Benchmarks — Balaur beside Godot, case for case

Balaur 0108981c0 on Apple M1, 8 cores, Darwin 25.6.0, 2026-09-05.

Every case is the scene the reference suite builds, with the same counts, the same 60 Hz tick and the same window: a settle, then 300 timed steps, reported as the median and the 99th percentile of one tick.

step p50 is a whole physics tick: script fixed_update, the solver, and writing every simulated pose back to the scene tree — what the Godot suite calls step_ms. rapier's own step is the solver alone, and the gap between the two is what the engine adds. script seam is what the case's own script cost inside that tick.

Across the 12 physics cases the engine itself costs 1% to 13% of a tick; the rest is rapier, and in the query cases the script's own calls.

Median tick, lower better:

case	bodies	Balaur	quickest in Godot	that engine's tick, over ours
3d/pyramid	3795	10.43 ms	Box3D 13.52 ms	1.30x
3d/mixed_pile	5000	5.04 ms	Box3D 9.72 ms	1.93x
3d/joint_grid	5000	3.08 ms	Rapier 3D 8.92 ms	2.90x
3d/smash	5001	7.99 ms	Box3D 9.62 ms	1.20x
3d/query_storm	2000	5.50 ms	Rapier 3D 5.93 ms	1.08x
3d/drop	8000	3.06 ms	Box3D 11.34 ms	3.70x
2d/pyramid	10000	7.10 ms	Rapier 2D 20.03 ms	2.82x
2d/mixed_pile	5000	2.86 ms	Box2D v3 6.71 ms	2.35x
2d/joint_grid	5000	2.17 ms	Box2D v3 4.93 ms	2.27x
2d/smash	5001	3.50 ms	Box2D v3 11.57 ms	3.31x
2d/query_storm	2000	12.45 ms	Box2D v3 4.56 ms	0.37x
2d/drop	8000	6.07 ms	Rapier 2D 16.54 ms	2.73x

The one case Balaur loses is 2d/query_storm, and the table below says why: nine hundred queries a tick cost 11 ms crossing the script seam and 1 ms inside rapier. A game asking that many questions per tick from script pays for the questions, not for the broad phase.

The Godot columns are 4.7-stable (official) with the addons the reference suite commits, run on this same machine.

3D

pyramid (3d)

3795 bodies. 300 timed steps after 60.

engine	step p50	step p99	rapier's own step	script seam
Balaur	10.43	15.28	10.30	0.002
Box3D	13.52	33.12	—	—
Rapier 3D	14.62	34.17	—	—
Jolt	26.92	106.63	—	—
Godot Physics 3D	48.03	133.30	—	—

mixed_pile (3d)

5000 bodies, 5 static. 300 timed steps after 60.

engine	step p50	step p99	rapier's own step	script seam
Balaur	5.04	7.41	4.87	0.003
Box3D	9.72	17.16	—	—
Rapier 3D	11.28	34.73	—	—
Jolt	13.70	37.14	—	—
Godot Physics 3D	84.90	149.55	—	—

joint_grid (3d)

5000 bodies, 9950 joints. 300 timed steps after 60.

engine	step p50	step p99	rapier's own step	script seam
Balaur	3.08	4.40	2.82	0.002
Rapier 3D	8.92	14.13	—	—
Box3D	10.27	17.44	—	—
Godot Physics 3D	26.08	52.31	—	—

Balaur: displacement_max 49.58. Fingerprint cc96e435.

smash (3d)

5001 bodies. 300 timed steps after 60.

engine	step p50	step p99	rapier's own step	script seam
Balaur	7.99	10.68	7.81	0.003
Box3D	9.62	14.37	—	—
Jolt	12.12	29.06	—	—
Rapier 3D	19.14	25.05	—	—
Godot Physics 3D	99.98	201.82	—	—

Balaur: non_finite 0, spread 430.5. Fingerprint 8a46de70.

query_storm (3d)

2000 bodies, 8000 static. 300 timed steps after 60.

engine	step p50	step p99	rapier's own step	script seam
Balaur	5.50	6.03	1.12	4.09
Rapier 3D	5.93	6.80	—	—
Box3D	6.67	8.86	—	—
Jolt	26.45	34.11	—	—
Godot Physics 3D	73.26	86.42	—	—

Balaur: hits 201175, queries 270000. Fingerprint dc3e3b83.

