Benchmarks

The numbers §9 of docs/ARCHITECTURE.md is measured against. Two renderer classes appear throughout:

	Renderer	Where
[Metal]	ANGLE (Apple, ANGLE Metal Renderer: Apple M2 Max)	the Electron app — what a user gets
[SwS]	ANGLE (Google, Vulkan 1.3.0 (SwiftShader Device …))	headless Chromium — the golden authority

Everything below was measured on an M2 Max with TETRAVOX_TESTDATA=…/derivatives/SimNIBS/sub-ernie. Reproduce with cargo bench -p <crate>, node scripts/bench-wasm-cut.mjs, cargo run -p tvx-geom --example measure, and the real-data Playwright specs (mesh-real, derived-r4, phase1-gate), which also write their own current figures into docs/benchmarks/ on every run.

Load to first frame

File	Bytes	Renderer	Load → first frame
m2m_ernie/T1.nii.gz	13,143,463	[Metal], app	404 – 524 ms (typ. 415)
m2m_ernie/T1.nii.gz	13,143,463	[SwS], engine harness	318 ms
m2m_ernie/ernie.msh	184,207,351	[Metal], app	1,221 – 1,353 ms (typ. 1,240)
m2m_ernie/ernie.msh	184,207,351	[SwS], engine harness	1,240 ms

Where T1.nii.gz's 400 ms goes (criterion medians, native):

Stage	ms
read_nifti	101.1
stats(0)	46.1
gpu_payload → R16	43.2
the three together	191.5
loadVolume end to end in the worker, incl. fetch + transfer	373 ([SwS])
… then GL upload + first draw + shell	~45

Mesh pipeline, native, ernie.msh

Step	ms	Note
read_msh	77	§6.2's bar is < 1.5 s native
orient_surface (1,177,213 tris)	100
morton_reorder (4,722,625 tets)	93	§9.1 row 8's 250 ms bar is WASM
build_tet_blocks (73,792 blocks)	62	§9.1 row 9's 500 ms bar is WASM
build_point_locator	188
tag_surfaces → 1,177,213 tris / 582,126 verts / 10 tags	36
locate_point (bbox centre)	2.1	§8 budgets the hover round trip at ≤ 50 ms
extract_boundary (topo = None)	624

plane_cut — §9.1 row 10

Plane	The op (mesh_cut, wasm)	Worker round trip (Chromium)	Native	Bar
mid-axial, 62,966 cap tris	12.9 ms	16.9 ms	10.4 ms	< 15 ms ✔
oblique [1,1,1], 76,217 cap tris	16.6 ms	21.2 ms	13.7 ms	< 30 ms ✔

Without the Morton block index the same planes are 24.3 / 27.2 ms native, so the index is worth ~2.1×. The cost tracks the cap triangle count, not the tet count:

Mesh	Tets	Cap tris	[Metal]	[SwS]
testdata/mesh_v2_binary.msh	48	32	4.5 ms	8.6 ms
grey_Thalamus_TI.msh	1,340,029	30,058	9.6 ms	66 ms
ernie.msh	4,722,625	70,757	18.7 ms	151 ms

The ~8× column gap is the two Chromium builds' WebAssembly tiers on identical wasm, not this feature.

Slice sweep with a mesh cross-section — R4's bar

ernie.msh alone, fillIn2D + contoursIn2D on, 20 × 1 mm, 768×768, DPR 1, [Metal]:

Sweep	Per step (median)	Worst step	End-to-end	Bar
1x1 axial	23.6 ms	26.3 ms	42.1 fps	≥ 30
2x2 (3 panes + 3D)	24.4 ms	26.6 ms	41.9 fps	≥ 30

The 2x2 sweep moves only the axial plane; the other two panes re-request an identical plane and the cut source drops the repeat, which is what keeps three panes from costing three cuts.

Render

Buffer	Pixels	CPU median	CPU p95	GPU median
1200 × 800 (1×)	0.96 MP	0.10–0.20 ms	0.20–0.60 ms	1.65 – 2.16 ms
2400 × 1600 (2×)	3.84 MP	0.10 ms	0.20–0.50 ms	2.48 – 3.92 ms

Orbiting ernie's tag surfaces: plain 1.18 M-tri pass 2.32 ms, with the wireframe 2.24 ms. Six clip planes: discard 2.89 ms vs gl_ClipDistance 2.07 ms.

Responsiveness — §9.1 row 6

	Measured	Budget
Open → load card on screen	0.8 ms	—
Open → moving progress bar	13 – 28 ms	< 200 ms
Cancel click → card reads cancelled	4 – 6 ms	< 500 ms

Memory — §9.2

wasm_heap_bytes() after loadMesh (resident, not live):

File	Bytes	Heap	× file	Bar
m2m_ernie/ernie.msh	184,207,351	341.8 MB	1.95 ×	≤ 380 MB ✔
m2m_ernie/ernie_seeg.msh	492,090,201	912.4 MB	1.94 ×	≤ 1.0 GB ✔
m2m_ernie/T1.nii.gz	13,143,463	105.3 MB	—	—

After buildTopology the same two reach 846.1 MB and 1,893.1 MB resident — 21 % and 47 % of wasm32's 4,032 MiB usable ceiling. Linear memory grows and never shrinks, so the observable peak is the load path's resident total plus the topology arena, not the larger of the two.