20 replies to this topic

[redline]

Hi All,

I believe the majority of aerodynmicists carrying out CFD work in racing are using Fluent as a solver. I was curious as to which pre-processor is most commonly used for mesh generation - GAMBIT or TGRID? What are the differences between the two in terms of meshing capabilities?

Thanks,

redline

[david_martin]

I think you will find just about everybody has switched to Gambit -- it has all of the mesh generation features of tgrid, plus a better import engine for Catia CAD data and Ideas solid models, which I would guess is the "killer feature" for the F1 guys. Gambit also offers a better environment for generating solver input files for parametric runs and sensitivity analysis than tgrid can do, which would be a bonus where a lot of the work would be solving the same mesh over different boundary and initial conditions.

[Chevy II Nova]

Any idea how much a CFD system costs to set up from scratch? As I understand, Fluent charges a hefty $100,000 license for a single machine. Factoring in a pre-proccessor and hardware, how much would one need to spend on the setup?

[gug]

dont know if this is relevant to your question or not, but we have a copy on every computer in the engineering suites (about 100 or so) of fluent and gambit. dont know how much the uni paid for them, but i suspect there is a student version available for a tiny fraction of that cost quoted above. or you could pay a hacker maybe $100 and they would crack it for you, working off a student copy. i wonder if f1 teams are secretive enough to get away with running illegal software?

anyway, fluent and gambit run fine on typical pcs. well, they are supposed to run fine. last time i saw someone trying to work it they were importing wing profiles from solid edge (a solid modeling program), so the wings (4 element, uniform profile across each element) consisted of several thousand flat faces that made up the curves. naturally, the computer got its nickers in a knot and stopped working.

well, 1 processor in a pc is fine for a student (a student that can be bothered modeling wings in gambit anyway ), but how would an f1 team go about it? i know windows makes a copy of win2k that uses dual processor motherboards. since fluent runs under linux (well, the student copy anyway), i imagine that linux can be programmed up to use as many cpus as required (to a point). more processing power means that it takes less time to analyze a model. i very much doubt they use a supercomputer or any such thing. so the amount required to set up the processor would be determined by how much accuracy they need, but i doubt it would exceed $20k.

er, quick glossary here for those who dont know much about computers.
cpu = part of the computer that does all the calculations for things such as fluent. average pcs have 1, a few specialised pcs have 2. they are the "pentium 3, 2.4 GHz" things.
linux = similar to microsoft's windows, but it is open code. this means that everyone can see the actual instructions used to write the software. you could say seeing microsoft's source code is a bit like seeing every internal on an f1 car. only, microsoft's source code is worth $50 billion or whatever microsoft is worth today.

im an engineer studying half electrical engineering and half mechanical engineering, so ive read a lot of books about this. thats code for "i think i know a lot, but i really dont know anything".;) so feel free to correct me if you think you know better.

[Chevy II Nova]

Wow. My understanding was that it all required a ton a proccessing power to run. How long do you rekon it would take to model and test something as detailed as an F1 car?

[DOHC]

Originally posted by Chevy II Nova
Any idea how much a CFD system costs to set up from scratch? As I understand, Fluent charges a hefty $100,000 license for a single machine. Factoring in a pre-proccessor and hardware, how much would one need to spend on the setup?

It depends on what kind of CFD you want to do. Fluent is expensive, and it takes time to learn how to use it. If you're going to do problems with turbulence, it might be your choice. For laminar flow and moderate Reynolds numbers Femlab could be a better choice. Getting started in Femlab is easy, modelling is a lot faster, meshing and remeshing is simple, and it interfaces nicely to Matlab. Depending on what kind of license you can get, Femlab will cost something like 2-5 kUSD for a single PC version.

