NeoSCAD benchmarks: method, results and caveats

Method

What is timed: the wall time of <binary> [--backend=…] -o out.stl model.scad, a full render to ASCII STL, including process startup.
Runs: one binary after another, never in parallel; the best of 3 (one run once a run takes over 60 s); a 300 s timeout per run.
Tool: conformance bench in the NeoSCAD repository, which records the machine, the binaries' versions and SHA-256s, and every run.
Correctness: each model's mesh is compared with NeoSCAD's (volume, area within 0.1%, bounding box). Every model agrees with the nightly.
Cached references: NeoSCAD is measured on every run. OpenSCAD's times come from a cache measured earlier on the same machine (27 September 2026; the BOSL2 test-suite time, 28 September) with the same binary (pinned by SHA-256), model, method and environment, so they aren't interleaved with NeoSCAD's runs.
Timeouts: excluded from the geometric means. The nightly with CGAL and OpenSCAD 2021.01 time out on two models, so their means cover 12 models, which understates NeoSCAD's lead over them.
Libraries and fonts: OPENSCADPATH points at BOSL2 (commit 9948313); each binary uses its own bundled fonts.

Machine and binaries

Machine	Mac16,7, Apple M4 Pro, 14 cores (10 performance + 4 efficiency), 48 GB, on AC power
OS	macOS 27.0 (26A428)
NeoSCAD	0.1.1 (commit `7c4ebc0`), `cargo build --release`, arm64
OpenSCAD nightly	2026.09.23, universal binary, run with `--backend=manifold` and `--backend=cgal`
OpenSCAD 2021.01	x86_64 only, so it runs under Rosetta 2 on this machine
Measured	NeoSCAD: 30 September 2026, 15:07 UTC. OpenSCAD: cached from 27 and 28 September 2026

The chart

Bar chart of the table below on a log time scale: for each model, bars for NeoSCAD, the nightly with Manifold, the nightly with CGAL, and OpenSCAD 2021.01, with NeoSCAD's time and its speedup over the nightly with Manifold on the right. Geometric means at the bottom: nightly Manifold 4.73× over 14 models, nightly CGAL 31.45× over 12, OpenSCAD 2021.01 69.31× over 12. — Log time scale. Hatched bars are timeouts; a red mark means the mesh differs from NeoSCAD's (only OpenSCAD 2021.01's, on two models). The right-hand columns are NeoSCAD's time and its speedup over the nightly with Manifold.

Every result

Best of 3, in seconds; × is the reference's time divided by NeoSCAD's.

† OpenSCAD 2021.01's mesh differs from NeoSCAD's and the nightly's (extrude_twist: volume and area; text_30lines: bounding box). Geometric means over the 14 models: 4.726× against the nightly with Manifold (14 models), 31.451× against the nightly with CGAL (12 models; bosl_fractal_tree and csg_spheres time out) and 69.313× against OpenSCAD 2021.01 (12 models, the same two excluded).
Model	NeoSCAD	Nightly, Manifold	×	Nightly, CGAL	×	2021.01	×
`bosl_fractal_tree` BOSL2 examples/fractal_tree.scad (recursive attach, depth 10)	0.615	11.73	19.07×	timeout	–	timeout	–
`bosl_gears__003` BOSL2 gears.scad doc example 3: helical spur_gear	0.046	0.214	4.68×	0.225	4.93×	0.411	9.00×
`bosl_isosurface__006` BOSL2 isosurface.scad doc example 6: metaballs sphere with connectors	0.604	2.471	4.09×	2.526	4.18×	5.501	9.11×
`bosl_screws__001` BOSL2 screws.scad doc example 1: selected UTS screws	0.164	0.668	4.07×	31.45	191.43×	70.91	431.56×
`bosl_spring_handle` BOSL2 examples/spring_handle.scad	0.204	0.729	3.58×	13.26	65.08×	41.90	205.70×
`csg_deep_union` 800 rotated/translated cubes, implicit top-level union	0.040	0.134	3.35×	7.617	190.44×	20.33	508.22×
`csg_spheres` cube(60) minus 5x5x5 grid of spheres ($fn=48)	0.396	1.130	2.86×	timeout	–	timeout	–
`ex_csg_basic` OpenSCAD examples/Basics/CSG.scad	0.007	0.062	9.04×	1.127	163.30×	3.050	442.04×
`ex_menger` OpenSCAD examples/Old/example024.scad (Menger sponge, n=3)	0.102	0.160	1.56×	15.19	148.37×	45.56	444.97×
`extrude_twist` linear_extrude(height=50, twist=720, slices=400) of square minus circle	0.035	0.218	6.25×	0.220	6.30×	0.327	9.36×
`import_stl` import() of a 21 MB ASCII STL (sphere $fn=256) minus two cubes	0.104	0.544	5.21×	10.17	97.37×	37.55	359.35×
`mink_convex` minkowski(cube([30,20,5],center=true), sphere(3,$fn=48))	0.022	0.074	3.42×	0.148	6.84×	0.102	4.72×
`mink_nonconvex` minkowski(L of two cubes, sphere(2,$fn=24))	0.010	0.064	6.38×	0.361	36.07×	0.901	90.13×
`text_30lines` linear_extrude(2) of 30 text() lines of 58 characters (Liberation Sans)	0.113	0.618	5.46×	0.592	5.23×	1.338	11.82×
`cold_start` cube(1) to ASCII STL: process startup plus the smallest render (not in the geometric means)	0.0029	0.0455	15.69×	0.0464	16.00×	0.0701	24.17×

