Skip to content

Performance

Kinesis does the expensive parts once and skips work nobody will see. None of this needs any setup.

  • Geometry goes straight to the card. Particles, trails and other shapes are written to native memory and streamed in one upload per batch. Meshes are uploaded once and only moved, coloured and deformed afterwards.
  • Off-screen work is skipped. Particle tracks keep their bounds, and model and mesh tracks know their size, so anything outside the view costs almost nothing. Model re-renders are skipped entirely when the mob is off screen.
  • Particles use every core. Particle tracks that only touch their own state are simulated in parallel on worker threads while the render thread waits. Each track still uses its own seed, so every client sees the same thing.
  • Big tracks move to the GPU. With "simulation": "auto", the default, large additive tracks are simulated on the graphics card. See Particles.
  • One capture per mob per frame. Several model tracks on the same mob that use the same root bone and detail share one re-render of it.
  • Bloom only when something glows. Frames whose effects have no glow skip the bloom passes.
  • Textures have mipmaps. Detailed textures stay smooth at a distance instead of shimmering.
  • Prefer additive blending for large particle counts: alpha-blended particles must be sorted every frame, and "simulation": "auto" keeps them on the CPU.
  • Collisions, trails and particle-spawned effects cost the most per particle. Use them on the tracks that need them.
  • max is an upper bound, not a target, but it sizes buffers. Keep it near what the track can reach.
  • Model tracks re-render the mob every frame. Snapshots ("snapshot": true) render it once.