Packed GEMM-style matrix multiplication benchmarks comparing NumKong against faer, matrixmultiply, ndarray, nalgebra, and NumPy.

Rust

Library	Precision	GSO/s
numkong::try_dots_packed_into	i8 → i32	2783.00
numkong::try_dots_packed_into	u8 → i32	2784.10
numkong::try_dots_packed_into	bf16 → f32	1250.80
numkong::try_dots_packed_into	f16 → f32	1249.70
faer::linalg::matmul::matmul	f32 → f32	117.50
matrixmultiply::sgemm	f32 → f32	116.77
ndarray::dot	f32 → f32	117.51
nalgebra::gemm	f32 → f32	118.72
numkong::try_dots_packed_into	f32 → f64	197.79
numkong::try_dots_packed_into	e4m3 → f32	495.25
numkong::try_dots_packed_into	e5m2 → f32	746.22
numkong::try_dots_packed_into	e2m3 → f32	1355.00
numkong::try_dots_packed_into	e3m2 → f32	693.43

Python

Library	Precision	GSO/s
numkong.dots_packed	i8 → i32	2621.97
numkong.dots_packed	bf16 → f32	1142.19
numpy.matmul	f32 → f32	1854.27
numkong.dots_packed	f16 → f32	1134.69
numkong.dots_packed	f32 → f64	194.15

Run It

Rust

``$\text{bash}

\text{Default} 2048 \times 2048 \times 2048 \text{workload}

\text{cargo} \text{bench} --\text{bench} \text{bench_dots} --\text{features} \text{bench_dots}

\text{Smaller} 512 \times 512 \times 512 \text{workload}

\text{NUMWARS_DIMS_WIDTH}=512 \text{NUMWARS_DIMS_HEIGHT}=512 \text{NUMWARS_DIMS_DEPTH}=512
\text{cargo} \text{bench} --\text{bench} \text{bench_dots} --\text{features} \text{bench_dots}

\text{Focus} \text{on} \text{float32}

\text{NUMWARS_FILTER}="\text{dots}/\text{f32}"
\text{cargo} \text{bench} --\text{bench} \text{bench_dots} --\text{features} \text{bench_dots} $``

Python

# Default 2048×2048×2048 workload, float32 only
python dots/bench.py --filter 'dots/numpy/f32/2048x2048x2048'