SPINACHNATIVE PROPAGATION ATLASEXPERIMENT 01 • NOV 2026
NATIVE PROPAGATION · BENCHMARK REPORT

Native propagation
benchmark

Measured runtime and numerical error for native Spinach propagation routes across eight spin-system and pulse fixtures.

Open interactive results ↓
1,204method / grid cases
6,020timed full solves
28native routes at 2 spins
8physical fixtures
17accuracy checkpoints
01 / RESULTS EXPLORER

Runtime and error

Each point represents a measured grid. Select a fixture, error criterion and target to compare routes.

Work–precision

LOG / LOG

Interpretation. The fastest qualifying measured grid is selected for each route. Lines connect measured grids. A missing qualification means the target was not reached on the tested grids, not that the method cannot reach it.
Inspect the measured data
Native methodStepsMedian, msMin–max, msMax errorFinal errorGram defect
02 / COMPARISON SUMMARY

Fastest qualifying route

Select a cell to inspect its fixture and target. The map uses the selected error criterion.

CF4 CF4-3 CF6-5
03 / METHOD CATALOGUE

Native routes

Route identifiers, descriptive names and native Spinach entry points.

28 native routes

Method labels follow the archived native source. “Added in branch” describes implementation history, not mathematical novelty. CF4-3 and CF6-5 are Alvermann–Fehske CF4:3Opt and CF6:5Opt: primary paper ↗.

04 / BENCHMARK PROTOCOL

Methods and limitations

01

Physical model and error

Four reproducible orthonormal complex state columns; a 1 ms pulse; identical spin-chain Hamiltonian and RF. Drift includes site offsets and nearest-neighbour ZZ interactions.

Accuracy is the Frobenius difference divided by the initial block norm of 2. No phase alignment or renormalisation. Timing requests the final state only.

02

Timing protocol

Waveform sampling, generator construction, derivatives and onsite jets, exponential actions, endpoint processors and physical decoding.

Static setup and reference integration are separate. Five interleaved repeats per point are recorded. Min–max ranges are descriptive, not confidence intervals.

03

Validation

625 correctness comparisons and 36 amplitude/frequency checks. All numerical calls completed. An independent ode113 reference was refined in tolerance and step limit.

4.47 × 10⁻¹³largest reference refinement difference
8.69 × 10⁻¹⁴largest long-run Gram defect, six additions
Coverage and interpretation limits
  • Finite grids: 16–512 steps at 2 and 6 spins; 16–256 at 8 spins. Eight-spin timing has nonflat and chirped pulses only. The three amplitude/frequency routes are timed at 2 spins only.
  • ODC4 is generally fourth order with interactions. Sixth-order onsite cancellation in the uncoupled case is not a general sixth-order claim. The implementation pays for full-basis embedded onsite commutators; no reduced local-action variant was timed.
  • Representation: coherent Hamiltonian wavefunction propagation. Four state columns at 6 and 8 spins do not establish a worst-case operator-norm bound. No dissipative, nonlinear, GPU, or all-formalism performance claim is made.
  • Statistics: five-repeat medians and ranges, one hardware configuration, two sessions on the same machine. Reference refinement is a consistency check, not a rigorous error certificate.
  • Regression scope: seven focused regression groups passed. The full repository suite was not run. Three AGROW diagnostics remain in untimed benchmark/check bookkeeping.
Environment, provenance and reproduction

Intel Xeon w9-3595X, 60 physical / 120 logical cores. Windows Server 2025 Standard. MATLAB 26.2.0.3320248 (R2026b) Prerelease Update 2. One native process worker and one computational thread after construction. The worktree used ReFS; Spinach recommends NTFS for locking and performance.

Source branch: research/odc4-native-benchmark-20261114. Source commit: dc0f6c3d384c1f7ca3b38e5fa09d835f1ddb7676. Base: d09596d5171cda3218147f0ae6d42034cf14528f. No merge or source-repository publication is implied by this dashboard.

The initial session ended after 779 complete rows due to an output-stream interruption. The continuation preserved those rows exactly and resumed missing full grid blocks. The failed singleCompThread startup is excluded from timing results.

Run native_validate, native_af_check and native_quality before native_benchmark, following the original benchmark README and source archive. Full source and execution evidence remain in the original benchmark delivery.

Download benchmark data

Measurements, target selections and validation records.

Full workbook ↓1,204 raw rows ↓1,254 target selections ↓48 winner records ↓625 validation rows ↓36 AF checks ↓Original report ↓
05 / PLANNED WORK

Example-workload
benchmarks

A future benchmark will measure complete Spinach example workloads, including observable error, runtime and method coverage.

PLANNED · NOT YET MEASURED

Not yet measured

Only the native propagation experiment is reported here. Example workloads will be reported separately with a defined reproducibility protocol.

No example-workload timings are reported.