Demo gallery 3D demo Physics quality comparison Rendered workflow benchmark

Rapier and AVBD, side by side

Scene Bodies

Rapier iterations Rapier sleeping AVBD iterations AVBD margin Both: one camera drives both views

Loading…

Rapier (WASM, CPU)

three-avbd (WebGPU)

Both engines run at the comparison harness defaults unless changed here: Rapier's own defaults (4 solver iterations, 1 internal PGS iteration, sleeping and CCD off), AVBD 4 iterations and a 0.01 m contact margin (the World API's default is 10 iterations). Equal iteration counts are different amounts of work. Changing a setting rebuilds both worlds from step 0.
Same scene manifest (src/comparison/scenes.ts), seed, masses, friction and dt; each animation frame advances each engine by one step. Neither engine's stepping is used to build the other's input.
Timings are from this machine and this tab, rolling over the last 120 steps, and include no rendering. Rapier runs as WASM on the CPU: its ms/step is the synchronous world.step(). AVBD runs on the GPU: the headline figure is completed time, from the start of the step's CPU encode until queue.onSubmittedWorkDone() resolves, with the queue drained first (batch size 1, as in compare.html; it includes completion-promise latency). The second figure is CPU encode + submit only, which is not finished simulation. Use compare.html for batched throughput with replay validation.
Quality is read about once a second (and on every manual step): max floor penetration, max body–body overlap from an independent sweep-and-prune check (skipped above 12k bodies), and max linear speed. Invalid scenes are marked ✗ per SCENE_VALIDITY: they cannot back resting-quality claims.