Duplicating comment on other thread: I've also talked with @rserban over a month ago about porting FEM to GPU, he said:
>But even more importantly, it is not clear to me that it is worth it and that one would get the performance gains you seem to expect from a port to GPU. FEA does not map well to the GPU programming paradigm (unlike granular dynamics or SPH). For the types and sizes of FEA problems we’re interested in, it is unlikely a port to GPU is worth it. Because, as far as I understand, he means that it because FEM is traditionally a sequential execution program, not parallel execution. But AMG methods are parallel for whichever equations are given. OK, I'm not super-versed in FEM solver mathematics, but from what I've explored in FEM (and I explored thoroughly), there is nothing that can't be done. Are we stopped by literally anything but plugging it in? Cheers. On Tuesday, July 16, 2024 at 9:23:40 PM UTC+1 Maksym Riabov wrote: > Sup Chrono, > > I’m writing an open-source simulation library dealing with multibody, > evolve-over-time scenes - basically Project Chrono but with a twist: I’m > using voxels mapped from SDFs. > > Throughout my research I’ve found that (theoretically) using voxels for > FEA simulation is an order of magnitude (10x) faster than meshes with > similar precision. Same goes for SDFs. Both are much faster in fluid sims > too. Using adaptive upvoxelization I think I’ll solve the problem of voxels > having problems with contact. > > I need the speed because I’m using it for a machine learning app, so, > ideally, a single good GPU (80 gb memory) should be able to run hundreds, > or even thousands of the simulations in parallel. > > So… You might notice things that I don’t - I’m a mecheng, and I’ve never > dealt with the computational part. > > SDFs for collision detect and voxels for bending and all the physics - is > this crazy? > > Thanks everyone, > Maksym Riabov > -- 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 [email protected]. To view this discussion visit https://groups.google.com/d/msgid/projectchrono/c6d31929-a145-484a-93a1-a97d5bfb851cn%40googlegroups.com.
