Case Study  ·  The Davis Manifold in Action

Curvature-Guided
Wavefront Execution
on Sudoku

GPU-accelerated constraint satisfaction guided by intrinsic Riemannian curvature. A geometry-aware solver that crushes the world's hardest Sudoku at 1,226× speedup over Python CPU — demonstrating that constraint satisfaction is fundamentally a geometry problem.

DOI: 10.5281/zenodo.18511755 · Patent Pending · February 2026
40,128×
faster than Yann LeCun's Kona 1.0 Energy-Based Model
Performance Summary
GPU throughput~270K puzzles/sec
Per-instance latency3.7 µs
Speedup vs. Python CPU1,226×
Hardest-known solved11 / 11
Avg solve time (v4)7.8 ms

Computational resources should flow toward regions of highest curvature — where they have the greatest marginal impact on constraint resolution. Classical GPU approaches treat every vertex equally. The Davis manifold tells you where the hard work is, before the work begins.

Three-Phase GPU Pipeline

The solver implements a three-phase CUDA pipeline derived from the Davis Field Equations. A trichotomy parameter Γ automatically classifies each instance by its geometric complexity and routes it to the optimal phase combination — no manual algorithm selection required.

I
Wavefront Propagation
Bitmask-based arc consistency and hidden-single detection, parallelized across warps. Converges in a single ballot pass per iteration.
II
Manifold Relaxation
Softmax-parameterized probability distributions minimized via curvature-adaptive gradient descent on the Davis energy functional. Fully data-parallel.
III
Curvature-Directed DFS
Iterative-deepening depth-first search with branch selection by information value and holonomy-based dead-branch pruning.

Classical heuristics — MRV, degree heuristic, checkerboarding — all emerge as degenerate cases of the curvature-guided ordering when specific curvature weight components are set to zero.

Intrinsic Curvature as Scheduling Signal

The constraint graph induces a discrete Riemannian manifold in the sense of Regge calculus — not a metaphor, but a genuine geometric structure with a metric tensor, connection, and holonomy group. The local curvature field measures three independent scalar invariants of the constraint fiber bundle.

Local Curvature
$$K_{\text{loc}}(v) = w_s \cdot \sigma(v) + w_r \cdot \rho(v) + w_c \cdot \kappa(v)$$
$\sigma$ = saturation (boundary curvature) · $\rho$ = scarcity (fiber dimension) · $\kappa$ = coupling norm (connection rigidity)
Information Value
$$V(v) = \frac{K_{\text{loc}}(v) + \displaystyle\sum_{u \in N(v)} K_{\text{loc}}(u)}{\left|D(v)\right|}$$
Curvature-weighted variable ordering — selects the vertex whose resolution propagates the most information globally.
Davis Energy Functional
$$E[\gamma] = \lambda_1 \int_\gamma ds + \lambda_2 \int_\gamma K_{\text{loc}}(s)\, ds + \lambda_3 \int_\gamma \left\| \mathrm{Hol}_\gamma(s) - I \right\| ds$$
Path length + curvature-weighted complexity + holonomy deficit. A geodesic of this functional resolves constraints with minimal total effort.

The trichotomy parameter $\Gamma = \dfrac{m \cdot \tau}{\hat{K}_{\max} \cdot \log |S|}$ classifies instances by the ratio of assigned structure to geometric complexity, automatically gating them into the optimal phase combination.

Performance Comparison

Evaluated on 15-clue extreme Sudoku (66 empty cells, Γ ≈ 0.19) — the hardest class of well-posed puzzles. All 11 hardest-known instances solved in under 9ms on a consumer laptop GPU (NVIDIA RTX 5070, Blackwell architecture).

GPU Performance · Hardest Known (v4)
AI Escargot8.5 ms
Inkala 20108.1 ms
Golden Nugget8.5 ms
Easter Monster8.5 ms
17-clue Coloin6.4 ms
CUDA kernel time, excluding host transfer · Avg 7.8 ms across 11 puzzles
Speedup vs. Baseline Solvers
vs. Kona 1.0 EBM (LeCun)40,128×
vs. DLX (Dancing Links)3.8×
vs. Python CPU1,226×
Batch throughput (10K)278K/s
Batch throughput (65K)268K/s
Kona: 313ms avg, 96.2% accuracy · Davis: 7.8ms, 100%
Davis GPU (v4)
7.8 ms
DLX
29.6 ms
Kona 1.0 EBM
313 ms
Python CPU
9,565 ms
Hardware
GPURTX 5070 Laptop
ArchitectureBlackwell (SM 12.0)
CUDA Cores4,608
Memory8 GB GDDR7

Reproduce the Results

$ git clone https://github.com/nurdymuny/sudoky-energy.git
$ cd sudoky-energy
$ pip install -e .
$ python scripts/benchmark_hardest.py

Sudoku is just one demonstration. The Davis Manifold framework applies to any finite-domain constraint satisfaction problem.
See the full framework →