference solutions against which approximate methods can be calibrated.

This software update extends Omnisolver with omnisolver-bruteforce, a first-class plugin for exact exhaustive search of QUBO and Ising instances on CUDA-enabled GPUs. The update contributes three components: packaging of the single-GPU brute-force kernel of [Computer Physics Communications 260, 107728, 2021] as a first-class Omnisolver plugin with a uniform Python and CLI interface, a Ray-based distributed framework that splits the search into 2 k fixed-variable subproblems and dispatches them across multiple GPUs and hosts, with a controller that gathers and merges partial results, and numerical stabilization of the fast float32 ground-state path through compensated incremental updates, periodic exact energy re-anchoring, and best-buffer refresh. Stabilization activates automatically for N at least 40 leaving the public sampler API is unchanged. On dense random Ising instances we measure exhaustive solve times up to N=60 on 8 times NVIDIA H100 (96 GB) GPUs, reaching about!3.15 days at N=60 in close agreement with the empirical t N+1=2,t N doubling rule, with the distributed sampler reaching the full about!8 times speedup over a single H100 from N! !44 onwards. The plugin thereby serves as a practical ground-truth oracle for state-of-the-art classical heuristics such as the Simulated Bifurcation Machine.