Similarity Benchmarks

March 23, 2026 · View on GitHub

Pairwise vector-vector distance and dot-product benchmarks comparing NumKong against scalar baselines and BLAS-backed libraries.

Rust

Library	Precision	GSO/s
Dot
`numkong::Dot`	u8 → u32	37.34
`numkong::Dot`	i8 → i32	37.41
`numkong::Dot`	bf16 → f32	14.83
serial code	f32 → f32	13.86
`ndarray::dot`	f32 → f32	12.77
`nalgebra::dot`	f32 → f32	12.95
serial code	u8 → u32	19.28
`numkong::Dot`	f32 → f64	2.50
serial code	i8 → i32	19.70
serial code	bf16 → f32	0.67
`numkong::Dot`	f16 → f32	6.16
`numkong::Dot`	f64 → f64	2.81
`numkong::Dot`	e4m3 → f32	4.43
`numkong::Dot`	e5m2 → f32	6.08
`numkong::Dot`	e2m3 → f32	26.64
`numkong::Dot`	e3m2 → f32	11.32
`ndarray::dot`	f64 → f64	11.62
`nalgebra::dot`	f64 → f64	11.66
Angular
`numkong::Angular`	f32 → f32	2.34
`numkong::Angular`	f64 → f64	2.30
`numkong::Angular`	i8 → f32	31.86
`numkong::Angular`	u8 → f32	31.86
`numkong::Angular`	f16 → f32	4.81
`numkong::Angular`	bf16 → f32	10.28
`numkong::Angular`	e4m3 → f32	2.19
`numkong::Angular`	e5m2 → f32	5.34
`numkong::Angular`	e2m3 → f32	2.60
`numkong::Angular`	e3m2 → f32	2.60
Euclidean
`numkong::Euclidean`	u8 → f32	34.85
`numkong::Euclidean`	i8 → f32	34.91
`numkong::Euclidean`	bf16 → f32	5.54
`numkong::Euclidean`	f32 → f64	2.50
`ndarray::norm`	f32 → f32	7.61
`nalgebra::norm`	f32 → f32	7.74
serial code	f32 → f32	11.15
serial code	u8 → f32	9.19
serial code	i8 → f32	9.26
serial code	bf16 → f32	0.66
`numkong::Euclidean`	f16 → f32	5.27
`numkong::Euclidean`	f64 → f64	2.60
`numkong::Euclidean`	e4m3 → f32	2.28
`numkong::Euclidean`	e5m2 → f32	5.80
`numkong::Euclidean`	e2m3 → f32	2.74
`numkong::Euclidean`	e3m2 → f32	2.75
`ndarray::norm`	f64 → f64	5.59
`nalgebra::norm`	f64 → f64	5.63
Squared Euclidean
`numkong::SqEuclidean`	f32 → f32	2.65
`numkong::SqEuclidean`	f64 → f64	2.76
`numkong::SqEuclidean`	i8 → f32	38.59
`numkong::SqEuclidean`	u8 → f32	38.80
`numkong::SqEuclidean`	f16 → f32	5.52
`numkong::SqEuclidean`	bf16 → f32	5.70
`numkong::SqEuclidean`	e4m3 → f32	2.29
`numkong::SqEuclidean`	e5m2 → f32	5.87
`numkong::SqEuclidean`	e2m3 → f32	2.76
`numkong::SqEuclidean`	e3m2 → f32	2.76
Hamming
`numkong::Hamming`	u1x8 → u32	45.20
Jaccard
`numkong::Jaccard`	u1x8 → f32	36.88
Kullback-Leibler
`numkong::KullbackLeibler`	f32 → f32	2.87
`numkong::KullbackLeibler`	f64 → f64	0.27
Jensen-Shannon
`numkong::JensenShannon`	f32 → f32	1.47
`numkong::JensenShannon`	f64 → f64	0.17

Python

Library	Precision	GSO/s
Dot
`numkong.dot`	u8 → f32	9.37
`numkong.dot`	i8 → f32	9.81
`numkong.dot`	bf16 → f32	1.06
`numkong.dot`	f32 → f32	2.02
`numkong.dot`	f64 → f32	2.48
`numpy.dot`	f32 → f32	3.81
`numpy.dot`	f64 → f64	3.68
Angular
`numkong.angular`	u8 → f32	10.07
`numkong.angular`	i8 → f32	10.04
`numkong.angular`	bf16 → f32	1.03
`numkong.angular`	f32 → f32	1.98
`numkong.angular`	f64 → f32	2.01
Euclidean
`numkong.euclidean`	u8 → f32	10.24
`numkong.euclidean`	i8 → f32	10.32
`numkong.euclidean`	bf16 → f32	0.92
`numkong.euclidean`	f32 → f32	2.15
`numkong.euclidean`	f64 → f32	2.14
SqEuclidean
`numkong.sqeuclidean`	u8 → f32	9.34
`numkong.sqeuclidean`	i8 → f32	8.60
`numkong.sqeuclidean`	bf16 → f32	0.93
`numkong.sqeuclidean`	f32 → f32	2.02
`numkong.sqeuclidean`	f64 → f32	2.14

Run It

Rust

# Default 2048-dimensional vectors
cargo bench --bench bench_similarity --features bench_similarity

# Smaller 512-dimensional vectors
NUMWARS_DIMS=512 cargo bench --bench bench_similarity --features bench_similarity

# One benchmark group
NUMWARS_FILTER="similarity/angular/f32" \
cargo bench --bench bench_similarity --features bench_similarity

Python

# Default 2048-dimensional pairwise distances
python similarity/bench.py --filter 'angular.*float32'

# Compare probability metrics
python similarity/bench.py --filter 'jensenshannon|kullback' --ndim 1536 --count 128