CFD processing power

Here are our CFD links and discussions about aerodynamics, suspension, driver safety and tyres. Please stick to F1 on this forum.
0

Post Mon Mar 03, 2008 9:14 am

slimjim8201 wrote:Game physics and large scale simulation matrix solving are two completely different things. There is so much more going on within a CFD solver than with simple game physics. Even the most high end games employ relatively "simple" game physics.


That is true enough, however sparse matrix solving is something extremely well suited to the GPU. That component is essentially a very fast vector processor with fast access for random memory reading. This is exactly what you want for CFD, but the technology isn't ripe enough yet to allow for simple adaptation of existing codes. GPUs are likely much better suited to the engineering applications than the usual out-of-order CPU designs, but ease of use is still lacking.

A very similar thing is true of the Cell processor in the PlayStation 3. It's excellent at processing a stream of information rather than a sequence of various different instructions, which is exactly what you usually need in physics and engineering problems. Look at this story for an example (which is really all about solving partial differential equations):

http://www.wired.com/techbiz/it/news/20 ... ercomputer

It is not that these kinds of processors are unsuitable for engineering applications, quite the contrary, the problem is that it's still emerging technology that will take a while to become adopted. I do expect many more vector processors being used for serious computations in the next few years, though.
Gecko
1
 
Joined: 5 Sep 2006

0

Post Mon Mar 03, 2008 3:26 pm

I did some reading last night that comes to similar conclusions on GPU's/PPU's and the Cell processors. So one day soon it may be possible to play video games and solve CFD at the same time.
:D Sounds good to me. :D

http://www.blachford.info/computer/articles/PhysX2.html
http://www.nytimes.com/2006/02/09/techn ... ref=slogin Cell
http://www.blachford.info/computer/articles/PhysX3.html PPU & Cell
http://en.wikipedia.org/wiki/Physics_processing_unit
http://en.wikipedia.org/wiki/Physics_engine
Carlos
8
User avatar
 
Joined: 2 Sep 2006
Location: Canada

0

Post Mon Mar 03, 2008 5:26 pm

I've worked in both the CFD and game industries.

There is no doubt that the GPUs on modern-day graphics cards make excellent computing engines compared to standard CPUs. However, current industrial CFD has not begun to exploit this resource. One of the reasons is that the GPU is optimized to process pixels on a regular grid matching the pixels on a screen. This is great for simulating simple 2D fluid flow in a box, but inadequate for 3D flow around a complex shape such as an F1 car. That said, I've seen 3D GPU fluid simulations that stack 2D slices to simulate 3D flow in a cube. The problem still remains that regular spaced grids for complex shapes are inadequate to resolve features, both geometric and flow related that may vary by over 3 orders of magnitude (1mm - 4m) on a car.

When the game and movie industries talk about CFD, it is a specialized form that is optimized to 'look right' and most often is related to free-surface water flow. It is typically far removed from a physically accurate simulation good for extracting quantities such as downforce and drag. Many so-called fluid simulations in games and movies are actually particle simulations sharing more with Lattice Boltzmann equations rather than the Reynolds-Averaged Navier-Stokes equations solved in industrial CFD packages.

To find out more try: http://www.symscape.com/blog/real_time_simulation

There is every chance that innovative uses of the GPUs will be made by industrial CFD, but that is not currently the case.
Symscape, Computer-Aided Engineering for all
syguy
0
User avatar
 
Joined: 22 Feb 2007
Location: USA

0

Post Mon Mar 03, 2008 6:59 pm

Return email from the OpenFOAM team concerning the use of GPU CrossFire. My return email was to ask about starting this as an OpenFOAM Team project. This email tells me that what little objection he has raised, that there are solutions available.

****************
Dear Chris Knopp,

exciting things seem indeed to come from the graphics hardware industry!

To answer your question in short: no, there is no support for ATI's CrossFire.

Our problems tend to consist of a setup phase (calculating matrix
coefficients) followed by a linear matrix solver phase. The solver will take
up anywhere from 20 to 90% of the overall time. We know of linear solvers
having been coded to run on GPU/cell processor architectures and expect the
ATI Physics engine to fall into the same category.

