Table of Contents
NumSharp performance lab

Benchmark Dashboard

A compact operating view of the NumSharp vs NumPy benchmark suite: 18 official op suites, all supported dtypes where applicable, three cache tiers, and six complementary scans for iterator, layout, operand, cast, fusion, and native OpenBLAS/LAPACK behavior.

Snapshot 2026-08-24 Commit 938d0449 NumPy 2.4.2 Ratio NumPy / NumSharp
API coverage
98.7%

455 / 461 benchmarkable NumSharp-compatible APIs have benchmark evidence

Effective geomean
—

Best valid implementation selected per comparable cell

Backends Supported
2/2

Managed C# Native Backend
OpenBLAS Backend

Cast subsystem
1,187

wins out of 1,568 comparable cast cells

Operation cells

Legend & How To Read

Benchmark row
Each result is one operation x dtype x size timing cell from the latest benchmark snapshot; rollups use comparable measured rows.
Timing fields
Ratios come from raw NumPy_ms and NumSharp_ms timings. A value above 1.00x is a NumSharp win; below 1.00x means NumPy was faster.
Drill-down
Click Status Mix segments, Suite Scoreboard rows, and Dtype Heatmap cards to open the first 25 best and worst benchmark rows; use Load More to reveal the rest.

Cell

One benchmark row: operation, dtype, and size tier. The dashboard uses best timed runs after warmup. Allocation-heavy families retain the 10M tier label while both NumPy and NumSharp use a symmetric 1M physical-work cap; 1M is not a separate dashboard tier.

Reading Ratios

Ratio is NumPy / NumSharp. Higher is better: 3.00x means NumSharp took 10s while NumPy took 30s; 0.50x means NumSharp takes about twice as long.

Performance Bands

Faster: 1.05x and above Close: 0.80x to 1.05x Slower: 0.33x to 0.80x Much slower: below 0.33x No data: pending C# measurement

Status Mix

Credible rows are grouped by NumPy / NumSharp ratio; negligible and no-data cells are separate

4 x100+ faster 933 faster, 1.05-100x 523 close, 0.80-1.05x 954 slower, 0.33-0.80x 516 much slower, <0.33x Negligible Failed 63 no data

Suite Scoreboard

Geomean across credible rows. Parity marker is 1.0x.

Statistics 2.24x 39 / 10
Reduction 1.81x 385 / 110
Broadcasting 1.10x 3 / 0
Sorting 1.08x 23 / 13
Creation 1.04x 23 / 21
Unary 0.83x 93 / 120
Arithmetic 0.80x 145 / 197
Selection 0.76x 2 / 4
Logic 0.68x 14 / 25
Comparison 0.64x 17 / 31
Linear algebra 0.57x 2 / 6
Bitwise 0.54x 45 / 68
Manipulation 0.40x 1 / 1

Dtype Heatmap

Credible operation-matrix rows by dtype and cache tier

uint82.00x61 rows
int81.84x63 rows
uint161.44x65 rows
int161.39x66 rows
float161.28x224 rows
uint321.09x65 rows
complex1281.04x64 rows
float321.01x233 rows
float640.95x257 rows
int320.93x111 rows
int640.75x118 rows
uint640.74x63 rows
bool0.30x8 rows

Subsystem Signals

Result models that the op matrix cannot express

NDIter

1.46x

Iterator operation geomean across 165 comparable cells: 108 wins and 57 losses, with no sections excluded by the access-violation policy.

Layout

0.54-9.24x

Published geomeans by layout span 0.54x to 9.24x: large copy and elementwise paths win, while strided/broadcast reductions and decimal sums trail.

Cast

1,187 wins

The full src-to-dst astype grid is broadly ahead. Remaining lag clusters around same-type diagonal copy and bool conversion cases.

Fusion

4.12x

The best fixed expression speedup for fused np.evaluate over NumSharp's unfused chain; broadcast fusion reaches 4.10x.

Function Explorer

Scores use credible rows only; negligible and no-data cells remain available in the raw tables

Loading np.* function performance surface...

Optimization Priorities Recommended

Current optimization priorities from the latest snapshot

  1. Shift kernels: vectorize 100K int left/right shift.
  2. Bool bitwise: lift invert, &, |, ^.
  3. Decimal reduce: fix broadcast/sliced sum axis cliffs.
  4. i32 broadcast: repair stride-0 axis sum.
  5. f64 sum: close the 100K reduction gap.
  6. Float predicates: accelerate 100K isnan/isinf/isfinite.
  7. f32 add/mul: improve 100K scalar/literal paths.
  8. f64 add/mul: improve 100K scalar/literal paths.
  9. Float abs/neg: remove 100K C/strided cliffs.
  10. Rounding: tighten 100K f32/f64 floor/ceil/trunc.
  11. Exp/log: reduce f32/f64 mid-tier overhead.
  12. f16 unary: lift sign/math scalar fallback.
  13. Int mean: improve 10M int64/uint64 mean.
  14. Linear algebra: revisit large f64 matmul/dot.
  15. Flatten: fix the 100K copy/cast trough.
  16. Astype small: reduce scalar/1K setup cost.
  17. Ravel T: close transpose ravel mid-size gap.
  18. less->bool: specialize bool output loops.
  19. Index math: speed scalar/1K unravel/ravel-multi.
  20. NDIter copy: raise 100K copy/cast geomean.
  21. Chunk width: optimize tiny inner width dispatch.
  22. f16 operands: improve strided/reversed/broadcast cases.
  23. Cast diagonal: speed same-type copy cells.
  24. Cast bool: clean remaining * -> bool losses.
  25. Fusion: broaden fused-expression coverage.

Full Reports

Detailed tables remain available for traceability

Edit this page