[gug]

well, all my experience comes from designing a fsae car. similar to one like this:
.
im not actually designing the wings, i just watch and laugh as the aero guys swear and hit the computer.

a ton of processing power is very much required. as far as i know, CFD (computational fluid dynamics) is the most processing power intensive thing an engineer can do on a computer. i would be surprised if a f1 car took more than 10 hours of processing time though. i suspect they would set the accuracy of the analysis to take that long, so they can leave it overnight. its sort of a natural thing to do when you know you can get a more accurate result by leaving the computer working for longer, i do it all the time.

but then again, having no experience at f1 engineering, i shouldnt really comment on how accurate they require their results.

to compare it to fsae, the rear wing in the picture above would take about 5 mins of processing time before you got about as accurate a test as CFD will give you. start modeling the whole car and it might take 30mins for a rough model, maybe several hours for a very accurate model (modeled radiator vanes, bolts, mounting brackets, yada yada yada, if anyone actually bothers to do this). the main errors in CFD come about because of things such as surface roughness (although surface roughness can be modeled, it cant be modeled exactly) and other small factors. CFD is very useful for developing an idea and backing up principals, such as "this standard NACA aerofoil will give us less downforce than this custom shape", but when it comes to things such as a complete f1 car, nothing beats a windtunnel. an f1 team could have all the processing power in the world and they would still put their car in the windtunnel, because it is much more useful.

actually, i just had a thought. there are computers designed purely for rendering images (making very lifelike pictures from computer models). as i understand, their processors can understand different instructions than standard ones which makes the rendering process much quicker. maybe there are processors out there for CFD like this?

anyway, im a student building my first racing car, so if someone out there is more informed than me, feel free to disagree!

oh yeah, you want to muck around with cfd? a simple 2D program called javafoil is available on the net, just use google. very simple, and contains some bugs, but its interesting anyway.

[david_martin]

Originally posted by gug
I would be surprised if a f1 car took more than 10 hours of processing time though

On what basis and on what class of machine? We regularly work with what I would consider to be small or medium sized meshes solving 3D steady state combined heat and mass transfer problems with combustion on either very fast two CPU machines or on a small cluster, both running IA-32 Linux. Solution times can easily run into hundreds of hours....

[Chevy II Nova]

Wow... so a team like Minardi would have to invest in some very high end hardware to run a CFD program for their designs?

[Ben]

Originally posted by gug
i would be surprised if a f1 car took more than 10 hours of processing time though.

I've done DOEs on rigid body ADAMS models that have taken 7.5hours and I don't think its useful to make such generalisations about CAE tools. As has been pointed out it depends so much on the hardware and how the model's set up

Ben

[Greg Locock]

Yeah, I'd have to say that almost everything gug wrote in his first post disagreed with my experience. We examine aero problems on cars with a /Cray/ for goodness sake.


hey Ben, 7.5 hours for an ADAMS DOE is nothing, I have to do one job for each model where I leave it running for a week! (it's about 170 runs, I think).

[gug]

fair enough! like i said, im a student without much experience. sorry for misleading you.

[Ned]

Hi!

I am new here. Greetings to everyone.

Besides the geometry, I think it depends on how fine of a grid and what turbulence model you are using, what speed you have and all kinds of stuff. It might depends less on the actualy size of the model but it actually depends more on the number of cells in the grid. It also depends on how manys steps you want to run it of course.

As far as I know some 3-D grid of 1 million cells can easily take 3 or 4 days to converge... averaging about 1000 steps per day? That's not the slowest I know so far...there are some slower ones...

Also, besides fluent and starCD maybe check this out? www.metacomptech.com
I remember hearing that this package is somewhat better than fluent..but I don't really know that much to conclude anything...

Cheers,
Ned

[Colin]

CFD++ (from Metacomp) is a development of a former internal-Rockwell code before they were bought out by Boeing. It has an Aerospace background in its development, unlike Fluent which was originally written to predict heat transfer in low-speed flows (i.e. pumps & pipe-flow). Consequently, CFD++ is better at calculating drag forces which make it more commonly used in Aerospace than Fluent. It is also a fair bit faster, and less expensive than Fluent. Fluent's forte is having a large development team which is continually expanding the code to offer a "full-service" package. CFD++ focuses more on excelling with a core set of capabilities.

