Do any of the GPU accelerators for OpenFOAM actually work for real cases?

kyle · January 26, 2014, 13:05

It seems they all fall into one or more of the following categories:

- Don't work with (or don't work well with) the solvers that most people use (GAMG), so they publish benchmarks using a less efficient solver for the CPU baseline (PCG)
- Suspiciously only have benchmarks for 2D structured meshes
- Require adding $5000 in GPUs to a $1000 computer to see a 10% speedup
- Third parties unable to reproduce published benchmarks

I'm sure we will get there one day, but it doesn't seem like we are at the point where you can cost-effectively accelerate OpenFOAM with a GPU. Is this your general feeling as well?

wyldckat · January 26, 2014, 13:59

Greetings Kyle,

Last year Symscape made this report: http://www.symscape.com/blog/cfd-per...etween-gpu-cpu - the conclusion was that a machine with a good CPU can leave almost any GPU in the dust. AFAIK, this methodology essentially replaces the CPU matrix related solvers for GPU based ones.

Vratis on the other hand noticed the same issue longer ago and went in another direction (if my memory doesn't fail me): to implement the complete SIMPLE and PISO loop directly in GPU.

More on these and similar topics: http://www.cfd-online.com/Forums/ope...tml#post403724 post #12 - have a look at the second link, which provides some of the links for presentations on Vratis' developments, namely AREAL and SpeedIt.

As for the feeling of things: I would say that in 1 to 2 years time, all of this confusion might be for not much. I say this based on this thread: http://www.cfd-online.com/Forums/har...aptop-cpu.html - if the CPU development market gets off their commercial comfort zones, we could start seeing CPU's with high powered embedded GPUs with some considerable RAM embedded on chip, acting as L4, therefore making external graphic cards redundant. And then there is the AVX feature on CPUs: http://en.wikipedia.org/wiki/Advanced_Vector_Extensions

On the other hand, it might be the GPU cards that get the jump on the CPUs, since there are already cards with 6 GB DDR5 on the 1000 Euro mark. I'm still waiting for NVidia to come out with an ARM+GPU card, which turns a graphics card into its own motherboard... then we SLI 4 cards together and then we do have a good competition against CPUs

edit: Only now did I notice it was you who created the thread about the Intel CPU!

Best regards,
Bruno

elvis · January 28, 2014, 06:05

Hi,

the OF workshop is always a good source to look for information on that topic
http://www.openfoamworkshop2013.org/...p?Main=2&sub=1
http://openfoam-extend.sourceforge.n...7_Program.html

I think there was also something from the japanese OF users on that topic but I do not have a link to that.

January 26, 2014, 13:59		#2
wyldckat Retired Super Moderator Bruno Santos Join Date: Mar 2009 Location: Lisbon, Portugal Posts: 10,982 Blog Entries: 45 Rep Power: 128	Greetings Kyle, Last year Symscape made this report: http://www.symscape.com/blog/cfd-per...etween-gpu-cpu - the conclusion was that a machine with a good CPU can leave almost any GPU in the dust. AFAIK, this methodology essentially replaces the CPU matrix related solvers for GPU based ones. Vratis on the other hand noticed the same issue longer ago and went in another direction (if my memory doesn't fail me): to implement the complete SIMPLE and PISO loop directly in GPU. More on these and similar topics: http://www.cfd-online.com/Forums/ope...tml#post403724 post #12 - have a look at the second link, which provides some of the links for presentations on Vratis' developments, namely AREAL and SpeedIt. As for the feeling of things: I would say that in 1 to 2 years time, all of this confusion might be for not much. I say this based on this thread: http://www.cfd-online.com/Forums/har...aptop-cpu.html - if the CPU development market gets off their commercial comfort zones, we could start seeing CPU's with high powered embedded GPUs with some considerable RAM embedded on chip, acting as L4, therefore making external graphic cards redundant. And then there is the AVX feature on CPUs: http://en.wikipedia.org/wiki/Advanced_Vector_Extensions On the other hand, it might be the GPU cards that get the jump on the CPUs, since there are already cards with 6 GB DDR5 on the 1000 Euro mark. I'm still waiting for NVidia to come out with an ARM+GPU card, which turns a graphics card into its own motherboard... then we SLI 4 cards together and then we do have a good competition against CPUs edit: Only now did I notice it was you who created the thread about the Intel CPU! Best regards, Bruno __________________ OpenFOAM: FAQ \| Getting started Forum: How to get help, to post code/output and forum guide Read this before sending me PM Last edited by wyldckat; January 26, 2014 at 14:04. Reason: see "edit:"

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
ATTENTION! Reliability problems in CFX 5.7	Joseph	CFX	14	April 20, 2010 16:45
Cases with small length scale work fine on a single processor but fail in parallel	adona058	OpenFOAM Bugs	5	April 17, 2009 05:41
New Nvidia gpu aimed at gpgpu	bmeagle	OpenFOAM	0	November 9, 2006 10:41
Why do the Plant library cases don't work?	Alumna	Phoenics	6	June 22, 2004 13:08

January 26, 2014, 13:05	Do any of the GPU accelerators for OpenFOAM actually work for real cases?	#1
kyle Senior Member Join Date: Mar 2009 Location: Austin, TX Posts: 160 Rep Power: 18	It seems they all fall into one or more of the following categories: - Don't work with (or don't work well with) the solvers that most people use (GAMG), so they publish benchmarks using a less efficient solver for the CPU baseline (PCG) - Suspiciously only have benchmarks for 2D structured meshes - Require adding $5000 in GPUs to a $1000 computer to see a 10% speedup - Third parties unable to reproduce published benchmarks I'm sure we will get there one day, but it doesn't seem like we are at the point where you can cost-effectively accelerate OpenFOAM with a GPU. Is this your general feeling as well?

January 28, 2014, 06:05		#3
elvis Senior Member Elvis Join Date: Mar 2009 Location: Sindelfingen, Germany Posts: 620 Blog Entries: 6 Rep Power: 24	Hi, the OF workshop is always a good source to look for information on that topic http://www.openfoamworkshop2013.org/...p?Main=2&sub=1 http://openfoam-extend.sourceforge.n...7_Program.html I think there was also something from the japanese OF users on that topic but I do not have a link to that.