Geospatial Benchmarks

Batch geodesic distance benchmarks comparing NumKong against the geo and geopy libraries for Haversine and Vincenty formulas.

Rust

Library	Precision	MP/s
Haversine
numkong::haversine	f32 → f32	491.98
numkong::haversine	f64 → f64	149.72
serial baseline	f32 → f32	137.83
geo::Haversine distance	f32 → f32	136.96
serial baseline	f64 → f64	94.33
geo::Haversine distance	f64 → f64	92.48
Vincenty
numkong::vincenty	f32 → f32	71.64
serial baseline	f32 → f32	18.20
numkong::vincenty	f64 → f64	13.73
serial baseline	f64 → f64	6.47
geo::Vincenty distance	f64 → f64	2.76

Python

Library	Precision	MP/s
numkong.haversine	f32 → f32	444.38
numkong.haversine	f64 → f64	132.85
numkong.vincenty	f32 → f32	65.89
numkong.vincenty	f64 → f64	11.93
geopy.distance.great_circle	f64 → f64	0.47
geopy.distance.geodesic	f64 → f64	0.03

Run It

Rust

# Default 2048 coordinate pairs
cargo bench --bench bench_geospatial --features bench_geospatial

# Smaller 256 coordinate pairs
NUMWARS_DIMS=256 \
cargo bench --bench bench_geospatial --features bench_geospatial

# Focus on one metric
NUMWARS_FILTER="geospatial/haversine/f32" \
cargo bench --bench bench_geospatial --features bench_geospatial

Python

# Run the Python suite
uv run --with numkong,numpy,geopy,tabulate python geospatial/bench.py --count 2048