Mesh Benchmarks

March 23, 2026 · View on GitHub

3D point-cloud alignment benchmarks comparing NumKong RMSD and Kabsch against scalar and nalgebra implementations.

Rust

Library	Precision	MP/s
RMSD
`numkong::MeshAlignment::rmsd`	f16 → f16	2864.47
`numkong::MeshAlignment::rmsd`	bf16 → bf16	2861.70
`numkong::MeshAlignment::rmsd`	f64 → f64	1859.32
`numkong::MeshAlignment::rmsd`	f32 → f32	1626.67
nalgebra-based RMSD	f32 → f32	634.04
Kabsch
`numkong::MeshAlignment::kabsch`	f16 → f16	696.00
`numkong::MeshAlignment::kabsch`	bf16 → bf16	691.01
`numkong::MeshAlignment::kabsch`	f32 → f32	396.52
`numkong::MeshAlignment::kabsch`	f64 → f64	331.70
nalgebra-based Kabsch	f32 → f64	283.16
Umeyama
`numkong::MeshAlignment::umeyama`	bf16 → bf16	673.50
`numkong::MeshAlignment::umeyama`	f16 → f16	614.06
`numkong::MeshAlignment::umeyama`	f32 → f32	376.48
`numkong::MeshAlignment::umeyama`	f64 → f64	325.16
nalgebra-based Umeyama	f32 → f64	255.14

Python

Library	Precision	MP/s
`numkong.rmsd`	f64 → f64	1311.77
`numkong.rmsd`	f32 → f64	1228.00
`numkong.kabsch`	f32 → f64	360.08
`numkong.umeyama`	f32 → f64	327.01
`numkong.umeyama`	f64 → f64	296.67
`numkong.kabsch`	f64 → f64	285.81
`numpy-based RMSD`	f32 → f64	124.48
`numpy-based RMSD`	f64 → f64	117.30
`biopython SVDSuperimposer (Kabsch)`	f32 → f64	2.88
`biopython SVDSuperimposer (Kabsch)`	f64 → f64	2.92

# Default 2048-point clouds
python mesh/bench.py

# Smaller 256-point clouds
python mesh/bench.py --count 256

# Focus on one operation
python mesh/bench.py -k "kabsch"

Run It (Rust)

# Default 2048-point clouds
cargo bench --bench bench_mesh --features bench_mesh

# Smaller 256-point clouds
NUMWARS_DIMS=256 cargo bench --bench bench_mesh --features bench_mesh

# Focus on one operation
NUMWARS_FILTER="mesh/rmsd/f32" \
cargo bench --bench bench_mesh --features bench_mesh