Public benchmarks
Our nesting densities, measured on our corpus of edge cases — open method, reproducible numbers.
Method
Every case is a standard product job (multi-sheet BPP, one –X direction, nesting-in-cutouts on). Material density is measured on the verified report: area of the placed parts over the area of the used sheets. Physical validation checks zero overlap, every part inside its sheet, and a minimal gap no smaller than the requested spacing.
Results by case
| Case | Geometry | Sheets | Spacing | Parts | Material density | Physics |
|---|---|---|---|---|---|---|
| T-A | Ringed part (100×100 mm, Ø35 cutout) + 4-lobed fan nesting inside the cutouts — the reference corpus (100 + 800 parts) | 2 × 1000×1000 | 0.1 mm | 900/900 | 55.4% | verified |
| T-B | Three close rectangle classes (300×200, 250×180, 120×90) — pure lattice packing | 3 × 1500×1000 | 2 mm | 80/80 | 84.4% | verified |
| T-C | Non-convex L (200×200) and U (240×200) parts | 2 × 1200×1000 | 1 mm | 60/60 | 60% | verified |
| T-D | Long thin parts (900×40) + 300 small rectangles — one dominant orientation | 2 × 1000×1000 | 1 mm | 330/330 | 76.5% | verified |
| T-E | Ringed part + fan with free rotations every 30° (12 orientations) | 3 × 1000×1000 | 0.1 mm | 460/460 | 61.5% | verified |
| T-F | 90 rectangles 200×150 on two sheet formats (cost ∝ area) — tight stock | 1000×1000 + 2000×1000 | 1 mm | 90/90 | 90% | verified |
| T-G | One near-full-sheet part (950×950) + 200 small parts around it | 2 × 1000×1000 | 0.5 mm | 201/201 | 60.1% | verified |
| T-H | Single class of 200 identical rectangles (120×80) — annealing at steady state | 3 × 1000×1000 | 1 mm | 200/200 | 64% | verified |
| T-J | Same as A but 1000 parts on ONE 1000×2000 sheet at 4 mm — over capacity on purpose | 1 × 1000×2000 | 4 mm | refused before compute | — | honest refusal < 1 s |
| T-K | 1000 parts at 2.4 mm on two 1000×1000 sheets — right at the measured packing limit | 2 × 1000×1000 | 2.4 mm | 1000/1000 | 58.5% | verified |
“Verified” = zero measured overlap, every part inside its sheet, minimal gap never below the requested spacing.
Robustness beyond the numbers
The internal torture corpus also runs a free-form ESICUP instance (shirts, 12 orientations) and a 12-orientation variant of case A. Those live in their own normalized units, so their density is not comparable to the product cases above — they exist to prove the engine never overlaps, never drops a part, and never returns an uncuttable layout on unfamiliar geometry.
What we do NOT claim: these densities are ours, on our corpus, at the stated budgets — not a third-party certification, and no benchmark makes one engine universally “best”. Every number above was produced by the exact Docker image deployed in production at the date shown, and re-verified after each engine release.