Menu

Discussion Forum

New Discussion
Filtrer

Discussion Closed This discussion was created more than 6 months ago and has been closed. To start a new discussion with a link back to this one, click here.

CUDA and COMSOL

Posted 11 sept. 2013, 05:30 UTC−4 MEMS & Nanotechnology, Fluid & Heat, MEMS & Piezoelectric Devices, Computational Fluid Dynamics (CFD), Microfluidics, Chemical Reaction Engineering, Cluster & Cloud Computing, Installation & License Management, Modeling Tools & Definitions, Parameters, Variables, & Functions, Studies & Solvers, Structural Mechanics Version 4.3b, Version 5.3 6 Replies

Stefano Firpi
Send Private Message Flag post as spam

Please login with a confirmed email address before reporting spam

Dear all, I have a new NVIDIA Quadro GPU and I would like to use it to accelerate my simulations, is it possible to use it somehow?
I looked for some informations and it seems to me that CUDA it's still unsupported, does someone have any fresh news?

I think this would be a very nice feature to be implemented in the future since nearly all the competitors (ansys, abaqus, nastran etc...) can benefit of GPU acceleration
6 Replies Last Post 31 juil. 2017, 09:40 UTC−4
COMSOL Moderator

Hello Stefano Firpi

Your Discussion has gone 30 days without a reply. If you still need help with COMSOL and have an on-subscription license, please visit our Support Center for help.

If you do not hold an on-subscription license, you may find an answer in another Discussion or in the Knowledge Base.

Martin Trebbin
Send Private Message Flag post as spam

Please login with a confirmed email address before reporting spam

Posted: 1 decade ago 27 nov. 2013, 06:29 UTC−5
+1!!!!!!!

yes, please integrate GPU solving. These days it's much easier to setup personal clusters and todays's GPU performance is astounding!
+1!!!!!!! yes, please integrate GPU solving. These days it's much easier to setup personal clusters and todays's GPU performance is astounding!
Stefano Firpi
Send Private Message Flag post as spam

Please login with a confirmed email address before reporting spam

Posted: 1 decade ago 27 nov. 2013, 06:54 UTC−5
I've talked extensively about this with the support.
Making a long story short, the problem with GPU performances is that they are NOT stable, so changing the physics, boudary conditions, geometry or whatever can give completely different results.
As far as I understood ( but I'm not affiliated with comsol in any way so this can be wrong) COMSOL decided not to support GPU processing because of those issues.
I've talked extensively about this with the support. Making a long story short, the problem with GPU performances is that they are NOT stable, so changing the physics, boudary conditions, geometry or whatever can give completely different results. As far as I understood ( but I'm not affiliated with comsol in any way so this can be wrong) COMSOL decided not to support GPU processing because of those issues.
Niklas Rom COMSOL Employee

Please login with a confirmed email address before reporting spam

Posted: 1 decade ago 3 déc. 2013, 09:33 UTC−5
Hi, just some clarifications about this interesting subject:

It is not COMSOL's opinion that GPUs are unstable or
unreliable.

We are following the GPU hardware development closely and when we find that
GPUs become useful to solve problems in COMSOL, we will support them.

Currently we see a number of factors that limit the
usefulness of GPUs for solving COMSOL problems. The main solvers in COMSOL
are the direct solvers and the multigrid solvers.

The direct solvers are very useful to handle strongly coupled smaller size
multiphysics problems, but they also require a lot of memory. Currently,
dedicated general purpose GPUs have about 5-16 GB of internal memory. This
corresponds to small problem sizes for the direct solvers, in the range
where they do not yield a high scalablity on GPUs.

For larger size problems, where multigrid is most efficient, the current
memory available in GPUs is not sufficient. In addition, multigrid solvers
use direct solvers on the coarsest grid, which also means that the
scalability of the coarse grid problem is limited on GPUs.

kind regards
Niklas Rom


I've talked extensively about this with the support.
Making a long story short, the problem with GPU performances is that they are NOT stable, so changing the physics, boudary conditions, geometry or whatever can give completely different results.
As far as I understood ( but I'm not affiliated with comsol in any way so this can be wrong) COMSOL decided not to support GPU processing because of those issues.


Hi, just some clarifications about this interesting subject: It is not COMSOL's opinion that GPUs are unstable or unreliable. We are following the GPU hardware development closely and when we find that GPUs become useful to solve problems in COMSOL, we will support them. Currently we see a number of factors that limit the usefulness of GPUs for solving COMSOL problems. The main solvers in COMSOL are the direct solvers and the multigrid solvers. The direct solvers are very useful to handle strongly coupled smaller size multiphysics problems, but they also require a lot of memory. Currently, dedicated general purpose GPUs have about 5-16 GB of internal memory. This corresponds to small problem sizes for the direct solvers, in the range where they do not yield a high scalablity on GPUs. For larger size problems, where multigrid is most efficient, the current memory available in GPUs is not sufficient. In addition, multigrid solvers use direct solvers on the coarsest grid, which also means that the scalability of the coarse grid problem is limited on GPUs. kind regards Niklas Rom [QUOTE] I've talked extensively about this with the support. Making a long story short, the problem with GPU performances is that they are NOT stable, so changing the physics, boudary conditions, geometry or whatever can give completely different results. As far as I understood ( but I'm not affiliated with comsol in any way so this can be wrong) COMSOL decided not to support GPU processing because of those issues. [/QUOTE]
John H.
Send Private Message Flag post as spam

Please login with a confirmed email address before reporting spam

Posted: 1 decade ago 4 déc. 2013, 05:15 UTC−5
Thank you, Niklas, for these very informative remarks. I've been asking myself this very question for quite some time (and even tried to run Comsol with the CudaBLAS libraries at some point).

