Hey there,

lucky me got a job as research associate at my university!

They buy me a computer and I get to configure it. I wanted to ask you about your opinion which components are best for cfd computations and generally numeric calculations (most likely large coupled equation systems) besides the usual office stuff. The OS will definitely be Linux. Price and (to me) heat dissipation (noise) is an issue. What would you use?

Greeting,
Jothan.

If you are on a budget, AMD is normally better then intel for price/performance.

Which CPU, and what else I would use would depend on your budget.

"They buy me a computer"

NSA I think or the US Air Force bought a bunch of game computers and hooked it up in a cluster that provided what was maybe the cheapest supercomputer.

If you are doing any kind of real computation, nothing you can buy at the local store would do. Start at the IBM or HP enterprise site for real systems.

Thanks,

I'm sorry, I think I should've given more details. My knowledge of hardware is about the age of intel core2 CPUs...

I fancied getting a Sandy Bridge i7 because I'm quite happy with my core2 on my private computer, both computational and thermal performance (I don't like noisy computers), but a friend of mine pointed out that this might be overkill and an i5 would do. For GPU I looked at GTX 560, mostly because I'd like to try out OpenCL and hear people complaining over the proprietary ATI-drivers. For nvidia there are nouveau drivers, which enable direct rendering and the proprietary drivers, the debian support of which I like (dkms).

Of course the real computations will be done on a machine built for it, but a lot of people here develop their own code which of course runs on their own machines first. What do you think of AMD CPUs? Are they still that hot (in celsius-terms) these days?

Greetings,
Jothan

FOR CFD you need a lot of power, so an i7 won't be overkill I think.

How hot a CPU gets is dependent on your cooling system. My Phenom II X6 1055T is almost never going above 40° Celsius, cooled by a Scythe Mugen 2 in a Coolermaster HAF920 case.

That's obvious, but certain CPUs tend to dissipate more power and thus generate more heat. Fitting the Scythe Mugen 2 on a mobile CPU will certainly result in lower temperatures than if it would be mounted on a quadcore i7. That's what I meant.

I read a lot more yesterday and am a bit turned down by the integrated GPU in the Sandy Bridge. Not only seems the Linux support to be bad, the CPUs with integrated GPU cost as much as the older models without GPU. Certainly the Intel Cores have a lot of features (extended whatever states and so on) but I ask myself if I really need this. Phenom IIs look a lot more attractive today, their price adds to this.

Greetings,
Jothan.

A cpu is rated either or both in watts and TDH total dissipated heat. The brand is unlikely to differ much in a comparable model.

If you want a better performance per watt then get a sparc system.

Believe it or not, but the top end i7 models are actually rated at higher TDP than the top end phenom IIs. "Sandy Bridge" linux support is fine, apart from the intergrated video (I wouldnt call it a 'GPU'). Which is something you wouldnt be using with an nVidia or AMD/ATI card.

If an i7 is overkill, a GTX560 would be as well. The GTX2XX, 4XX and 5XX cards are pretty power hungry, and in general not quiet at all....

I'd probably go for a Phenom II X6, even with a nice aftermarket CPU cooler it would work out cheaper than the i5s. Video card....hmm...difficult, and would depend on details I dont know. Possibly a GT430/440, they are cheap, fairly quiet, dont need huge wattage power supplies and will run CUDE/Open CL.

It was the air force, and PS3s-

http://www.physorg.com/news/2010-12-...rcomputer.html

The air force is using (or at least was using) off the shelf games consoles for some pretty neat number crunching.

I know you are a fan of server hardware, but you dont need server hardware for 'real' computation.

I can't imagine what sort of a University would do with cfd computations that were not from some very high end system.

Are we talking about the same type of work??

http://en.wikipedia.org/wiki/Computa...fluid_dynamics

I'm a bit surprised that you aren't passing the real serious compute work to some system other than your desktop, and getting it to do the heavy-duty number crunching. However, assuming that you are going to use your desktop for the number crunching...

Firstly, determine whether you can do any of the number crunching on your GPU. If you can do that, let that determine what video card you use (what is supported, what do others have success with, using the software that you want to use).

If not,, you will be doing all of your number crunching on your general purpose CPU. Fortunately, you main problem is highly paralisable, so it seems that the 'more cores' option should work well for you (for some problems, the more cores option isn't very good, as the task itself isn't that easy to parallelise).

In that case, a Phenom II X6 ought to do a decent job, without costing a fortune. Intel will have higher performance options, but, by the time that you have beaten the AMD option by a worthwhile margin, it could cost significantly more. But, if you have the budget to spend...

(I'd still imagine the option of a 'departmental compute server' with a multi-socket arrangement, thus offloading several desktops from real compute duties would be better, but it may be an organisational issue as to whether that would work within the organisation that you have.)

Actually, this machine seems only to be for testing the code. So I would recommend to built this machine as a small version of the number-crunching machine. It wouldn't make sense to write your code for parts that are not used in the number-cruncher.

+1 to TobiSGD above.

Depends on the department, the student/lecturer level of the creator and what 'future' there is in the software.

A post grad student who is making CFD software that might have commercial or educational impact will get a lot more access to resources than an undergrad making CFD software 'just to see how hard it is'.

BTW, not with CFD AFAIK but there are some serious projects out there running on very average systems. The best example is folding....

http://folding.stanford.edu/