Benchmark snapshot — 2026-08-29 · 9b200075

Official NumSharp-vs-NumPy comparison with Managed C# / OpenBLAS profiles and complementary subsystems, persisted for provenance. Auto-generated by benchmark/scripts/snapshot_history.py.

Provenance


Run timestamp	`20260829-005949`
Git HEAD	`9b200075` — docs(skill): document master-code-data dashboard-data deliv…
Working tree	DIRTY — benchmarked HEAD + uncommitted changes: test/NumSharp.Tests.Build.Analyzer/NumSharp.Tests.Build.Analyzer.csproj
Date	2026-08-29

Environment


CPU	13th Gen Intel Core i9-13900K
OS	Windows 11 (10.0.26200)
.NET SDK	10.0.101 (net10.0, Release)
Python	3.12.12
NumPy	2.4.2

Convention

Ratio = NumPy_ms ÷ NumSharp_ms (NPY/NS) → >1.0× = NumSharp faster (higher is better).

Methodology

C#: BenchmarkDotNet, OfficialBenchmarkConfig — InProcessEmit toolchain, 50 measured iterations / 5 warmup, iteration time capped at 25 ms. MemoryDiagnoser on.
NumPy: time-budgeted per op — a ~200 ms window (a >20 ms/call op runs exactly 100 times), 10 warmup (warm long-lived interpreter).
Timing basis: best window (min) of each side's per-case sweep — min-based harnesses run each case to a ~200 ms time budget (a >20 ms/call op runs exactly 100 times); the C# op-matrix keeps BenchmarkDotNet's 50 iterations. merge-results.py compares the mins; per-case means are retained in the JSON (numpy_mean_ms/numsharp_mean_ms) as interference/tail diagnostics.
Sizes: 1,000 / 100,000 / 10,000,000 elements. Same seeds both sides.
Join keyed on (op, dtype, N, scenario); effective timing is the fastest valid backend profile.
Availability: MissingBackendException and NotSupportedException are recorded, never timed.
Backend profiles: Managed C# and OpenBLAS are merged into the canonical matrix.
Subsystems appended to benchmark-report.md: NDIter, Layout, Operand, Cast, Fusion.

NDIter: HEADLINE — operation matrix: 1.46× geomean · 69%🕐 of NumPy's time · 108 win / 57 lose over 165 cells

Cast: 381 / 1568 comparable cells lag (<1.0); 1187 win (≥1.0).

Files

file	what
`benchmark-report.md`	backend-aware op matrix + appended complementary subsystems
`benchmark-report.json` / `.csv`	combined backend-aware machine-readable / spreadsheet form
`benchmark-report.managed.json` / `.openblas.json`	separate profile measurements using the same schema
`numpy-results.json`	raw NumPy timings (merge input)
`nditer_results.*` + `cards/`	iterator benchmark sheet + summary cards (ops/cat)
`layout_/operand_/cast_/fusion_/openblas_results.*`	complementary subsystem sheets and targeted backend profiles