Our problem is that our solver uses indirect addressing (and distributed data
exchange at the solver level). The indirect addressing means that the memory
gets accessed all over the place so can badly be cached. We have not heard of
indirect addressing based solvers running (efficiently) on GPU type engines
but this is not really our field. Do you know what currently is going on? Is
there a single language that both NVidia and ATI support? Running on Linux?
Is double precision supported? What open-source solvers are available?

Also interesting could be to use the GPU as a multi-core machine with fast/low
latency memory access and run all of the code on it instead of just the
linear solver. (our code does run in parallel, using e.g. MPI to do
communication) The current generation of generic multi-core processors seems
to be memory limited (e.g. on a 5300 series quad core we get say a factor of
2 speedup) but this might get solved with newer generations.

On our forum there have been discussions on using the cell processor but I am
not sure what the state of that is.

Kind regards,

Mattijs

- Show quoted text -

On Monday 03 March 2008 02:21, Chris Knopp wrote:
> Is there support for ATI's CrossFireX Physics Processing Unit platform as
> of yet? I am incredibly interested in any info that you may have on this
> subject, since I cannot find ANY information that it has even been tried as
> of yet. This could be industry breaking if it works as well as the
> whitepaper states.
>
> http://ati.amd.com/technology/crossfire ... s_Processi
>ng_with_ATI_CrossFire.pdf
>
> I am willing to build the machine to test it on, so please let me know
> ASAP!
>
> Chris Knopp


--

Mattijs Janssens

OpenCFD Ltd.
9 Albert Road,
Caversham,
Reading RG4 7AN.
Tel: +44 (0)118 9471030
Email: M.Janssens@OpenCFD.co.uk
URL: http://www.OpenCFD.co.uk
********************
Conceptual
0
 
Joined: 15 Nov 2007

0

Post Mon Mar 03, 2008 7:46 pm

Maybe these instead of the Video cards?

http://en.wikipedia.org/wiki/AMD_Stream_Processor

Chris
Conceptual
0
 
Joined: 15 Nov 2007

0

Post Mon Mar 03, 2008 8:05 pm

Saribro wrote:
Conceptual wrote:If you read that white paper (http://ati.amd.com/technology/crossfire ... ssFire.pdf) you will see that they specifically mention fluid dynamics as what it does incredibly well, by their standards.

It's not a white paper, it's advertisement :).
Also, notice how they only mention "Game Computing" and "Game Physics Processing".
GPUs can do a lot more than they used to, but not everything yet.

This better?

http://ati.amd.com/companyinfo/research ... _Guide.pdf

Chris
Conceptual
0
 
Joined: 15 Nov 2007

0

Post Mon Mar 03, 2008 10:35 pm

Looks like the OpenFOAM project cannot happen due to the fact that the CrossFire SDK is WinXP only.

Does anyone have any other ideas on if there are any opensource windows platfrom CFD software? And if so, please post a link.

Thanks!

Chris
Conceptual
0
 
Joined: 15 Nov 2007

0

Post Mon Mar 31, 2008 2:33 pm

I've recently been working on a native Windows version of OpenFOAM, built using a cross-compiler running under Linux. Though, it would still be a lot of work to support a GPU (stream processor), it's a step in the right direction.

For details visit: http://www.symscape.com/openfoam-on-windows
Symscape, Computer-Aided Engineering for all
syguy
0
User avatar
 
Joined: 22 Feb 2007
Location: USA

0

Post Mon Mar 31, 2008 11:07 pm

syguy wrote:I've recently been working on a native Windows version of OpenFOAM, built using a cross-compiler running under Linux. Though, it would still be a lot of work to support a GPU (stream processor), it's a step in the right direction.

For details visit: http://www.symscape.com/openfoam-on-windows



Awesome site, and shows the most potential to be used on the crossfire platform.

Chris
Conceptual
0
 
Joined: 15 Nov 2007

0

Post Tue Apr 01, 2008 12:19 am

I don't want to hi jack this thread but is there anybody out there, anyone at all that would be able to run my model through of my electric car, im just finalising the body at the moment and i don't want to discover its rubbish aero wise.

I can convert my model into the following formats,
VRML, Parasolid transmit file, Stereo Lithography File, IGES File, Medusa Model file. I made it in pro/Desktop a couple of months ago.