The edit loop

An agent or a person edits one line and looks again. Each edit gives the line a new value, so no cached whole model answers it. Best of 10 edits, in milliseconds, from the same run. The nightly has no server mode, so it starts cold each time, exporting STL or a 1024×1024 rendered PNG.

Case	Warm `serve`: render	Warm `serve`: snapshot	CLI, no server: STL	CLI, no server: snapshot	Nightly: STL	Nightly: PNG
`bosl2` A BOSL2 part: rounded cuboid, chamfered cylinder and prismoid; the edit changes the cuboid's height	8.4	14.2	35.9	76.1	163.0	235.6
`csg` OpenSCAD's examples/Basics/CSG.scad; the edit moves the difference (the third object)	1.0	6.3	7.1	31.4	60.0	127.9

On a heavier model the edit time depends on what the edit touches. On the gearbox on the home page, in a warm neoscad serve, an edit to the planet carrier re-renders in 0.25 to 0.27 s; an edit to the plinth takes 1.55 s, because its isosurface is evaluated again. So: an edit to one part re-renders in about 0.25 s, not every edit.

Evaluation only: BOSL2's test suite

BOSL2's 976 tests, one process per test, run one after another and written to .echo, with no geometry: NeoSCAD takes 29.6 s and the nightly 145.7 s. Both pass all 976.

Caveats

The 4.7× is mostly small models. 8 of the 14 finish in under 0.11 s in NeoSCAD, where process startup counts for a lot (2.9 ms for NeoSCAD against 45.5 ms for the nightly). Subtract startup and the geometric mean is 3.63×; over the three models that take over 0.3 s it is 6.06×. One model lifts both: bosl_fractal_tree is 19×, and without it the startup-subtracted mean is 3.19×. The gearbox on the home page is 2.80× (2.09 s against 5.85 s; re-timed on 30 September, 2.03 s against 6.04 s). That is why we say "about 3.6× on heavy models", and why a given heavy model may see less.
We know of no model where NeoSCAD is slower, but we have only measured so many. The ones earlier builds were slower on, which this set doesn't include, are now faster (one run each): the Menger sponge at level 4 (1.40 s against the nightly's 2.03 s), 200 lines of text() as 2D exported to SVG (0.47 s against 1.34 s), and the same 200 lines extruded to 3D (2.65 s against 32.38 s). If you find a slower one, please open an issue.
One machine. An Apple M4 Pro, arm64, with a load average of 2.8 to 3.2 while NeoSCAD was measured. Other machines, and x86_64 in particular, haven't been measured here; users' own runs are collected in results from other machines.
The references weren't interleaved. OpenSCAD's times are from a cache measured earlier with the same binary, not alternated with NeoSCAD's runs.
OpenSCAD 2021.01 runs under Rosetta 2 on this machine, and is five years older than the nightly. It is here as a reference series only; compare against the nightly with Manifold.
The browser build will be slower. WebAssembly runs single-threaded, and these numbers don't apply to it.

Reproduce it

In a clone of the repository, build with cargo build --release, then run ./target/release/conformance bench. It writes a JSON record and a chart like the ones here to progress/bench/.

How fast, measured how