drop (3d)

8000 bodies, 5 static. 300 timed steps after 60.

engine	step p50	step p99	rapier's own step	script seam
Balaur	3.06	4.24	2.77	0.002
Box3D	11.34	41.58	—	—
Rapier 3D	12.26	18.59	—	—
Jolt	22.14	37.79	—	—

2D

pyramid (2d)

10000 bodies. 300 timed steps after 60.

engine	step p50	step p99	rapier's own step	script seam
Balaur	7.10	8.53	6.71	0.002
Rapier 2D	20.03	28.12	—	—
Box2D v3	21.33	52.38	—	—
Godot Physics 2D	147.26	442.54	—	—

mixed_pile (2d)

5000 bodies, 3 static. 300 timed steps after 60.

engine	step p50	step p99	rapier's own step	script seam
Balaur	2.86	4.69	2.60	0.002
Box2D v3	6.71	11.73	—	—
Rapier 2D	9.25	15.28	—	—
Godot Physics 2D	71.12	136.33	—	—

joint_grid (2d)

5000 bodies, 9950 joints. 300 timed steps after 60.

engine	step p50	step p99	rapier's own step	script seam
Balaur	2.17	2.75	1.88	0.002
Box2D v3	4.93	8.75	—	—
Rapier 2D	10.16	19.44	—	—

Balaur: displacement_max 190.9. Fingerprint f779b873.

smash (2d)

5001 bodies. 300 timed steps after 60.

engine	step p50	step p99	rapier's own step	script seam
Balaur	3.50	4.39	3.29	0.002
Box2D v3	11.57	28.63	—	—
Rapier 2D	11.86	21.93	—	—
Godot Physics 2D	55.09	104.59	—	—

Balaur: non_finite 0, spread 2999. Fingerprint 7c549757.

query_storm (2d)

2000 bodies, 8000 static. 300 timed steps after 60.

engine	step p50	step p99	rapier's own step	script seam
Box2D v3	4.56	5.49	—	—
Rapier 2D	6.25	7.70	—	—
Balaur	12.45	15.57	1.14	11.03
Godot Physics 2D	57.24	67.48	—	—

Balaur: hits 280111, queries 324000. Fingerprint 5ff47d2c.

drop (2d)

8000 bodies, 3 static. 300 timed steps after 60.

engine	step p50	step p99	rapier's own step	script seam
Balaur	6.07	9.75	5.75	0.003
Rapier 2D	16.54	31.08	—	—
Box2D v3	22.66	46.56	—	—
Godot Physics 2D	103.34	286.81	—	—

Nodes

Scene tree

Milliseconds for the whole loop, lower better. Godot's column is its own published run on another machine, so it says which operations are in a different class, not how the two machines compare.

operation	Balaur p50	Balaur min	Godot release
add_children	46.18	45.85	28.19
delete_children_in_order	543.27	542.12	9.82
delete_children_reverse	463.37	461.49	6.10
delete_children_random	513.89	512.12	19.63
get_node	66.98	66.35	29.60

move_child has no row: the scene tree has no sibling-reorder operation to measure.

Destroying is two orders of magnitude off, and it is the scene tree, not the script: scene::free_subtree unlinks each node from its parent by scanning every sibling, so freeing a flat container of fifty thousand is quadratic. Adding and looking up are within a small factor of Godot's on a faster machine.

What is not the same

• Balaur's rapier is built without SIMD; the godot-rapier addon's is built with it. Same version, same enhanced-determinism, same threaded solver.

• The worlds are statistically alike, not bit-identical: the scattered cases draw from each engine's own random stream, so a fingerprint compares two Balaur runs, never Balaur against Godot.

• The 2D cases are the GDScript's coordinates with y negated, because Godot's 2D plane points down and ours points up. Distances, contacts and gravity are unchanged, and length_unit is 100 as godot-rapier's 2D default is.

• mixed_pile's convex hull is one fixed size rather than a random one: a hull collider is built from a mesh asset and has no scale.

• Godot's own suite runs its scene-tree cases on its CI machine, not this one.

Running it

cargo build --release -p balaur_cli --bin balaur
python3 scripts/bench_compare.py --godot-results <benchmarks-repo>/results

One case on its own, or in the editor:

target/release/balaur run examples/benchmark --headless --fixed-tick -- --case=3d/pyramid
target/release/balaur edit examples/benchmark