Video board as additional memory

Cleiton · November 6, 2009, 11:21

Dear users,
I've received an information that it's possible to use the video card as additional memory to run simulations, and they would run about 15 times faster. If this is possible, is there some way to configure CFX to run serial/parallel simulations on video card?

stumpy · November 6, 2009, 11:41

Short answer - no :-)

ghorrocks · November 7, 2009, 05:48

Here's a slightly longer answer -

You are referring to the GPU based systems such as NVidia's CUDA (http://www.nvidia.com/object/cuda_home.html#). CFX does not support CUDA or anything like it, and this will not change for some time. These systems need highly parallelised software to run well and in CFD that means explicit solvers. CFX is an implicit solver (and so is 95% of all the commercial CFD codes) so it will never run well on GPU systems. I know ANSYS is doing a research project to investigate this technology but it will not be implemented for many years yet (if at all).

There is no commercial CFD software currently available to my knowledge which runs on GPU systems. This is a big opening for new players in the CFD market as the established players have such a big investment in CPU based systems so a nimble footed new player could grab some market share here. There are some preliminary open source codes around on GPU systems, surely a more general purpose commercial code will be around sometime soon.

And always keep in mind that GPU based systems will work best with explicit solvers and that means simulations of very fast dynamics. The traditional implicit solvers running on CPU based systems are likely to be best for slower time scale simulations and steady state simulations.

Definitely a thing to keep a watch on as things develop.

November 6, 2009, 11:21	Video board as additional memory	#1
Cleiton New Member Cleiton Porciuncula Join Date: Nov 2009 Posts: 1 Rep Power: 0	Dear users, I've received an information that it's possible to use the video card as additional memory to run simulations, and they would run about 15 times faster. If this is possible, is there some way to configure CFX to run serial/parallel simulations on video card?

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
How to optimize the memory usage when using FEM	vasilis	Main CFD Forum	11	August 24, 2009 23:57
Memory error?	Young	CFX	3	September 30, 2008 11:33
CFX CPU time & real time	Nick Strantzias	CFX	8	July 23, 2006 17:50
Temperature Distribution on a circuit board	Blackadder	Main CFD Forum	1	March 27, 2003 11:32
Replace periodic by inlet-outlet pair	lego	CFX	3	November 5, 2002 20:09

November 6, 2009, 11:41		#2
stumpy Senior Member Join Date: Apr 2009 Posts: 531 Rep Power: 21	Short answer - no :-)

November 7, 2009, 05:48		#3
ghorrocks Super Moderator Glenn Horrocks Join Date: Mar 2009 Location: Sydney, Australia Posts: 17,703 Rep Power: 143	Here's a slightly longer answer - You are referring to the GPU based systems such as NVidia's CUDA (http://www.nvidia.com/object/cuda_home.html#). CFX does not support CUDA or anything like it, and this will not change for some time. These systems need highly parallelised software to run well and in CFD that means explicit solvers. CFX is an implicit solver (and so is 95% of all the commercial CFD codes) so it will never run well on GPU systems. I know ANSYS is doing a research project to investigate this technology but it will not be implemented for many years yet (if at all). There is no commercial CFD software currently available to my knowledge which runs on GPU systems. This is a big opening for new players in the CFD market as the established players have such a big investment in CPU based systems so a nimble footed new player could grab some market share here. There are some preliminary open source codes around on GPU systems, surely a more general purpose commercial code will be around sometime soon. And always keep in mind that GPU based systems will work best with explicit solvers and that means simulations of very fast dynamics. The traditional implicit solvers running on CPU based systems are likely to be best for slower time scale simulations and steady state simulations. Definitely a thing to keep a watch on as things develop.