Thanks in advance!!!
______________________________________
greenpower dude reloaded
6
User avatar
 
Joined: 29 Dec 2005
Location: Portsmouth, UK

0

Post Tue Apr 01, 2008 1:01 am

If you want you can have a go yourself with a 30 day free trial of SymLab Professional. Or if you prefer I can give it a go for you. If you don't mind I'd like to publish a picture or two of the final results. If you are interested contact me at:
http://www.symscape.com/contact

In terms of geometry exchange formats I'd recommend using STEP and fall back to IGES if STEP isn't available. For more thoughts on geometry exchange try: http://www.symscape.com/node/606

Note that I'd be using a Panel Method and not OpenFOAM.
Symscape, Computer-Aided Engineering for all
syguy
0
User avatar
 
Joined: 22 Feb 2007
Location: USA

0

Post Sat Apr 12, 2008 2:42 am

Conceptual wrote:OpenFOAM sounds decent, and if there is a support forum, I'm sure that I can learn how to use it. Although I am kinda bummed by being told that I should leave this for the CFD experts.


It's not because we don't think you can learn CFD - it's because there is no way you can use open foam without a SERIOUS understanding of the theory and programming.

Trust me - you will be wasting a lot of time with openfoam. Get a commercial package or a crack for a commercial package or whatever. Openfoam = bad idea.
AeroGT3
0
 
Joined: 29 Mar 2006

0

Post Sat Apr 12, 2008 3:58 am

AeroGT3 wrote:
Conceptual wrote:OpenFOAM sounds decent, and if there is a support forum, I'm sure that I can learn how to use it. Although I am kinda bummed by being told that I should leave this for the CFD experts.


It's not because we don't think you can learn CFD - it's because there is no way you can use open foam without a SERIOUS understanding of the theory and programming.

Trust me - you will be wasting a lot of time with openfoam. Get a commercial package or a crack for a commercial package or whatever. Openfoam = bad idea.


I ended up dropping this idea, since I was looking at a $5000 PC driving my projector in my living room. I'm sure they blu-ray would have looked nice, but way more than is necessary without someone that is willing to actually use it.

Thanks tho, I appreciate your input.

Chris
Conceptual
0
 
Joined: 15 Nov 2007

0

Post Sat Apr 12, 2008 8:59 am

Panel methods are borderline worthless for F1 type simulations. Without the ability to predict viscous flows and flow separation, you might as well draw an F1 car on a napkin and blow air across it.

syguy wrote:If you want you can have a go yourself with a 30 day free trial of SymLab Professional. Or if you prefer I can give it a go for you. If you don't mind I'd like to publish a picture or two of the final results. If you are interested contact me at:
http://www.symscape.com/contact

In terms of geometry exchange formats I'd recommend using STEP and fall back to IGES if STEP isn't available. For more thoughts on geometry exchange try: http://www.symscape.com/node/606

Note that I'd be using a Panel Method and not OpenFOAM.
slimjim8201
9
User avatar
 
Joined: 30 Jul 2006

0

Post Sat Apr 12, 2008 3:22 pm

Sorry I should have made it clearer that my comment on the Panel Method above was directed at the electric racecar (closed-wheel) comment by greenpower dude reloaded, not F1.

slimjim8201 wrote:Panel methods are borderline worthless for F1 type simulations. Without the ability to predict viscous flows and flow separation, you might as well draw an F1 car on a napkin and blow air across it.


You're comment concerning Panel Methods for intricate F1 geometries isn’t far off the mark - if a little harsh. You must also be aware of boundary layer methods that can enhance Panel Methods by accounting for localized viscous effects near walls - obviously no match for full CFD and its ability to predict massive separations.

Anyway, for closed-wheel (streamlined) cars Panel Methods are useful as a first pass simulation, due to their speed and ease of setup - i.e., no volume grid and short run-times.

By the way your comment applies equally well to poorly resolved (meshed) models using full CFD. To produce an accurate F1 CFD simulation (such as those produced by Advantage CFD) is a non-trivial undertaking, as you no doubt know given your affiliation with CFdesign.
Symscape, Computer-Aided Engineering for all
syguy
0
User avatar
 
Joined: 22 Feb 2007
Location: USA

PreviousNext

Return to Aerodynamics, chassis and tyres

Who is online

Users browsing this forum: Paul, turbof1 and 7 guests