Benchmarks

Benchmarks

This page records CPU compile time, draw time and peak memory for bench/catalog.json. The same size and frame count run in three.js. The previous Mojo release is compared when that compiler is installed.

The draw column names the faster frame loop. Whole-process time is summarized, because startup can hide the draw. Displayed times are rounded. The results file keeps the recorded values before display rounding.

The dates below describe saved runs, not exact source revisions. These retained files do not record the tested commit. They do not establish performance for later source changes, including the merge of #594. Rerun the benchmarks to measure a later revision.

Who wins

Draw time is the frame loop inside the process. A gap under 10% is a tie. A gap of 30% or more is a large win.

cpu-flat fills triangles with a flat color and a depth test. It does no lighting, no textures and no file write, so it is not the language comparison.

WebGL is three.js drawing with WebGL 2 when the context runs. The table uses that draw when WebGL runs. The renderer, backend and work differ, so this does not isolate language speed. A host with no GPU compares the CPU flat fill instead.

Linux

WebGL did not run on this host. cpu-flat is a flat color fill, so the comparison is the CPU fill.

cpu-flat is the draw comparison on this host. ThreeMojo wins 6 draws, 6 of them by 30% or more. cpu-flat wins 67 draws, 67 of them by 30% or more. 3 draws are within 10% and count as ties. The median cpu-flat draw takes 0.42 times the ThreeMojo draw. The flat fill does less work than the ThreeMojo frame.

ThreeMojo uses less memory on 75 of 76 examples. The median resident set is 45 MiB for ThreeMojo and 64 MiB for cpu-flat.

Whole-process time includes startup and writing the image. The median ThreeMojo process takes 0.096 s. The median cpu-flat process takes 0.073 s. A Mojo program that does nothing takes 0.007 s. A Node process that does nothing takes 0.019 s. The median Mojo 1.2.0 compile takes 26.4 s.

Mojo 1.1.0 runs 2 programs. Refused programs: 75. A refused compile is not a faster compile. On those programs, Mojo 1.2.0 compiles faster on 1 of 2. 1 compile is within 10%.

2 runs are within 10%.

The catalog lists 76 examples. This host measures 76 of them.

macOS

This file has no ThreeMojo draw times. The draw winner stays blank until the next measurement.

ThreeMojo uses less memory on 61 of 61 examples. The median resident set is 46 MiB for ThreeMojo and 87 MiB for WebGL.

Whole-process time includes startup and writing the image. The median ThreeMojo process takes 0.058 s. The median WebGL process takes 0.099 s. A Mojo program that does nothing takes 0.010 s. A Node process that does nothing takes 0.023 s. The median Mojo 1.1.0 compile takes 8.79 s.

Mojo 1.0.0 runs 5 programs. Refused programs: 57. A refused compile is not a faster compile. On those programs, Mojo 1.1.0 compiles faster on 5 of 5.

Mojo 1.1.0 runs faster on 4 of 5. 1 run is within 10%.

The catalog lists 76 examples. This host measures 61 of them.

Linux against macOS

For these saved runs, the median Linux/macOS whole-process time ratio is 1.6. The median Linux/macOS compile-time ratio is 3.0. Each ratio is Linux time divided by macOS time.

The Linux date is 2026-10-23. The macOS date is 2026-09-26. The source and software can differ between these snapshots. These ratios do not isolate hardware speed. The comparison uses 61 shared examples.

Machines

Linux

• Latest refresh: 2026-10-23
• OS: Linux 6.12.94+
• CPU: Intel(R) Xeon(R) Processor
• Mojo: Mojo 1.2.0 (2016d0cd)
• Previous Mojo: Mojo 1.1.0 (8189361e)
• Node: v22.23.4
• three.js backends: cpu-flat (webgl did not run)
• A Mojo program that does nothing: 0.007 s
• A Node process that does nothing: 0.019 s

macOS

• Latest refresh: 2026-09-26
• OS: macOS 26.7 (arm64)
• CPU: Apple M4 Max, 16 cores
• Mojo: Mojo 1.1.0 (8189361e)
• Previous Mojo: Mojo 1.0.0 (ed45d567)
• Node: v24.20.0
• three.js backends: cpu-flat and webgl
• A Mojo program that does nothing: 0.010 s
• A Node process that does nothing: 0.023 s

The macOS record is from Mojo 1.1.0. No Mac was available for the move to Mojo 1.2.0, so that record is not measured again. Run make bench-examples on a Mac to refresh it.

