ThreeDream - GPU soft bodies

no adapter cloth 0 nodes cpu-upload 0 fps

Probing adapters…

Tier

running on: —
asked for: —
renderer backend: —
device shared: —
vertex path: —
blitted / frame: —
race-free solve: —
feature level: —

Not built yet.

Simulation

solver iterations 8

stiffness 1.00

Iterations are how far a correction travels, not how stiff the material is: each sweep moves a constraint fix about one row of the mesh, so the 10k cloth hung from its top edge keeps stretching near the pinned row until it has had enough of them. Watch max stretch and kinetic energy as you drag.

Graph

nodes: —
constraints: —
islands: —
color batches: —
node workgroups: —
dispatches / step: —
draw mode: —
workgroup size: —

Counters

steps / frames: —
steps per frame: —
ms / frame: —
ms / step: —
draw calls: —
triangles: —
escaped: —
max stretch: —
max speed: —
islands awake: —
kinetic energy: —
digest: —

Log

The solver and the renderer share one GPUDevice, and the publish buffer is copied straight into the GPUBuffer three.js already allocated for the position attribute — no node crosses the bus, and the surface and its wire overlay share that one attribute so a single blit fills both. Constraints are solved in colour batches, never two edges of one colour at once, which is what makes the GPU solve race-free without being bit-reproducible. SoftView never writes the simulation, so replay and training stay independent of this layer; e2e/soft_gpu.spec.ts asserts CPU/GPU parity.