I am a new bee to the CFD world. Why should we use 64 bit computing resources for CFD analysis.

Because then you can access more than 2GB of RAM and deal with files greater than 2GB in size.

This may not seem a problem, because most CFD codes come with parallel capabilities that allow models to be spread over many CPUs (each chunk within the 2GB limit).

However, when it's time to postprocess your model, you're always going to need to use a computer than can address the entire model.

In practice, you often see (32-bit) Linux clusters used to run the solver in parallel, with (hugwely expensive) high-end 64-bit workstations (e.g. J-class HPs) to postprocess.

With AMD64 processors (Opteron and Athlon 64) there is no conflict with 32-bit versus 64-bit, since they can run both 32-bit and 64-bit x86 code at full speed and performance without any emulation (Itanium can only run 32-bit code in emulation and is thus *VERY* slow).

AMD64 dosen't just offer larger adress space with regards to bigger memory, it also means bigger registers that are important for making efficient applications. Read more about this great 64-bit architecture here: http://www.devx.com/amd/Article/16101

Hey, this isn't an advertising board!

If you want "great 64-bit architecture", you could also try computers from all of the regular UNIX vendors: for example HP, SGI, Sun and IBM. In the UNIX world, 32-bit is old hat and has been for some time.

Hehe, this is just my humble opinion and I've tried systems from most major vendors so I'd say I know what I'm talking about..

But I'd argue against your claim that HP and SGI offer "great 64-bit architectures" anymore, since their systems are now IA64 based which is as far from a great 64-bit architecture you can get in my opinion (inefficient and extremely expensive). Sun and IBM offers far better 64-bit solutions based on SPARC and Power4, and recently also systems based on AMD64.

Anyways, I think AMD64 is a great 64-bit architecture and without question the best performing solution for PCs according to numerous recent reviews.

The Alpha chips from HP (DEC/Compaq/whatever) ain't Intel-based. Let's just hope market forces (e.g. huge US government contracts) keep them alive. It would be such a shame if these chips are allowed to die.

In CFD, there are other factors to consider besides raw computing power.

A few years ago, we upgraded to one of the first 1 GHz AMD Athalon computers for running a commercial code (FLUENT). The AMD was faster than our old 300 MHz Intel system for small test cases, but for larger real cases, it wasn't any faster.

Emails to AMD went unanswered. We worked on the issue with our code vendor off and on for a month or so with no solution. They offered to recompile the code with a different compiler - reported to give better results - but we already had three Fortran compilers and didn't want to move to a fourth.

Then there was the issue (as AMD recognized) that after doing some graphic intensive operations, with some graphic cards, there could be problems with the main memory storage. So if I stopped a run in the middle to look at the results, then the solution would start to diverge, whereas it would converge if you didn't do any viewing in the middle.

After wasting lots of time doing benchmarks and testing, we cut our losses and bought an Intel-based machine for CFD and relegated the AMD to the office scanner.

Intel has certainly had its problems - remember when the Pentiums came out and couldn't add? And the security issues with the chip ID? But, there are also advantages to sticking with proven technologies.

Anyway, when you buy hardware for CFD, consider what it is you want to do. Develop and run CFD programs or fool with hardware and compilers.

Hi,

My company is currently looking into buying new hardware to run large mutliphase simulations using FLUENT. What do you suggest would be the best option at the moment. Some people are suggesting the AMD Opteron and others have suggested going for a Dell Precision 530 Workstation with Dual Xeon processors. What are your thoughts?

Yes, I agree that Alphas are nice 64-bit processors but it seems like HP have killed off the them too according to this story.

Soon HP will only provide IA64 for their 64-bit systems, a pretty bad 64-bit solution if you ask me with its poor performance and extremely high price.

I'd definately go for a dual Opteron system, as Xeon is yesterdays technology already. Fluent will also release AMD64 optimized versions of its software early next year, even though the current 32-bit version performs very well on Opteron.
In the November issue of Cadalyst Magazine, nine dual powered workstations were reviewed (see here) and the dual Opteron systems came out on top (also read this article for more background information).
"Xeon is hemorrhaging performance because its northbridge-based architecture is out of date. Opteron's ability to scale has made today's Xeon yesterday's technology."

If you look at former discussions about this topic you will find that is is much better to build the cluster using single processor machines. A P4 with 3.2GHz and 400 MHz Bus is fast, cheap and known to work.

The AMD stuff might be good but anybody should proof it first. If a 64bit version is announced, it is still vapourware. Once it is released and approved by the users we can go on talking ...

Allan, when you compared the 1Ghz AMD with the 300Mhz Pentium on large problems, could it be that the computations were hindered by I/O since your RAM may have been too small? When I do CFD on a PC, I try to install as much RAM as possible since I want to put the whole grid in RAM, otherwise the computation time may increase by a factor of 5. I must mention that I was using W2000 Pro which did not allow my CFD case to be run in parallel on other machines.