nVidia's new Temporal Gauss-Seidel (TGS) solver in PhysX 4

Post Reply
mpm
Posts: 6
Joined: Mon Apr 17, 2017 12:11 pm

nVidia's new Temporal Gauss-Seidel (TGS) solver in PhysX 4

Post by mpm » Wed Dec 18, 2019 9:11 pm

So Nvidia introduced a PGS derivative called TGS with much improved convergence rate over their old PGS solver and adding support for much larger mass ratios. Videos look impressive, esp. for difficult constraint types like prismatic, and so far some testing (IIRC our very own Erwin Coumans found some instabilities with contacts) shows that it during iterations overshoot a little before converging, resulting in instabilities for constraints with very low or zero tolerances.

Has anyone gained any in-depth knowledge of how this algorithm works?
I suppose it anticipates impulse responses along critical paths sort of like auxiliary constraints used for closing open-loop chains, or tries to propagate impulses between multiple bodies faster.

The most official statement discussing its workings that I could find is this https://devtalk.nvidia.com/default/topi ... gs-solver/ which states that they should release a paper on it in the near future, but that has yet to happen.

Dirk Gregorius
Posts: 875
Joined: Sun Jul 03, 2005 4:06 pm
Location: Kirkland, WA

Re: nVidia's new Temporal Gauss-Seidel (TGS) solver in PhysX 4

Post by Dirk Gregorius » Tue Jan 28, 2020 10:37 pm


Post Reply