I was wondering what setup would be good for a home CFD machine? I have found a dual Phenom (AMD quad core) motherboard that holds 32 GB of Ram.
Are the CFD programs processor dependent, or are they video card dependent? With the Quad Crossfire/SLI now available, would that be the better way to go?
I am pretty interested in this, as I have an $1800 budget to build a HTPC for my living room. I was gonna save a ton of money by just getting another dual-core machine, but maybe I'll go a bit further.
1) You'll need a decent graphics card to modify those complex surfaces to run in CFD...
2) You can always run the CFD while You sleep (which is definitely processing speed driven).
3) It will be easier (cheaper) to add significant processing power in the future (serverclusters/ computer farming/ etc) than to add graphics cards (motherboard compatibility/ $/ CAD program compatibility)
Figuring out what programs You want to use (CAD & CFD) will help guide You in a hardware direction. You may find $1800 is a little overkill for freeware... You may also find some legal personal (non-business) use programs for grocery money.
There is actually very good free CFD software out there, called OpenFOAM ( http://www.opencfd.co.uk/openfoam/ ). It is not very straightforward to figure out and it's really Linux only, but I use it and I am very satisfied with it. The main problem is actually that there is little good free pre-processors available, namely programs that can be used to generate meshes. OpenFOAM does support the most common commercial programs however.
Your question is too broad. What are you going to be using CFD for?
I can't think of any CFD program that has higher graphics hardware requirements than most CAD packages.
OpenFoam is anything but straightforward to use. Traditional CFD should be left to the CFD analysts. And if you are a CFD analyst, you probably need a much nicer machine than you can purchase with $1800.
slimjim8201 wrote:Traditional CFD should be left to the CFD analysts. And if you are a CFD analyst, you probably need a much nicer machine than you can purchase with $1800.
Why such a defensive attitude? You are of course not going to be doing cutting edge CFD with any desktop machine, but a $1800 machine today will be worth roughly a $10k cluster from 4 years ago in terms of processing power. In order to learn CFD and actually already perform some useful calculations, a $1800 machine is going to be more than enough. It is of course not going to allow you to process a whole F1 car with any accuracy, but a quad core with that sort of RAM will certainly lead to proper analysis of many simpler systems if so desired.
And to answer the original question, a CFD program will be CPU and RAM dependent mostly. RAM will limit the detail of the systems you can study, and CPU will determine the time needed for the computation. Any good graphics card will likely suffice for pre- and post processing.
Dual (2) Quad core processors, AMD or Intel (Probably AMD for cost)
8 GB of ECC Ram
Now, with the crossfire on the ATI cards, you can actually use 1 card for video output, but then link up to 4 cards total, using 3 of tehm for physics processing.
My question is is it better to go Dual Quad core, with 1 graphics card or is it better to go single quad core with 4 graphics cards, 3 devoted to physics processing?
And about Linux, I am going to use Vista home Premium on the box for the media center as well as the free dish TV with the skywalker-1 interface. I also am thinking of putting a stand alone hard drive in for Ubuntu if necessary.
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. I didnt really expect that coming from this forum.
I just want something that I can play around with on my own, and maybe devote my sleeping hours to running some of the F1Technical communities projects.
If the general consensus is for my n00bness to stay away from CFD, then I wont waste my time, effort, money or interest into this project.
Why should you let someone's opinion influence your final decision? If you want to try CFD, try CFD. Suggestion to leave CFD for pros sounds like a sign of fear. Even if there was a consensus on this forum saying NO, why would that stop you from doing what you like? It's your money and your decision.
Gecko, Conceptual, not trying to be defensive or to put anyone off from learning CFD. Far from it. I just wanted to get more information about what the computer was going to be used for. Valves, pumps, electronics, or full-scale F1 simulations.
As far as machine specs, Gecko is correct in that the size of the simulation model that can be handled is dependent upon RAM, while simulation speed is dependent upon CPU speed.