What the columns measure

Column	Meaning
Compile	mojo build, in seconds, rounded
Mojo draw	The frame loop inside the ThreeMojo process, in seconds
cpu-flat draw	The three.js flat fill, in seconds
WebGL draw	The three.js WebGL 2 draw, in seconds
Draw winner	The faster draw. tie means a gap under 10%
Mojo RSS	Peak resident set of the ThreeMojo run, in MiB
Measured	The date of that row

Whole-process time and the cpu-flat resident set stay in bench/results-linux.json and bench/results-macos.json. The refresh command is in How to measure examples.

Catalog examples

Linux

Example	Size	Frames	Compile (s)	Mojo draw (s)	cpu-flat draw (s)	WebGL draw (s)	Draw winner	Mojo RSS (MiB)	Measured
triangle	320×240	1	2.91	0.001	0.006	unavailable	Mojo	11	2026-10-23
spin	160×120	24	3.02	0.002	0.007	unavailable	Mojo	18	2026-10-23
cube	240×180	36	4.95	0.009	0.015	unavailable	Mojo	41	2026-10-23
cubes	260×200	48	25.9	0.052	0.026	unavailable	cpu-flat	56	2026-10-23
uv	320×200	2	18.1	0.008	0.008	unavailable	tie	14	2026-10-23
textured	260×200	36	26.0	0.15	0.017	unavailable	cpu-flat	48	2026-10-23
glass	260×200	36	25.6	0.068	0.021	unavailable	cpu-flat	48	2026-10-23
floor	320×200	30	25.9	0.20	0.018	unavailable	cpu-flat	43	2026-10-23
photo	260×200	36	26.5	0.092	0.017	unavailable	cpu-flat	48	2026-10-23
lamps	260×200	36	25.7	0.063	0.032	unavailable	cpu-flat	51	2026-10-23
first_scene	320×240	1	25.3	0.003	0.007	unavailable	Mojo	14	2026-10-23
lit_scene	320×240	36	25.7	0.13	0.020	unavailable	cpu-flat	65	2026-10-23
rotations	240×180	36	26.6	0.033	0.015	unavailable	cpu-flat	44	2026-10-23
ortho	240×180	36	25.7	0.033	0.012	unavailable	cpu-flat	43	2026-10-23
geometry	240×180	36	25.9	0.061	0.060	unavailable	tie	44	2026-10-23
instances	240×180	36	25.6	0.043	0.028	unavailable	cpu-flat	43	2026-10-23
raycast	240×180	36	26.7	0.063	0.041	unavailable	cpu-flat	45	2026-10-23
curves	240×180	36	26.3	0.035	0.023	unavailable	cpu-flat	44	2026-10-23
keyframes	240×180	36	28.3	0.025	0.016	unavailable	cpu-flat	45	2026-10-23
skinning	240×180	36	26.2	0.037	0.013	unavailable	cpu-flat	43	2026-10-23
phong	240×180	36	25.4	0.082	0.040	unavailable	cpu-flat	44	2026-10-23
fog	240×180	36	25.9	0.021	0.015	unavailable	cpu-flat	43	2026-10-23
culling	240×180	36	26.1	0.025	0.027	unavailable	tie	43	2026-10-23
clipping	240×180	36	25.6	0.039	0.015	unavailable	cpu-flat	45	2026-10-23
gpu_backend	240×180	36	27.6	0.061	0.027	unavailable	cpu-flat	43	2026-10-23
exposure	240×180	36	26.9	0.077	0.015	unavailable	cpu-flat	46	2026-10-23
model	240×180	36	27.2	0.037	0.015	unavailable	cpu-flat	44	2026-10-23
orbit	240×180	36	26.4	0.036	0.016	unavailable	cpu-flat	43	2026-10-23
clock	240×180	36	27.4	0.032	0.015	unavailable	cpu-flat	43	2026-10-23
chain	240×180	36	26.3	0.038	0.015	unavailable	cpu-flat	45	2026-10-23
additive	240×180	36	26.3	0.057	0.015	unavailable	cpu-flat	45	2026-10-23
normals	240×180	36	26.7	0.071	0.035	unavailable	cpu-flat	45	2026-10-23
fragments	240×180	36	26.0	0.060	0.023	unavailable	cpu-flat	44	2026-10-23
