Benchmark snapshot — 2026-06-29 · 2d16f477

Official NumSharp-vs-NumPy 3-size comparison + the five matrix subsystems, persisted for provenance. Auto-generated by benchmark/scripts/snapshot_history.py.

Provenance

Run timestamp	20260629-083502
Git HEAD	2d16f477 — test(drawing): edge-case and data-loss/corruption coverage …
Working tree	clean (HEAD exactly)
Date	2026-06-29

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: 50 timed iterations / 10 warmup per op (warm long-lived interpreter).
• Sizes: 1,000 / 100,000 / 10,000,000 elements. Same seeds both sides.
• Join keyed on (op, dtype, N).
• Subsystems appended to benchmark-report.md: NDIter, Layout, Operand, Cast, Fusion.

Headline — op-matrix geomean by size (NPY/NS, >1 = NumSharp faster)

Size	geomean	%NumPy🕐	✅ / 🟡 / 🟠 / 🔴
1,000	1.13x	89%	116 / 58 / 35 / 11
100,000	0.98x	102%	295 / 135 / 124 / 35
10,000,000	1.39x	72%	427 / 136 / 24 / 4

Overall op-matrix: 1851 ops | ✅ 838 | 🟡 329 | 🟠 183 | 🔴 50 | ▫ 388 | ⚪ 63.

NDIter: HEADLINE — operation matrix: 1.20× geomean · 83%🕐 of NumPy's time · 77 win / 53 lose over 130 cells

Cast: 118 / 1568 comparable cells lag (<1.0); 1450 win (≥1.0).

Files

file	what
benchmark-report.md	op-matrix (per-(op,dtype,N) ratio) + appended NDIter/Layout/Operand/Cast/Fusion
benchmark-report.json / .csv	unified machine-readable / spreadsheet form
numpy-results.json	raw NumPy timings (merge input)
nditer_results.* + cards/	iterator benchmark sheet + README cards
layout_/operand_/cast_/fusion_results.*	the four matrix-subsystem sheets

Raw BenchmarkDotNet per-class JSON (~tens of MB) is not persisted here (regenerable). Reproduce with python benchmark/run_benchmark.py.