And how long it takes something to converge to a solution is entirely dependant on how many processors you can throw at it. At work there are often jobs run on 50+ processors in a Beowulf cluster. There is a limit since it doesn't scale linearly with # of processors. Some jobs would only be marginally faster with 30 processors than 15. Larger numbers of blocks work better with larger number of processors. But 15 is damn better than 1 or 2.

[Ned]

Hi Colin,

Ah i see. I don't know about the backfground of the code. Do you know what behind the mechanism / numerical methods makes fluent better at predicting low speed heat transfers?

And if CFD++ is faster and better in prediciting drag forces... why is it seems that it's not popular in motorsports?

Btw do you mind telling me where did you know those things?

Thx a lot,
Ned

[MrAerodynamicist]

Well memory is the real killer for running CFD. If you don't have enough it won't run full stop 9or you cut down your grid size and sacrifice the quality of results) . But you can always substitute patience a decent cpu. Ah, the joys of residual watching.

As for grid generation, if you're sticking to fluent software then Gambit's not bad. It's what I've always used for 2D, and I quite like its interface/working methodology. However, I've had problems when scaling that to 3D (getting overwhelmed on complex geometry ie poor productivity, plus frequent crashes.), so for 3D I've used ICEMCFD. But then that'd be a whole bunch of more money!

[Colin]

Originally posted by Ned
Hi Colin,

Ah i see. I don't know about the backfground of the code. Do you know what behind the mechanism / numerical methods makes fluent better at predicting low speed heat transfers?

And if CFD++ is faster and better in prediciting drag forces... why is it seems that it's not popular in motorsports?

Btw do you mind telling me where did you know those things?

Thx a lot,
Ned

I use Fluent as well as a company-developed code and a different part of my company uses CFD++. We have compared solutions. You can check my profile to see where I work. I don't
know much about CFD++ besides what I have heard from the other users, so I don't know why
it is faster or better at predicting drag forces. Fluent also is a marketing machine. I don't think
it's the best code out there for motorsports-related activities, but there are a couple things in
its favor:

1) They spend a lot of money to be able to brag that most teams use their code.

2) This has developed a pool of motorsports engineers that know Fluent, so I'm sure teams are less willing to start over with a new code for they cannot find trained people. Incorrect CFD results could set a team back for months.

[Christiaan]

There is a thread I was looking for on a CFD conference held in the Netherlands by a represantative from Renault. The paper said that Renualt does a complete CFD computation in 24hrs using a million cell grid.

On what processes the speed my CFD lecturer told me about a friend of his who monitored the demands on a PC performing a complex computation. According to him though a faster processor was good, more important was how fast the RAM could be read, written and cleared. At my university they use CFX5 and they prefer to invest in RAM than processors. With CFX5 it seems to have made quite a difference.

[kilcoo316]

A million element mesh in 24hrs!!!! Thats fecking amazing, Im working on turbulence modelling in CFD at the moment, my 3D grids will be the far side of 3 million elements (to get Y+ right), and it would take upwards of 3000 hours to do on one computer, so the renault team must be using around 40-50 machines simultaneously or a couple of supercomputers or something. Thank god Im getting time on a super computer, or I would never get this damn thing done

[MrAerodynamicist]

Can't remember which team is was, but I've heard the figure 20 million cells overnight. Even if you factor in a Sales Speak bullshit factor.... I've also seen a pretty impressive demonstration by Advantage CFD (a consultancy who are linked with BAR) in which Villeneuves BAR is driving down the track while just about every flow visualisation method is demonstrated.

[kilcoo316]

There is simply no way a team could run 20 million cells overnight, for example 18.5 million cells for an Ahmed body simulation at Karlsruhe uni computing centre in germany using 2*64 IBM SP-SMP (power 3) processors took around 30,000 CPU hours, thats 10 days on 128 processors!!!! I seriously doubt any F1 teams has 20 supercomputers like these. Flow visualisations may look nice, but mean f all at the end of the day, numbers are what counts, and numbers require alot of time and CPU effort.