We have been using top-end PC's for our CFD work over the last two years, currently using a couple of dual AMD Athons and a 12 CPU cluster. I am quite happy that I understand the performance issues of our hardware very well. However, there is always a need for more performance, and my question then is: What is the best option for a sub $40 000 CFD computer, and how fast can one expect such a system to be on real world CFD applications, when compared to the X86 "PC-based" systems.

Actually in our company we have moved away from the "sub US$ 40000" Workstation and use now a simple Dual NT PC we got for "sub US$3000". We have had a test with our old workstations and the PC was a little bit faster. I am sure the newest Workstation will be faster than your Athlon or my PC, but you have to really think about the price/performance ratio. Now of course the real Unix afficionados will go into detailed discussion about memory management and bus speeds, but I can assure you, with your equipment you will be very dissapointed of the benefit from your investment. If you want a real speed up of ten times or more you have to invest an order of magnitude more.

I think this benchmark calculations are sometimes overrated. In real world you usually don't stare into the monitor with a stopwatch in your hand until the results are ready, because there are a whole lot of other things to do (like writing reports to the management).

With $40000 you can buy over the time several Dual PCs of the newest generation and before the Workstation is halfly depreciated they have certainly become faster.

Same with us. More and more machines are Win2000 or Linux dual processor boxes. These boxes can have multiple roles, e.g. used for office activities some days and as calculator overnight and/or at the WE. And we can buy quite a few for the same money!

This is pretty much in line with the conclusion that I had already arived at. Not having had the opportunity to benchmark a 64-bit workstation on a CFD problem, I just wanted to make sure that I wasn't missing something. Thanks guys.

Have you compared Athlon performance vs Pentium? About a year ago we did a test and it was pretty bad, that is to get equivalent number crunching for CFD you would need Athlon processor rougly 1.5-2 times faster than Pentium, i.e. e.g. 1000 MHz Athlon vs 600 MHz Pentium.

Yes, we've been down that road in quite a lot of detail. The results are summarised in a couple of graphs at http://home.mweb.co.za/cc/ccrosby/cf...enchmarks.html

The conclusions I've come to are: 1. A DDR-memory equipped Athlon is very competitive with even a Rambus equipped Pentium 4. 2. Cache size is important. The new 2.4 GHz P4's with 512 kB of L2 cache are probably the best proposition at the moment. 3. If the code is compiled to make proper use of SSE-2 instructions (unique to the P4) the P4 does much better relative to the Athlon.

I think your results are atypical - it is certainly not in line with any benchmarks I've either seen or performed myself.

We had a strange case. It involves the use of commercial CFD packages: The cost of a CFD license is often related to the number of processors you have. So using a few PC's instead of a single fast UNIX animal drives your license costs up. As we are using commecrial CFD software it is much cheaper to: 1. lease your UNIX equipment 2. get a fast multiprocessor unix workstation 3. Have them run 24 hours a day (UNIX will do that)

If you have your own CFD code (parallel) without any restrictions by licenses, a cluster of linux PC's will always be cheaper (and faster for the same amount of money)

Charles,

In the benchmark graphs it is only said the CPU. What about the other parameters (motherboard and memory) ?

Thanks,

Daniel.

Very odd. This is not in line with any benchmarks that I've seen. Additionally, with CFX-5 and two of my own codes the Athlon was at least 30% faster (at the same clock speed). I think you sacrifice some stability with the Athlon/VIA setups though.

Neale.

Well, it's been my experience that you should select the CPU and chipset for performance, and the brand of motherboard for its features (such as number of slots, etc.) In the tests we did the Pentium 4 had an i850 chipset with PC800 RDRAM and the Athlon had an AMD 760 chipset with PC2100 DDR SDRAM. Interestingly, shortly before those tests, we also benchmarked a 1.2 GHZ Athlon with PC2100 DDR SDRAM on a different motherboard. It was at a computer show, and we didn't have time to poke around on the motherboard, but I'm pretty sure it used the Ali Magik chipset. The performance was very disappointing, in fact, it was WORSE than our old 1.1 GHz Athlon with PC133 SDRAM! Since then of course, the VIA KT266A motherboard has been shown to outperform the AMD760. I've concluded that the performance of the memory controller is really VERY important in determining performance when running CFD applications, and it I think it's very promising that the AMD 64-bit CPU's (Opteron, formerly Hammer) will incorporate the memory controller on the CPU chip. It's likely to give a significant boost in performance. I gather that the Opteron will also support SSE2 instructions, which is another very important consideration. At the moment, if your software has been compiled for SSE-2 you should go for the P4, if not, the Athlon.