Why it matters: when budgeting a crowd of animated characters in three.js, the number to control is bones × characters. Switching to WebGPU does not change that.

Context

A browser prototype drew a crowd with one SkinnedMesh and one AnimationMixer per character (three.js r187, WebGPURenderer, Chrome, Apple M-series).

Insight

Processor work per frame scales with bones × characters.

Bones per character Processor time per frame at 400 characters
12 about 5 ms
24 7 to 9 ms
48 16.6 ms
  • Triangles: going from 300 to 3,000 triangles changed only graphics-card time.
  • Backend: forcing the renderer’s WebGL 2 backend gave the same processor cost, so the WebGPU backend by itself bought nothing for this technique. The official per-instance skinning example still samples every pose on the processor.
  • Shadows: a shadow-casting point light roughly tripled the render call (425 to 2,915 draw calls at 400 characters), because the scene is drawn six more times.

How to measure

  • Measure milliseconds, split into mixer updates and the render call. Frame rate hides the margin behind the display cap.
  • Run-to-run noise was about 2 ms, so smaller differences mean nothing.
// Graphics-card time for the render pass
const renderer = new WebGPURenderer({ trackTimestamp: true });
await renderer.resolveTimestampsAsync("render");
// Same technique on the WebGL 2 backend, to compare
new WebGPURenderer({ forceWebGL: true });