A reproducible classical reference for D-Wave Advantage2's 2024–2026 industrial benchmarks. Ground-state 3D ±J Ising at N = 10⁶ spins on a $1.57/hour GPU droplet, with SHA-256-pinned artefacts and a "Benchmark Gap" audit.
DSC-3 beats matched-compute-intensity SA-only on every fully-connected MaxCut cell we measured up to N = 10,000 vertices — over 2× past D-Wave Advantage2's 4,400-qubit embedding ceiling. σ-error bars confirm the advantage is many standard deviations from zero on every cell.
D-Wave Advantage2 capex sits at $10–15M list; DSC-3 runs on a $1.57/hour cloud droplet (or a ~$5K consumer workstation). The ratios on capex, $/solve, power, and energy span 10² to 10⁶.
Across all six benchmarks we examined, every D-Wave reference is missing at least one of four reproducibility artefacts (instances, per-instance wall-times, classical baseline pipeline, quantum/classical work split). This paper releases all four for every benchmark it runs.
DSC-3 reaches a one-million-spin 3D ±J ground-state approximation on a $1.57/hour cloud droplet (
| Axis | D-Wave Advantage2 | DSC-3 (this work) | Ratio |
|---|---|---|---|
| Max embeddable problem size | 4,400 qubits | 1,000,000 (droplet, n=4 seeds) | ~227× |
| Hardware capex / hourly | $10–15M list | $1.57/hour droplet | 10⁴–10⁵× |
| Power continuous | 12.5 kW | 0.30 kW | 42× |
| $/solve at N=1,728 | $0.05–$1.30 (Leap floor) | $0.024 | 10²–10⁵× |
| 3D EA quality vs. literature | sampling, not GS | Hartmann ±1% (L ≤ 40) | matched |
| MaxCut Δ vs SA at N=10,000 | not embeddable | +0.13–0.20%, σ ≤ 0.02% | DSC-3 only |
| Cryptanalysis (SHA/AES/RSA/GNFS) | not addressed | production encoders | DSC-3 only |
| Quantum-coherent sampling | yes (Science 2025) | no (classical engine) | D-Wave only |
main.pdf— the paper, 40 pages, ~9 figures, ~22 tablesmain.tex— full LaTeX sourceresults/*.json— every measured datapoint cited in the paper (SHA-256 manifest in Appendix E ofmain.pdf)figures/*.pdf— every plotted resulttables/*.tex— generated LaTeX tables (regenerable viaaggregate_results.py)make_plots.py— regenerates all figures fromresults/*.jsonaggregate_results.py— regenerates all tables fromresults/*.jsonrun_*.sh/run_*.bat— the exact commands used on the droplet + workstationAUDIT*.md,PLAN*.md— methodology audit trail and operational-section drafting notes
| Bench | D-Wave reference | DSC-3 result | Fidelity |
|---|---|---|---|
| B1 3D ±J Ising spin glass | King et al. Science 2025 (sampling, N≈5000) | Ground-state arg-min, N ≤ 10⁶, ±1% Hartmann at L≤40 | matched-class |
| B2 Currency arbitrage | Cococcioni et al. 2025 (Advantage2 Prototype 2.6) | Hamiltonian-cycle variant, 100% feas at N≤8, +5–22% recovered profit | matched-class |
| B3 Stride 45-instance + extension | Booth et al. 2024 (10× metaheuristic claim) | Matched-spec ensembles + N=5k/10k beyond-embedding probe with σ-error bars | matched-spec |
| B4 SCM (5 verticals) | SCM survey 2025–2026 (12–18% cost reduction) | Uncapacitated Facility Location (1 of 5); +5–30% gap to exact DP | partial |
| B5 Drug discovery + PoQW | JT/D-Wave LLM-molecular-generation; PoQW conceptual | Fragment-selection sub-problem (0–5% gap to DP); r=4 SHA-256 preimage | matched-class / functional-class |
| B6 Cryptanalysis differentiator | no D-Wave publication | SHA-256, AES, RSA-256 Boneh–Durfee, GNFS Phase C+ encoders | capability-only |
A pattern across all six benchmarks: every D-Wave reference is missing ≥1 of four reproducibility artefacts (instances, per-instance wall-times, classical baseline pipeline, quantum/classical work split). This paper releases all four for every benchmark it runs, with SHA-256 manifest in Appendix E. See §15.4 for the full synthesis.