edges	240×180	48	2.87	0.007	0.011	unavailable	Mojo	47	2026-10-23
lines	240×180	36	26.0	0.036	0.017	unavailable	cpu-flat	43	2026-10-23
sprites	240×180	36	26.0	0.043	0.024	unavailable	cpu-flat	43	2026-10-23
stereo	240×120	36	29.1	0.049	0.021	unavailable	cpu-flat	34	2026-10-23
television	240×180	36	27.3	0.10	0.017	unavailable	cpu-flat	49	2026-10-23
mirror	240×180	30	27.2	0.25	0.032	unavailable	cpu-flat	41	2026-10-23
split	240×180	36	27.0	0.068	0.027	unavailable	cpu-flat	47	2026-10-23
physical	480×220	1	25.5	0.022	0.009	unavailable	cpu-flat	39	2026-10-23
outlines	240×180	36	26.1	0.032	0.013	unavailable	cpu-flat	43	2026-10-23
shadows	240×180	36	25.3	0.13	0.021	unavailable	cpu-flat	46	2026-10-23
wide	240×180	36	26.1	0.053	0.030	unavailable	cpu-flat	45	2026-10-23
bloom	240×180	36	38.1	0.23	0.048	unavailable	cpu-flat	52	2026-10-23
gizmo	240×180	36	27.1	0.057	0.021	unavailable	cpu-flat	47	2026-10-23
json_scene	240×180	36	54.7	0.034	0.014	unavailable	cpu-flat	44	2026-10-23
reloaded	240×180	36	55.6	0.10	0.051	unavailable	cpu-flat	46	2026-10-23
gem	240×180	36	27.9	0.25	0.034	unavailable	cpu-flat	51	2026-10-23
distance	240×180	36	25.7	0.093	0.039	unavailable	cpu-flat	47	2026-10-23
unfogged	240×180	36	26.2	0.033	0.014	unavailable	cpu-flat	44	2026-10-23
targets	240×180	36	26.5	0.12	0.035	unavailable	cpu-flat	44	2026-10-23
layers	320×180	36	26.5	0.18	0.048	unavailable	cpu-flat	57	2026-10-23
graph	240×180	36	27.2	0.13	0.036	unavailable	cpu-flat	45	2026-10-23
basis	320×180	36	32.6	0.062	0.012	unavailable	cpu-flat	54	2026-10-23
coats	320×180	36	26.2	0.16	0.049	unavailable	cpu-flat	64	2026-10-23
skyjson	240×180	36	52.7	0.22	0.029	unavailable	cpu-flat	51	2026-10-23
daylight	240×180	36	39.6	1.50	0.100	unavailable	cpu-flat	48	2026-10-23
faces	240×180	36	26.0	0.045	0.020	unavailable	cpu-flat	43	2026-10-23
utah	240×180	36	28.9	0.11	0.061	unavailable	cpu-flat	49	2026-10-23
blobs	240×180	36	36.0	0.034	0.080	unavailable	Mojo	43	2026-10-23
flipbook	240×180	36	28.4	0.048	0.027	unavailable	cpu-flat	45	2026-10-23
ripples	240×180	36	37.5	0.22	0.053	unavailable	cpu-flat	51	2026-10-23
baked	240×180	36	30.7	0.62	0.022	unavailable	cpu-flat	46	2026-10-23
diagram	240×180	36	12.0	0.036	0.026	unavailable	cpu-flat	42	2026-10-23
bricks	240×180	36	30.7	0.046	0.017	unavailable	cpu-flat	44	2026-10-23
terrain	240×180	36	26.3	0.059	0.038	unavailable	cpu-flat	45	2026-10-23
override	240×180	36	26.1	0.083	0.042	unavailable	cpu-flat	47	2026-10-23
clay	240×180	36	29.6	0.057	0.041	unavailable	cpu-flat	44	2026-10-23
cells	240×180	36	28.6	0.70	0.053	unavailable	cpu-flat	44	2026-10-23
cloud	240×180	36	26.1	0.061	0.003	unavailable	cpu-flat	44	2026-10-23
bounce	240×180	36	27.6	7.64	0.030	unavailable	cpu-flat	49	2026-10-23
sunlight	240×180	36	27.2	0.15	0.024	unavailable	cpu-flat	46	2026-10-23
vase	240×180	36	27.1	0.045	0.032	unavailable	cpu-flat	45	2026-10-23
sapling	240×180	36	27.0	0.034	0.014	unavailable	cpu-flat	45	2026-10-23
particles	240×180	36	28.6	0.041	0.023	unavailable	cpu-flat	44	2026-10-23

