Game Physics Just Got 170 Times Faster — Key Takeaways
A new elastic simulation method using precomputed co-rotated local perturbation subspaces runs 30–170x faster than Vertex Block Descent, enabling real-time deformable physics at 2.5 million elements.
Key takeaways
New FEM solver runs 30–170x faster than Vertex Block Descent
New FEM solver runs 30–170x faster than Vertex Block Descent
- 2.5M-element barbarian ship scene runs at 3 fps; 400K-element house of cards hits 30 fps in real time.
- Where VBD fails to converge entirely, this method completes the simulation successfully.
Precompute rest-shape Hessian per asset to enable real-time deformation
Precompute rest-shape Hessian per asset to enable real-time deformation
- Per-asset precomputation (7 min for small scenes, 67 min for large) runs once at build time, not at runtime.
- Baking this into game shipping pipeline means players see only the real-time result with no startup cost.
Co-rotated local perturbation subspace lets GPU parallelize deformable sims
Co-rotated local perturbation subspace lets GPU parallelize deformable sims
- Each mesh slice predicts its effect on the whole object before moving, eliminating the overshoot instability that broke prior parallel approaches.
- This finally makes GPU-parallel deformable simulation both stable and accurate — previously considered incompatible goals.
This Dig holds the full set of insights, 4 flashcards, and 2 quotes — free in Homestake.
Unlock this Dig freeFree forever · No credit card required
In this video
- 1mIntroduction and the core simulation challenge
- 1mThe overshoot problem and GPU opportunity
- 3mHow the new method works: local perturbation subspace
- 4mPerformance results and comparison to VBD
- 5mLimitations and pre-computation trade-offs
- 6mSponsor segment: Weights & Biases Weave
“This is about 30 to 170 times faster than VBD. What? That is insanely fast.”
This page is a partial, transformative summary produced by Homestake. All rights to the original content remain with its creator — please support them at the source link above.