DSC-3 vs D-Wave Advantage2 — Industrial Benchmark Comparison (2024–2026)

v0.15.1-paper CC-BY-4.0 40 pp SHA-256 manifest

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 of D-Wave's published reproducibility deficits.

Open the paper (PDF, 40 pp) View on GitHub Live demo

Headline findings

Axis	D-Wave Advantage2	DSC-3 (this work)	Ratio
Max embeddable problem	4,400 qubits	1,000,000 (droplet, n=4 seeds)	~227×
Hardware capex	$10–15M list	$1.57/hour droplet	10⁴–10⁵×
Power continuous	12.5 kW	0.30 kW (GPU TDP)	42×
$/solve at N=1,728	$0.05–$1.30 (Leap floor)	$0.024	10²–10⁵×
3D EA quality vs Hartmann (2001)	sampling, not GS	±1% at L≤40 production preset	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

Six benchmarks covered

Bench	D-Wave reference	DSC-3 result
B1 3D ±J Ising	King et al. Science 2025 (sampling, N≈5000)	Ground-state arg-min, N ≤ 10⁶, ±1% Hartmann at L≤40
B2 Currency arbitrage	Cococcioni et al. 2025 (Advantage2 Prototype 2.6)	Hamiltonian-cycle variant; 100% feas at N≤8; +5–22% recovered profit
B3 Stride 45-instance + extension	Booth et al. 2024 (10× metaheuristic claim)	Matched-spec ensembles + N=5k/10k beyond-embedding probe (σ error bars)
B4 SCM (5 verticals)	SCM survey 2025–2026 (12–18% cost reduction)	UFL (1 of 5); +5–30% gap to exact DP
B5 Drug discovery + PoQW	JT/D-Wave LLM-molecular-generation; PoQW conceptual	Fragment-selection sub-problem (0–5% gap); r=4 SHA-256 preimage
B6 Cryptanalysis differentiator	no D-Wave publication	SHA-256 / AES / RSA-256 BD / GNFS Phase C+ encoders

Read the paper inline

Quick navigation

FAQ — pre-emptive answers to reviewer and procurement questions
Contributing — how to file a falsification attempt or factual correction
Result artefacts — load-bearing JSONs with SHA-256 manifest
Open an issue — falsification / reproduction / correction templates

Authors: Bryan W. Daugherty, Gregory Ward, Shawn Ryan — Origin Neural — originneural.ai
Companion paper: DSC-3 Benchmark Suite: 500 Million Spins on a Single GPU.