A legacy date is from an older aggregate file; its per-row measurement date is unknown.

Draw time is the frame loop inside the process. A gap under 10% is a tie. cpu-flat is a flat color fill with a depth test. WebGL did not run, so the draw winner uses cpu-flat.

macOS

Example	Size	Frames	Compile (s)	Mojo draw (s)	cpu-flat draw (s)	WebGL draw (s)	Draw winner	Mojo RSS (MiB)	Measured
triangle	320×240	1	1.39	—	0.002	0.014	—	14	unknown (legacy: 2026-09-26)
spin	160×120	24	1.37	—	0.003	0.020	—	20	unknown (legacy: 2026-09-26)
cube	240×180	36	2.33	—	0.006	0.032	—	45	unknown (legacy: 2026-09-26)
cubes	260×200	48	9.46	—	0.012	0.035	—	51	unknown (legacy: 2026-09-26)
uv	320×200	2	6.87	—	0.002	0.016	—	16	unknown (legacy: 2026-09-26)
textured	260×200	36	9.06	—	0.007	0.032	—	53	unknown (legacy: 2026-09-26)
glass	260×200	36	8.91	—	0.010	0.035	—	53	unknown (legacy: 2026-09-26)
floor	320×200	30	9.03	—	0.008	0.029	—	42	unknown (legacy: 2026-09-26)
photo	260×200	36	9.18	—	0.007	0.034	—	53	unknown (legacy: 2026-09-26)
lamps	260×200	36	8.66	—	0.014	0.032	—	45	unknown (legacy: 2026-09-26)
first_scene	320×240	1	8.61	—	0.002	0.015	—	17	unknown (legacy: 2026-09-26)
lit_scene	320×240	36	8.73	—	0.009	0.032	—	70	unknown (legacy: 2026-09-26)
rotations	240×180	36	8.69	—	0.006	0.030	—	46	unknown (legacy: 2026-09-26)
ortho	240×180	36	8.69	—	0.004	0.030	—	46	unknown (legacy: 2026-09-26)
geometry	240×180	36	8.75	—	0.035	0.030	—	46	unknown (legacy: 2026-09-26)
instances	240×180	36	8.65	—	0.012	0.031	—	47	unknown (legacy: 2026-09-26)
raycast	240×180	36	9.03	—	0.019	0.035	—	46	unknown (legacy: 2026-09-26)
curves	240×180	36	8.82	—	0.011	0.032	—	46	unknown (legacy: 2026-09-26)
keyframes	240×180	36	9.63	—	0.007	0.030	—	46	unknown (legacy: 2026-09-26)
skinning	240×180	36	8.84	—	0.005	0.030	—	46	unknown (legacy: 2026-09-26)
phong	240×180	36	8.62	—	0.018	0.097	—	46	unknown (legacy: 2026-09-26)
fog	240×180	36	8.69	—	0.006	0.028	—	46	unknown (legacy: 2026-09-26)
culling	240×180	36	8.67	—	0.013	0.031	—	46	unknown (legacy: 2026-09-26)
clipping	240×180	36	8.74	—	0.006	0.028	—	46	unknown (legacy: 2026-09-26)
gpu_backend	240×180	36	9.31	—	0.013	0.033	—	46	unknown (legacy: 2026-09-26)
exposure	240×180	36	8.68	—	0.006	0.030	—	46	unknown (legacy: 2026-09-26)
model	240×180	36	9.13	—	0.006	0.032	—	46	unknown (legacy: 2026-09-26)
orbit	240×180	36	8.73	—	0.006	0.026	—	46	unknown (legacy: 2026-09-26)
clock	240×180	36	8.74	—	0.006	0.025	—	46	unknown (legacy: 2026-09-26)
chain	240×180	36	8.68	—	0.006	0.029	—	40	unknown (legacy: 2026-09-26)
additive	240×180	36	8.53	—	0.006	0.026	—	46	unknown (legacy: 2026-09-26)
