Hi everyone, I've seen that there are quite a few CFD codes which will run on normal PC's. But what type of system would be more appropiate? For example, is it better the Intel architecture or the AMD? Is it important the speed of the RAM, or just the amount, any suggestions?

Thanks for your help

Fernando,

This has been debated on cfd-online before, and you can find that by searching the archives.

I'll try to summarize, and will surely be corrected by the forum when I get it wrong.

I guess first you split it into "workstation" class - PCs because they have Intel or AMD chips - but (probably) using Linux. They may also have other hardware (high-speed bus, specialized graphics, two or more CPU's for parallel processing, ...), and they will be very expensive. The consensus seemed to be that for, really big problems, the extra investment could pay off.

For any PC, I read the previous discussions to say that the bus speed is important. And the memory speed should be matched to the bus speed. Also, the amount of memory matters a great deal; I think the rule of thumb was 'as much as you can afford and the system design can handle.' You'll want a big and fast hard drive. For backup, a CD/RW (or even a DVD/R) will really help. I read the previous posts to say that 'your computation speed will not double if you double your processor (CPU) speed, but buy the fastest processor you can afford anyway.' Your graphics card and monitor will need to able to handle your output displays.

Thanks Jim,

I saw some previous posts about it, but as the processors have changed, and also there are so many different type of RAM memory available, I thought that maybe one of them is more suitable for this task.

I'm interested in a single CPU, maybe two processors, but no more.

Thanks again,

Fernando

If it is your only machine I would recommend a dual cpu, so that you may mesh/work while running a solution. You could also run parallel while away at night should your code/license support that. I think the general consensus on memory is at least 100 mb per 100k nodes... I would get the fastest dual cpu with an adequate amount of RAM for your largest model and enough to be able to work on other things..

How much cells do you have ? How much licenses do you have ?

Intel P4 1.7GHz with linux, RDRAM (at least 1GB, that's enough for 1000000 cells (fluent), with the possibility for more ram), HD as much as possible. Think about a storage-possibility: dat or cd ? (-->scsi or not ?) graphics: good geforce2-card is enough.

advantage: cheap. disadvantage: only one cpu, that means: it's difficult to work while a job is running. But if you only have one license --->doesn't matter.

Thanks John, that's a very good answer (I hope ).

Just one thing, do you suggest P4 for anything? I mean, does it perform better this task than a new Athlon XP processor?

Thanks

I have posted benchmarks of P4, PIII and Athlon machines here before - just search the archives (use the site-wide search function it the upper right corner).

I'd recommend you to get a P4 with PC800 RDRAM memory - this configurations runs CFD applications very well, much better than more "general purpose" benchmarks indicate.

How much memory you want depends on how large cases you will run and which code you uses.

If you can afford a dual-CPU machine that is nice - you can mesh/post-process and solve at the same time without any problems. Dell sells a very nice dual-CPU machine - Precision 530.

If budget is critical a single CPU machine is a better choice though - but make sure you get PC800 RDRAM (not PC600!).

Different CFD codes seem to favour different makes of CPU. I've found that CFDRC codes (Fastran and Ace) work extremely well on Athlon CPU's. However, I've been told that Star-CD works significantly better on a Pentium 4. Bottom line is probably that you can't go wrong with either, they are both really amazingly good value for meoney.

Other advice is sound, as much RAM as you can afford (for the type of work these CPU's are capable of, anything less than 1 GB is probably not justified), at least two BIG (40GB+, CFD data files are huge) hard drives, CDR/W, decent graphics card. For a P4 you need PC800 RDRAM, for an Athlon PC2100 DDR-SDRAM, CL 2 (latency) if possible. There are many motherboards out there, for an AMD Athlon system the best board would be one with 4 DIMM slots and the VIA KT266A chipset (the "A" is important, but I believe that they have now changed there nomenclature slightly. Just check!)

Dual CPU systems are a realistic and attractive option for both P4 (Xeon) and Athlon (MP).

The new Athlon XP processors are identical to the MP processors, and much much cheaper. And, the cost of the Athlon XP chip is much cheaper than a P4 chip of similar ratings. You should probably check out the website anandtech.com which I find gives a good discussion of hardware issues out there, including price, performance, etc. Do a bit of other research before investing in a given system.

I've been looking to the prices, and RDRAM memory is just twice as expensive as DDR. As I don't have too much to spend, I guess it'd be better to get double amount of memory (DDR). Do you think this is right?

I don't know yet what software I'll be using, I will in Jannuary though. I will be doing tests for yacht design only (sorry I didn't mention it before).

Thanks to everyone for your advice.

If you deal with 3D problems, you'll never have enough memory. From my point of view, if you don't have the extra money, go for the latest AMD athlon with DDR memory. If you give a close look to the benchmarks posted on: http://www.spec.org/osg/cpu2000/results/cfp2000.html you will notice that the latest processors produced by AMD give results that are not to far away from P4 results. These are not synthetic bencharks that can be ran in cache. But they reflect real life applications.

You should at least buy one GIG of RAM (this is the minimum required). But, as it was said (and depending on your software), 2GIG is more realistic.

As I know, all the floating specs are averages of a lot of single tests. In cfd you do not have a wide range of applications, but only the solving of large matrices. If you look in this forum for different benchmarks, you find, AMD sometime looks very bad. So in my opinion, intel P4 with rdram is more sure to work fast than amd. Benchmarks like povray or an overall spec (int and float) are not very useful..

The best would be a benchmark of meshes with different sizes with different cfd-codes and different cpu.

Athlon XP: I do not know much about it. Did someone benchmark it with cfd-codes ? Typical benchmarks like specs or povray or quake do not say much about the speed with cfd-codes.

For this site (www.spec.org), they use actual "real life" cfd codes. You can rely on the results posted their. For CFD, I wouldn't pay much attention to povray or quake or any of these benchmarks, they are aimed for gamers and doesn't realy reflect the Needs of cfd applications.

Regards

Would you kindly point out where in www.spec.org one can find _direct_ CFD-related specs? I browsed through the site quickly and didn't see anything of the sort.

Thanks in advance

Adrin Gharakhani

www.spec.org/osg/cpu2000/CFP2000/178.galgel/docs/178.galgel.html

www.spec.org/osg/cpu2000/CFP2000/

These floating points benchmarks:

http://www.spec.org/osg/cpu2000/results/cfp2000.html

The results are not in flops (such as linpack1000). But reflects "scores" obtained running these applications:

http://www.spec.org/osg/cpu2000/CFP2000/

Regards.

Thanks to everyone.

I don't know yet which system I'm buying, because I'm waiting for a confirmation about how much I can spend. I'm also waiting for an answer from the software company, because there are so many options: now you can even have P4 with DDR for example!

Hopefully everything is clear by Jannuary, then I'll let you know what the choice was and how it's performing.

Thanks again.

Fernando