Where I thought GPU support would be the most useful is the Particle Tracing module. Especially if there are no particle–particle interactions, this should be a highly parallelizeable problem. Then again, the specialized (competitor) software that we predominantly use for these kinds of simulations also doesn't support GPUs, so I guess I'm missing something.
Thank you, Niklas, for these very informative remarks. I've been asking myself this very question for quite some time (and even tried to run Comsol with the CudaBLAS libraries at some point). Where I thought GPU support would be the most useful is the Particle Tracing module. Especially if there are no particle–particle interactions, this should be a highly parallelizeable problem. Then again, the specialized (competitor) software that we predominantly use for these kinds of simulations also doesn't support GPUs, so I guess I'm missing something.
GuRu@BFI
Send Private Message Flag post as spam

Please login with a confirmed email address before reporting spam

Posted: 10 years ago 26 janv. 2015, 07:10 UTC−5
Hi, the CG-iterative solver is perfectly suited for GPUs. At least I would expect that Comsol support this particular CG-solver maybe even with a separate module. This will at least reduce the pure linear equation solution time drastically for such cases. Probably then, the major time would be consumed by the assemble stage.
Hi, the CG-iterative solver is perfectly suited for GPUs. At least I would expect that Comsol support this particular CG-solver maybe even with a separate module. This will at least reduce the pure linear equation solution time drastically for such cases. Probably then, the major time would be consumed by the assemble stage.
Mattias Skog
Send Private Message Flag post as spam

Please login with a confirmed email address before reporting spam

Posted: 7 years ago 31 juil. 2017, 09:40 UTC−4

Hi, just some clarifications about this interesting subject:

It is not COMSOL's opinion that GPUs are unstable or
unreliable.

We are following the GPU hardware development closely and when we find that
GPUs become useful to solve problems in COMSOL, we will support them.

Currently we see a number of factors that limit the
usefulness of GPUs for solving COMSOL problems. The main solvers in COMSOL
are the direct solvers and the multigrid solvers.

The direct solvers are very useful to handle strongly coupled smaller size
multiphysics problems, but they also require a lot of memory. Currently,
dedicated general purpose GPUs have about 5-16 GB of internal memory. This
corresponds to small problem sizes for the direct solvers, in the range
where they do not yield a high scalablity on GPUs.

For larger size problems, where multigrid is most efficient, the current
memory available in GPUs is not sufficient. In addition, multigrid solvers
use direct solvers on the coarsest grid, which also means that the
scalability of the coarse grid problem is limited on GPUs.

kind regards
Niklas Rom


I've talked extensively about this with the support.
Making a long story short, the problem with GPU performances is that they are NOT stable, so changing the physics, boudary conditions, geometry or whatever can give completely different results.
As far as I understood ( but I'm not affiliated with comsol in any way so this can be wrong) COMSOL decided not to support GPU processing because of those issues.



Is this something that is looked upon as a possible way to go for this? Or is that kind of memory not fast enough?

www.amd.com/en-us/press-releases/Pages/amd-radeon-pro-2016jul25.aspx

kind regards
Mattias
[QUOTE] Hi, just some clarifications about this interesting subject: It is not COMSOL's opinion that GPUs are unstable or unreliable. We are following the GPU hardware development closely and when we find that GPUs become useful to solve problems in COMSOL, we will support them. Currently we see a number of factors that limit the usefulness of GPUs for solving COMSOL problems. The main solvers in COMSOL are the direct solvers and the multigrid solvers. The direct solvers are very useful to handle strongly coupled smaller size multiphysics problems, but they also require a lot of memory. Currently, dedicated general purpose GPUs have about 5-16 GB of internal memory. This corresponds to small problem sizes for the direct solvers, in the range where they do not yield a high scalablity on GPUs. For larger size problems, where multigrid is most efficient, the current memory available in GPUs is not sufficient. In addition, multigrid solvers use direct solvers on the coarsest grid, which also means that the scalability of the coarse grid problem is limited on GPUs. kind regards Niklas Rom [QUOTE] I've talked extensively about this with the support. Making a long story short, the problem with GPU performances is that they are NOT stable, so changing the physics, boudary conditions, geometry or whatever can give completely different results. As far as I understood ( but I'm not affiliated with comsol in any way so this can be wrong) COMSOL decided not to support GPU processing because of those issues. [/QUOTE] [/QUOTE] Is this something that is looked upon as a possible way to go for this? Or is that kind of memory not fast enough? http://www.amd.com/en-us/press-releases/Pages/amd-radeon-pro-2016jul25.aspx kind regards Mattias

Note that while COMSOL employees may participate in the discussion forum, COMSOL® software users who are on-subscription should submit their questions via the Support Center for a more comprehensive response from the Technical Support team.

Suggested Content