normals	240×180	36	8.50	—	0.017	0.024	—	46	unknown (legacy: 2026-09-26)
fragments	240×180	36	8.60	—	0.010	0.026	—	46	unknown (legacy: 2026-09-26)
edges	240×180	48	1.30	—	0.004	0.027	—	51	unknown (legacy: 2026-09-26)
lines	240×180	36	8.80	—	0.007	0.026	—	40	unknown (legacy: 2026-09-26)
sprites	240×180	36	8.77	—	0.010	0.028	—	47	unknown (legacy: 2026-09-26)
stereo	240×120	36	8.84	—	0.008	0.025	—	36	unknown (legacy: 2026-09-26)
television	240×180	36	9.04	—	0.006	0.025	—	50	unknown (legacy: 2026-09-26)
mirror	240×180	30	9.18	—	0.011	0.024	—	41	unknown (legacy: 2026-09-26)
split	240×180	36	9.23	—	0.011	0.026	—	46	unknown (legacy: 2026-09-26)
physical	480×220	1	8.95	—	0.003	0.085	—	76	unknown (legacy: 2026-09-26)
outlines	240×180	36	8.99	—	0.005	0.025	—	40	unknown (legacy: 2026-09-26)
shadows	240×180	36	8.79	—	0.009	0.033	—	47	unknown (legacy: 2026-09-26)
wide	240×180	36	8.99	—	0.015	0.026	—	47	unknown (legacy: 2026-09-26)
bloom	240×180	36	12.1	—	0.027	0.032	—	43	unknown (legacy: 2026-09-26)
gizmo	240×180	36	9.24	—	0.008	0.026	—	41	unknown (legacy: 2026-09-26)
json_scene	240×180	36	19.5	—	0.005	0.025	—	47	unknown (legacy: 2026-09-26)
reloaded	240×180	36	21.3	—	0.028	0.025	—	47	unknown (legacy: 2026-09-26)
gem	240×180	36	8.77	—	0.015	0.049	—	52	unknown (legacy: 2026-09-26)
distance	240×180	36	8.59	—	0.019	0.024	—	46	unknown (legacy: 2026-09-26)
unfogged	240×180	36	8.64	—	0.005	0.029	—	46	unknown (legacy: 2026-09-26)
targets	240×180	36	8.44	—	0.015	0.026	—	40	unknown (legacy: 2026-09-26)
layers	320×180	36	8.70	—	0.022	0.040	—	53	unknown (legacy: 2026-09-26)
graph	240×180	36	9.25	—	0.016	0.029	—	46	unknown (legacy: 2026-09-26)
basis	320×180	36	12.4	—	0.005	0.029	—	58	unknown (legacy: 2026-09-26)
coats	320×180	36	8.94	—	0.027	0.036	—	50	unknown (legacy: 2026-09-26)
skyjson	240×180	36	19.7	—	0.012	0.040	—	40	unknown (legacy: 2026-09-26)
daylight	240×180	36	11.7	—	0.044	0.028	—	47	unknown (legacy: 2026-09-26)
faces	240×180	36	8.61	—	0.008	0.027	—	46	unknown (legacy: 2026-09-26)
utah	240×180	36	9.13	—	0.036	0.025	—	47	unknown (legacy: 2026-09-26)
blobs	240×180	36	15.7	—	0.051	0.024	—	46	unknown (legacy: 2026-09-26)

A legacy date is from an older aggregate file; its per-row measurement date is unknown.

Draw time is the frame loop inside the process. A gap under 10% is a tie. cpu-flat is a flat color fill with a depth test. WebGL is the three.js draw used for the language comparison.

The pinned Mojo against the previous release

The table lists only programs both compilers run. The column names give the version of each compiler on that host. A refused compile is not a faster compile.