I would avoid the current quad-core processors for computationally intensive simulation programs like CFD. In our testing, we have found that the front side bus bandwidth will be exceeded before the processors are fully cooking. A four core unit with two cores solving may produce a ~70% improvement in speed. Enabling a third core may result in 75%. A fourth, 76%. Until the FSB technology is dramatically improved (or eliminated altogether) these quad core processors can't be fully utilized.
I believe both Intel and AMD are working on new processor/memory pipeline technologies (point to point interconnect) that will replace the current, out dated FSB technology.
I'd recommend buying a faster dual core processor over a slower quad core unit. Or if it's in your budget, two dual core units will provide better performance than one quad core unit.
8GB of RAM will allow you to run a simulation model with 10-15 million elements or 3-5 million nodes (depending on the type of mesh you are working with).
I wouldn'nt be overwhelmed about learning CFD but I would be if it involved learning how to use an independent mesher, what all of the solvers mean and their capabilities, pros, and cons, how to interpret results, etc.
Telling people that they should back away from CFD is not what I intended when I said that traditional CFD should be left to the analysts. The company I work for was founded on the belief that there is a better way for mechanical design engineers to use CFD in their design process without being CFD experts. There is no reason why CFD has to be difficult. Sure, there's a lot of stuff going on behind the scenes and a basic fluids course only scratches the surface of the physics involved, but frankly, MEs don't care about partial differential equations and "this solver works better for this application" sort of stuff. They just want to know if their design changes are improving performance without testing them manually.
If you are just looking to run random projects here and there for your interests, I think your only really feasible option is to download something like OpenFoam since the costs to get a decent all in one CFD package would be considerably more than your machine alone. There are better ways to learn CFD, but unfortunately, they aren't free.
Manchild - Better to ask as to the overall feasibility than to get $1500 into it and find out that it is still gonna cost $5000 more to make it work.
SlimJim - As per the AMD Tech Tour of 2007 in New Jersey, the AMD chip engineer that I spoke to said that they had eliminated the traditional FSB architecture with the first gen x64 chips. He also went on to explain that the latest dual cores and the quad cores, as well as the upcoming Octal Cores!!! all use some sort of dedicated interface with load balancing. What it really comes down to in the AMD system is the RAM speed. Remember an Intel Gigahertz is not measured the same as an AMD Gigahertz, and the FSB design is directly the cause of that.
Back to my question of the CrossFire video cards. They clearly state that in a 4x setup (4 cards) that you can devote 1 to video output, and the other 3 to really complex physics processing. These cards are 128bit and are all sbout modelling 3D objects. I find it hard to believe that the CFD software isnt written to specifically take advantage of that processing power. Especially with 512Meg of onboard DDR3 RAM per card, that is the UberFast FSB workaround that you mentioned SlimJim, and I'm surprised that your cutting edge company hasn't looked into it.
slimjim8201 wrote:Until the FSB technology is dramatically improved (or eliminated altogether) these quad core processors can't be fully utilized.
I believe both Intel and AMD are working on new processor/memory pipeline technologies (point to point interconnect) that will replace the current, out dated FSB technology.
AMD hasn't been using an FSB since 2003 (April for Opteron server chips, Semptember for desktop Athlon 64s).
Intel's variant is for the end of this year/early 2009 IIRC. (and their current chips have upto 12.8GB/sec of FSB bandwidth)
Conceptual wrote:Manchild - Better to ask as to the overall feasibility than to get $1500 into it and find out that it is still gonna cost $5000 more to make it work.
Indeed, that's why the reply should have stated possible costs rather than lack of professionalism as an obstacle.
For $1350, I think that this will do an aweful lot of CFD work... What do you guys think?
Chris
EDIT:
I agree Manchild, but I dont want to start any beef over this. Especially because I need SlimJim to help me setup a Web Based interface for a command line CFD program so people can upload their models, and have them processed via the web.
After his comments, I'm sure that he will be more than willing to help!