Performance

All numbers are medians over repeated runs of benchmarks/benchmark.py, measured on an NVIDIA GeForce RTX 5090 with CUDA 12.8, extracting a sphere SDF at level 0. Timings include the full call from Python: case classification, extraction, and vertex welding.

Extraction time (median, ms)

Grid	Algorithm	128³	512³
uniform	marching_cubes	0.55	5.4
uniform	marching_tetrahedra	0.68	6.7
uniform	surface_nets	0.81	6.9
uniform	dual_contouring (ju)	0.83	7.0
uniform	dual_contouring (carrera)	380	4800
uniform	dual_marching_cubes	1.21	9.0
uniform	get_intersection	0.26	5.3
sparse	marching_cubes	0.48	1.9
sparse	marching_tetrahedra	0.61	3.2
sparse	surface_nets	0.73	2.2
sparse	dual_contouring (ju)	0.75	2.3
sparse	dual_contouring (carrera)	200	2600
sparse	dual_marching_cubes	0.44	3.5
sparse	get_intersection	0.13	0.53

Sparse grids only touch cells near the surface, so their cost scales with the surface area rather than the volume. At 512³ that is 300k cells instead of 134M.

The carrera variant of dual contouring is an iterative optimization (100 outer iterations of up to 100 inner iterations per cell by default) and sits three orders of magnitude above the one-pass methods. Its time scales with the number of crossed cells and the samples near the surface; fewer outer iterations trade accuracy for time.

Marching cubes variants (median, ms)

Variant	uniform 512³	sparse 512³
lorensen	4.9	1.4
nagae	5.0	1.5
lewiner	5.5	2.0
vega	5.4	1.9

Reproducing

The environment setup is described in Development. Then:

pixi run bench

# or directly, with options:
pixi run python benchmarks/benchmark.py \
    --res 32 64 128 256 512 --json results.json

The harness reports the median, 10th and 90th percentiles, and the first call separately.

Notes

• The first call after installation is slow (a few seconds): the driver compiles the shipped PTX for your GPU and caches the result. See Installation. The benchmark warms up before timing, so the table excludes this.

• The kernels compute cell corners on the fly instead of storing them, so memory use grows with the extracted surface, not the grid volume. A dense 1024³ grid fits on a consumer GPU.

• Numbers depend on the GPU, driver, and field. Expect different results on other machines.