Linux

Program	1.2.0 compile (s)	1.1.0 compile (s)	Compile winner	1.2.0 run (s)	1.1.0 run (s)	Run winner
probe	1.97	2.06	tie	0.042	0.043	tie
diagram	12.0	14.0	1.2.0	0.081	0.078	tie

Programs in this table: 2. Refused catalog rows: 75. A refused compile is not a faster compile. Those rows stay out of this table.

macOS

Program	1.1.0 compile (s)	1.0.0 compile (s)	Compile winner	1.1.0 run (s)	1.0.0 run (s)	Run winner
probe	0.34	3.19	1.1.0	0.025	0.029	1.1.0
triangle	1.39	6.53	1.1.0	0.012	0.011	tie
spin	1.37	6.65	1.1.0	0.014	0.018	1.1.0
cube	2.33	7.65	1.1.0	0.026	0.034	1.1.0
edges	1.30	6.25	1.1.0	0.030	0.042	1.1.0

Programs in this table: 5. Refused catalog rows: 57. A refused compile is not a faster compile. Those rows stay out of this table.

Mojo 1.1.0 against Mojo 1.2.0 on the whole catalog

Mojo 1.2.0 compiles the catalog 10% faster than Mojo 1.1.0. It draws at the same speed.

The previous-release table above lists only 2 programs, because Mojo 1.1.0 refuses the 1.2 source. So a second run measures the whole catalog. That run builds the source of commit 9b910ded, which is the source before the port to 1.2, with Mojo 1.1.0. The 1.2.0 numbers come from make bench-examples on the port. To repeat the 1.1.0 run, check out 9b910ded, link a Mojo 1.1.0 environment as .venv, and run sh tools/tool.sh .venv/bin/mojo bench_examples --skip-mojo10 --no-write-wiki. Both runs used the Linux host above, one after the other, and no test or build ran beside them.

Measure, 76 examples	Mojo 1.1.0	Mojo 1.2.0
Median compile	29.3 s	26.4 s
Total compile	2288 s	2050 s
Total draw	15.67 s	15.73 s
Median whole process	0.095 s	0.096 s
Median ThreeMojo resident set	45.0 MiB	44.6 MiB
Median compiler resident set	826 MiB	791 MiB

The next table counts examples. A gap under 10% is a tie.

Measure	1.2.0 faster	Tie	1.1.0 faster
Compile	35	41	0
Draw	11	63	2
Whole process	3	66	7

The two draws that are slower on 1.2.0 are uv (0.0065 s to 0.0080 s) and gpu_backend (0.053 s to 0.061 s). The gaps are 1.5 ms and 7 ms, so they can be noise. The largest draw gain is sapling (0.045 s to 0.034 s). The largest compile gains are first_scene, faces, lamps, bricks, ripples and utah, each 16% to 21% faster. A compile runs once, and a draw is the fastest of three runs. The seven slower whole-process times are gaps of 12 ms or less.

The source differs between the two columns, so this is a toolchain and port comparison, not a compiler-only one. The port touches 56 library, example and benchmark files, with 113 lines added and 97 removed. The changes include __async def, std.hash and the Float64(String) parse. Mojo 1.2.0 also moves a few std.math results in the last bits, and the anatomy reports record the changed floats.

PMREM and the coverage run

The heaviest suites now write a quarter to a half of the coverage records they wrote before issue #157. Every image is the same to the bit.

A coverage run writes one record for each statement that runs. So the cost of a suite under coverage follows the statements in its innermost loops, not its run time. test_pmrem takes a fifth of a second without coverage, but it wrote 14 GB under coverage. See Coverage tool.

Suite, every CPU module instrumented	Before	After
test_pmrem	13.95 GB, 390 s	3.63 GB, 99 s
test_renderer	3.22 GB, 84 s	2.35 GB, 70 s
test_environment	2.34 GB, 63 s	1.24 GB, 38 s
test_rasterizer	0.65 GB, 16 s	0.49 GB, 15 s

The times come from a Linux container with four shared cores, so they are approximate. The sizes are exact.

The changes:

• The PMREM blur finds its weights, sines, cosines and copy position once per pass. CubeUvCopy holds the position, and cube_uv_coordinate uses it too.
• A bilinear read wraps each of its two columns and two rows once.
• FloatColor and the CLAMP case of wrap_index have no statements to probe. Vector3.cross and the shadow texel clamp are one statement each.
• Spherical harmonics find only the weight they use.
• test_pmrem blurs four environments, not seven. Three tests only need a valid layout, and now use one made by hand.

