Algebraic Multigrid GPU integration #527
Maksym Riabov
Hello Project Chrono,
I'm digging deep on how to make Project Chrono faster, and work on GPUs, specifically 3d FEM.
Is there a reason Chrono does not use Nvidia AmgX instead of it's CPU solvers? Algebraic multigrid on GPUs is 10-100x faster than the CPU solvers, and considering that solving equations takes around 70% of the solution time (to my knowledge), it's the easiest way to increase performance.
I know Algebraic Multigrid is successfully used in Comsol, for example, where it is a standard because of it's speed. Ansys uses it too.
Is there anything that stops us from integrating it to Chrono? All FSI, FEA, magnetic and other simulations can be vastly sped up with a simple API connection.
Do I not know something?
Cheers.
Dan Negrut
Thanks for the comment.
There is a place where AmgX might help the FSI solver and it seems that that might be investigated.
There is no active work on speeding up the 3D FEM. For very large problems, this might help but unfortunately, we don’t have the bandwidth to work in that part of the code right now.
Dan
---------------------------------------------
Bernard A. and Frances M. Weideman Professor
NVIDIA CUDA Fellow
Department of Mechanical Engineering
Department of Computer Science
University of Wisconsin - Madison
4150ME, 1513 University Avenue
Madison, WI 53706-1572
---------------------------------------------
--
You received this message because you are subscribed to the Google Groups "ProjectChrono" group.
To unsubscribe from this group and stop receiving emails from it, send an email to
projectchron...@googlegroups.com.
To view this discussion visit
https://groups.google.com/d/msgid/projectchrono/96781d03-73ad-455a-a32f-fd3964dbfd2an%40googlegroups.com.
Radu Serban
Indeed, we plan on looking at AmgX to replace the previous preconditioned iterative solvers for Chrono::FSI (because of deprecated functionality in other NVIDIA libraries). Of course, in FSI that’s relevant only for the implicit SPH solver.
At some point, we may consider wrapping an AMG solver for use in Chrono, but only after implementing some upcoming changes to the implicit integrators which meant to precisely address performance bottlenecks associated with linear system solves.
--Radu