Discussion Forum

For example, a Colfax CXT300N uses an AMD processor, can have up to 4 C1060's and will run RedHat Linux. This meets some of COMSOL's requirements, but what other kind of glue software might be needed to make the connection to the hardware?

Hi,

I don't think that COMSOL 4.0 can use the GPU technology. You need an application for that. The architucture is different.
Comsol people should think about that soon, because this is the future for parallel computation. Especially that the GPU prices are so low comparing to the clusters.
But I think that Matlab has some application which can be ran on GPU station.


Regards

Thanks. I found "Jacket" which allows MATLAB to use GPU's. A program called Accelerware has some GPU glue to FDTD EM solvers, so I'll be looking at that too.

Hi,

I don't think that COMSOL 4.0 can use the GPU technology. You need an application for that. The architucture is different.
Comsol people should think about that soon, because this is the future for parallel computation. Especially that the GPU prices are so low comparing to the clusters.
But I think that Matlab has some application which can be ran on GPU station.


Regards

Indeed. However, it seems to me that CUDA is still in not mature enough to be used for C++ development and requires much code changes only aimed at NVidia chips. So it costs a lot to port a code. However, soon enough I think CUDA will come closer to c++, and then more people will be willing to port their i.e. MPI codes to CUDA.

I am looking forward to it myself.

Hate to say this, since I would love to see this feature in COMSOL, but who wants to maintain two versions of their code? To get maintstreat adoption I think it is up to NVIDIA to make this REALLY easy, not on software manufacturers to rewrite their programs.

Ideally, the software would have no changes. Realistically, maybe you surround routines that are appropriate to offload to the GPU in some kind of designator block. I don't see why you should need to do more than that. That being said, I don't know C++ or CUDA, so it would be nice if COMSOL could provide information on their plans here.

I will REALLY show my age here.

The nvidia cpus remind me so much of the old CDC 60-bit machines. For those of you too young to know, CDC is the predecessor to CRAY. The CDC 7600 had a high-performance region of the machine that fit in 256K of memory. The object of the game was to fit your part of the code that used the most cpu into this faster-running core and gain performance. It required a major rewrite, or development, of code to take advantage of this. Of course, this machine is long gone, and now the codes that were written for that machine are obsolete.

My opinion is that if you want to take advantage of the nvidia cpus, it should be done in the solver libraries that COMSOL calls, but do not rewrite the major part of the code to use this new hardware. 10 years from now, it will be a different hardware out to use.

Did anyone of you tried comson on PS3/Linux machine? If so, is there any remarkable performance incraase for those models with little memory requierements. PS hat 256MB of ram but its cpu has 6 cores that Linux can access.