Without coverage, a 256-texel cube prefilters in about 210 ms, from about 270 ms before.

The sums keep their order, so each blur gives the same bits. The Mojo compiler fuses a multiply and an add into one instruction when it can. Thus moving a product out of a loop can change the last bit. The blur keeps the cross product per tap for that reason. A scratch program compared the images before and after, bit for bit, at face sizes 16 to 256 and from a panorama.

Other CPU benches

The catalog compares ThreeMojo with three.js. Anatomy, buildings, CARLA towns and the viewer stay outside it. They have no matching three.js scene, or they need a fetched asset cache. Every other CPU bench is named here.

Program	Record	Measures
bench/raster_bench.mojo	This page	CPU and GPU raster time by image size
bench/scene_bench.mojo	This page	CPU renderer stages, one worker and every core
bench/probe.mojo	Previous-release table	A triangle fill with no ThreeMojo import
bench/mojo10/probe.mojo	Mojo 1.0 table	The same fill, built by Mojo 1.0
bench/noop.mojo	Host baselines	A program that does nothing
bench/animation_loop_bench.mojo	Animation	Actions and the mixer
bench/renderer_sort_bench.mojo	Renderer	Draw-list sort against the previous sort
bench/image_decode.mojo	Image decode queue	glTF and CARLA image preload
bench/periodic_remainder_bench.mojo	Math	Euclidean remainder and pingpong
bench/exact_predicates_bench.mojo	Geometry	Exact orientation predicates
bench/convex_hull_bench.mojo	Geometry	Ordinary convex hulls
bench/ray_query_bench.mojo	Raycasting	Ray bounds queries
bench/ray_query_consumers.mojo	Raycasting	Callers of those queries
bench/physics_static_bench.mojo	Physics	Static primitive sweep
bench/physics_static_bvh.mojo	Physics	Snapshot BVH experiment
bench/physics_snapshot_bench.mojo	Physics query snapshots	Large snapshot capture and query
bench/physics_snapshot_small_bench.mojo	Physics query snapshots	Small snapshot capture and query
bench/physics_ccd_bench.mojo	Continuous collision	Sphere and mesh steps
bench/physics_ccd_index_bench.mojo	Continuous collision	CCD triangle index
bench/carla_force_restore_bench.mojo	Tick force restoration	Force restore on a tick
bench/carla_moving_supports_bench.mojo	CARLA physics	Tires on moving supports
bench/carla_model_cache_bench.mojo	CARLA assets	Vehicle model cache
bench/carla_sensor_cache_bench.mojo	CARLA rendering	Semantic and depth capture cache
bench/carla_lane_orientation_bench.mojo	CARLA lane orientation	Lane heading correction
bench/carla_road_fixed_s_bench.mojo	CARLA fixed-s nearest	Fixed-s nearest lane query
bench/carla_segment_numerics_bench.mojo	CARLA maps	Segment distance numerics
bench/carla_search_queue_bench.mojo	CARLA agents	Search queue operations
bench/carla_route_search_bench.mojo	CARLA agents	Complete route search
bench/carla_navigation_budget_bench.mojo	CARLA agents	Search budgets and heap bytes
bench/carla_navigation_headroom_bench.mojo	CARLA agents	Fixture search headroom

The catalog comparison is make bench-examples. make bench times raster sizes. make bench-scene times renderer stages. Those two Mac tables below are from the Apple M4 Max. They are not from the Linux host above.

On the Mac

make bench on the Apple M4 Max, one triangle, best of five runs. The GPU column includes allocation and the copy back:

Size	CPU	GPU (Metal)
320×240	0.3 ms	17.6 ms
640×480	1.2 ms	19.0 ms
1280×720	3.4 ms	22.7 ms
1920×1080	8.3 ms	34.6 ms
3840×2160	30.3 ms	89.8 ms

make bench-scene on the same Mac, the sphere of 12096 triangles at 1280×720:

Workers	prepare	rasterize	resolve	frame
1	1.9 ms	82.7 ms	8.9 ms	96.8 ms
16	1.9 ms	9.1 ms	1.0 ms	15.4 ms