Builds and tracing

Measured build times, tree quality and ray rates, with how each number was taken.

Every number here says where it was measured. The package has two speed targets, the RTX 3060 Laptop and the Apple M5 Pro. The Intel UHD runs every check for correctness only. Numbers from the Apple M5 are indicative: its clocks drift by up to about 30% within a run, so those benches report ratios taken within one round, not absolute rates.

Method

Rays: a grid of probes fills the scene's bounds, and each probe fires 256 rays in random directions from a fixed seed. Any-hit rays go from each closest hit toward a fixed direction. The rays are generated once and stored in a buffer, so every tracer gets the same input.
Timing: GPU timestamps around the trace dispatch, the median of several runs after warm-ups. Without timestamps, a fence-based wall time over 20 dispatches, labelled as such.
The GPU is loaded and clocks are checked. A run counts only if it started cool, fewer than 5% of its samples were thermally throttled, and it had enough samples ("the clock rule").
three-mesh-bvh runs as a three WebGPURenderer compute node reading the same ray buffer, its per-mesh trees built with { strategy: SAH } explicitly (its default is CENTER).
The pages are tests/gpu/pages/trace-bench*; results go in bench/results/trace.csv.

Sponza on the RTX 3060 Laptop

5,744,384 triangles in 132 meshes. Headless Chrome 156 on Windows.

What	Number	Conditions
GPU H-PLOC build, whole scene	46.6 ms	Median of the 2 runs valid under the clock rule, loaded SM clock about 1,000 MHz, power-limited
CPU tinybvh build, whole scene	1.6-1.9 s	In a browser worker, single-threaded; medians of 3 runs per session
GPU LBVH build, whole scene	31.2 ms	Median of 3, before the clock rule existed
H-PLOC tree quality	SAH cost 1.032x tinybvh's	Lower is better; the LBVH's is 1.217x
Trace speed on H-PLOC trees	closest-hit 0.95x, any-hit 1.02x of tinybvh's trees	Same rays, both tracers in each run, median of 3. The runs were flagged low-clock, so the ratio is more reliable than the rates
Trace speed on LBVH trees	closest-hit 0.82x, any-hit 0.80x	As above
Closest-hit rate, tinybvh trees	104.3 Mrays/s, 1.20x three-mesh-bvh	Median of 3
Closest-hit rate, Intel UHD	12.8 Mrays/s, 1.97x three-mesh-bvh	Median of 3; correctness device, for scale only
GPU top-level refit, 10,000 moving instances	about 0.09 ms (0.086-0.091)	The refit's own page, two measurements

three-mesh-bvh's trees converted to our layout traced at 96.7 Mrays/s and took 7.4 s to build, against 101.6 Mrays/s and 5.3 s for tinybvh's, in the same session. These runs predate the 2026-11 levers: instanced trees, f32 nodes.

The 2026-11 levers on the RTX 3060 Laptop

Sponza and Bistro interior surface-ray pages (trace-bench-surface-*), Windows, headless Chrome 157, 2026-11-13, three runs each. Each lever is timed in paired rounds against the layout before it (instanced, f32 nodes, records, the 40-byte record, anyHitOrder: 'far'), with the order reversed every other round. A figure is the median over the runs of each run's median-of-rounds ratio; the range covers surface and air rays, instanced and merged trees. Above 1 is faster.

Lever	Sponza	Bistro interior	Default now
Merged static instances, over instanced	closest 1.44-1.67x, any 1.24-1.48x	closest 1.19-1.20x, any 1.16-1.35x	on
Quantized nodes, over `f32`	closest 1.04-1.22x, any 0.98-1.03x	closest 1.02-1.08x, any 0.98-1.01x	on
Indexed leaves, `vec4`, over records	0.92-0.97x	0.92-0.97x	off
Indexed leaves, packed, over records	0.89-0.95x	0.82-0.88x	off
40-byte hit record, over 24-byte	0.88-0.92x	0.93-0.94x	40-byte
`anyHitOrder: 'larger'`, over `'far'`	instanced 1.11-1.45x, merged air 1.07x, merged surface 0.87x	instanced 1.18-1.48x, merged air 1.03x, merged surface 0.80x	`'far'`

Compact against full record, indexed against records and every any-hit order against far agreed on every ray. No quantized t differed from the f32 one. Merged against instanced: closest hits named the instanced tree's instance and triangle on 99.949-99.996% of rays; the rest are near-ties at a shared distance, where the merged walk can pick another instance or triangle. Any hits blocked exactly the same rays.

On the Apple M5 (indicative)

What	Number	Conditions
`SceneTrees` first trees, Sponza	100-190 ms after the first update	`scene-trees-sponza` page, Chrome, several runs and caps
`SceneTrees` every tinybvh tree in, Sponza	285-370 ms after the first update	the same
Merged static instances, Sponza surface rays	closest 1.85x, any 1.36x of instanced	Merged static instances
GPU batch of 65k rays	about 1.9 ms (35 Mrays/s)	`three/raycast` example, a 100k-triangle knot from outside, timestamps
Path tracer	about 120 Mrays/s	`three/pathtracer` example, wall time of one dispatch
CPU raycast, 100k triangles	692,630 per second `firstHitOnly`	Raycast, Node 22

The variants' trace speed on the M5 (surface-ray pages, paired rounds with the order reversed, against f32 nodes and records): indexed leaves 0.83-0.99x with vec4 positions and 0.75-0.92x packed; quantized nodes 0.93-0.97x. The 40-byte hit record measured 0.96-1.15x the 24-byte one.

Not measured yet

On the RTX 3060: the correctness pages for the 2026-11 work, SceneTrees memory and the build-peak cap on D3D12, shader compile times, and quantized nodes and indexed leaves together.
quantizedDecode: 'folded' (the folded decode of quantized nodes, opt-in) on any GPU: the surface-ray pages time it against the exact decode (folded_over_exact); it has only passed the correctness pages on SwiftShader.
The Intel UHD has not run the 2026-11 work at all.
The comparison suite against three-mesh-bvh and others (compare/) has run on the Apple M5 only.

Reproducing

pnpm test:gpu:page trace-bench-sponza (and the other trace-bench-* pages, run by exact name) runs a trace bench on the machine's GPU. Sponza and Bistro need their assets; see tests/gpu/README.md.
pnpm test:gpu:page scene-trees-sponza reports SceneTrees load times and memory.
node bench/three-raycast.ts, bench/query-parity.ts and bench/closest-point.ts run the CPU benches in Node.