Every numerical claim in the paper traces to a results/*.json file. SHA-256 digests are pinned in Appendix E of main.pdf. To verify:
sha256sum results/*.jsonThen compare against the manifest in Appendix E.
The full pipeline (build + run + aggregate + plot + compile) is described in Appendix A of main.pdf. In short, given the isomorphic-engine Rust crate built with --features gpu,full,tsp:
# B1 3D Ising production preset, L = 4..20
./target/release/examples/dwave_b1_tfim_spin_glass \
--L 4,6,8,10,12,14,16,18,20 --seeds 0,1,2,3 --preset production \
--with-gpu --sa-baseline \
--out results/b1_full.json
# B3 Stride 45-instance + N=5k/10k beyond-embedding
./target/release/examples/dwave_b3_stride \
--seeds 0,1,2,3 --preset quality --with-gpu --gpu-batch 4 \
--maxcut-sizes 20,40,60,100,200,500,1000,2000 \
--out results/b3_gpu_batched.json
./target/release/examples/dwave_b3_stride \
--only maxcut --maxcut-sizes 5000,10000 \
--seeds 0,1,2 --preset production --with-gpu --gpu-batch 4 \
--out results/b3_maxcut_xlarge.jsonThen:
python aggregate_results.py # JSON → LaTeX tables
python make_plots.py # JSON → matplotlib PDFs
pdflatex main.tex # Rebuild paper.pdfBryan W. Daugherty¹, Gregory Ward¹, Shawn Ryan¹ ¹Origin Neural — https://originneural.ai
Three paths, in order of effort:
| Path | Effort | What you get |
|---|---|---|
| Live REST endpoint — https://dsc3.originneural.ai/ | Zero install | POST /v1/solve and POST /v1/mega-benchmark accept QUBO / Ising problems up to N = 5 × 10⁸ spins. Same engine binary as this paper. |
| Replicate the paper | ~30 min on a comparable GPU droplet | Follow Appendix A of main.pdf or the run_*.sh scripts in this repo. Verify your SHA-256s against Appendix E. |
| Adapt to your workload | Hours–days | Map your problem to Ising/QUBO using one of 16 encoders in the parent isomorphic-engine repo; run the same DSC-3 ensemble; compare to your existing classical solver at matched compute budget. |
- 🌐 Browsable landing page: https://originneuralai.github.io/DSC3-DWave-Comparison-2026/
EXECUTIVE_SUMMARY.md— one-page summary for CTOs / procurement officersCHANGELOG.md— version historyCONTRIBUTING.md— how to file a falsification attempt, reproduction failure, or factual correctionFAQ.md— pre-emptive answers to reviewer and procurement questionsresults/README.md— artefact taxonomy (load-bearing vs auxiliary)
Zenodo DOI: 10.5281/zenodo.20192275
@misc{daugherty_ward_ryan_dsc3_dwave_2026,
author = {Bryan W. Daugherty and Gregory Ward and Shawn Ryan},
title = {A Reproducible Classical Reference for D-Wave Advantage2's
2024--2026 Industrial Benchmarks: Ground-state 3D $\pm J$
Ising at $N=10^{6}$ Spins on a \$1.57/Hour GPU Droplet},
publisher = {Zenodo},
year = {2026},
month = may,
doi = {10.5281/zenodo.20192275},
url = {https://doi.org/10.5281/zenodo.20192275},
note = {GitHub: \url{https://github.com/OriginNeuralAI/DSC3-DWave-Comparison-2026}}
}- That classical methods invalidate D-Wave's sampling-class quantum supremacy demonstration. We solve a different problem (ground-state search).
- That DSC-3 outperforms every classical algorithm at every scale. Concorde and LKH3 dominate pure TSP; exact DP dominates Knapsack; Goemans–Williamson dominates MaxCut at small–medium scale.
- That this paper exhausts the relevant benchmarks. Four of five SCM verticals, the drug-discovery LLM-training pipeline, and reverse-annealing protocols remain future work.
- Companion paper: DSC-3 Benchmark Suite: 500 Million Spins on a Single GPU — separate single-instance capability demonstration at N = 5×10⁸. Live demo at https://dsc3.originneural.ai/.
Source code and result JSON: see the parent isomorphic-engine repository.
LaTeX source and PDF: released for review and citation; please contact authors before re-using figures or quoting verbatim.
Bryan W. Daugherty, Gregory Ward, Shawn Ryan — Origin Neural https://originneural.ai