In the near future, I will have to buy a new PC for running CFD programs (FLUENT). Typical problems will require about 500k-1M+ cells with turbulence and chemical reactions. In most cases only one license and one machine will be used for running.

Could somebody recommend a viable PC that could handle this kind of jobs? I am not looking for top of the line or most exepensive machine, rather something that has a good ratio of performance/price, assuming that probably in one year I will be upgrading and buying something better anyway.

What would be the important features of "ideal" CFD-PC?

- lots of fast memory, 2GB or more should probably be enough

- fast and big HDD, maybe 200GB SATA/PATA?

- fastest possible MB/CPU? intel/ADM?, dual cpu, dual core, front bus speed, hypherthreading, cache...? here I am confused and not sure what is of first importance for running CFD jobs

- Graphics card? something like ATI Radeon 9600XT should be good enough?

- display? probably something like 21" or Dell 24" LCD with enough of pixels "real estate" for concurrent working on geometry/meshing, running, previewing etc; or maybe two smaller displays (17"-19") side by side with dual graphics card?

- brandname? Dell, HP?

any comments, recommendations, own experience would be appreciated

thanks

Paul

Ask the Fluent folks. They should have some experience, and they may be able to put you in touch with other customers who have already answered your question.

well, as I mentioned in another thread, you MUST have a nice video card to display Fluent graphics or you will run into problems. - 256k.

tomshardware.com will explain to you all about motherboards, cpus, videocards, RAM, etc etc.

I have had good experiences with PCUSA.com, Tigerdirect.com and Monarchcomputers.com. I have used Giga-byte, Asus, and MSI mobo's with my favorite being MSI because of the extras the boards come with.

Also, pricewatch.com is a good place to check on prices.

At home I use a Dell XPS gen 5 to run star-cd. It has a 3.0 GHz dual core pentium chip, 3 gig memory, 800 MHz front side bus, nvida video card and a 24" Dell LCD monitor and dual 200 gig SATA HDD. My idea was to get a gamer's machine and adapt it, and it works well. Gamers want fast machines, so I read gamer's magaizines (hardware type only) and went with the fastest of everything I could find. My OS is XP/Pro. The problem with that is that it is a 32 bit system, so will only use 3 gig memory. I have heard that the 64 bit XP/Pro doesn't have all the drivers for gaming machines available.

By the way, you do not need dual video cards to run dual monitors.

Tom

If you are often going to be doing running and meshing at the same time, or using parallel processing part of the time, the most cost-effective solution will be a dual-core CPU. If it will mostly be doing just runs on a single serial (no parallel) license, the best option would be the fastest single core CPU. The performance of the dual-core CPU's seems to scale more or less with clock speed on a given platform, so a dual-core 3.0 GHz P4 runs a single process job at about 80% of the speed of a single core 3.6 GHz chip. But of course you have the option of running parallel with the dual core. Although the widely published games benchmarks show a large performance advantage for AMD over Intel, the difference in Fluent is smaller, but it is still there, and worth having. The SPECfpRATE benchmark for the dual-core Intel CPU's (such as the 830D) suggest that a single 830D chip is as good as two single core 3.6 GHz Xeon CPU's. This is not true in Fluent, where the 3.6 GHz Xeons lead by almost exactly their clock speed margin.

I would suggest getting 4GB of RAM, but in 32-bit Windows mode, your Fluent jobs will be limited to 2GB, unless you can use parallel. It sounds like your jobs could get close to the 2GB limit. If you are comfortable with Linux, it makes sense to go with 64-bit Linux and Fluent. But in that case get an NVidia graphics card, their Linux driver support is very good. Might be worth asking Fluent when a 64-bit Windows version is due out.

200 GB SATA drive sounds fine, but consider getting two for backup purposes.

A half decent 3d games card is likely to be plenty good enough, unless you are aiming to do some very sophisticated graphics post-processing.

On monitors, you really need at least 1600 X 1200 resolution.

Jim_Park:

Asking Fluent folks might work, and it might not. In my case they did not know, when I asked about a year ago. What we did was to contact two providers and ask for a test computer. We then benchmarked the two computers against each other (using Fluent). Although my personal preference is AMD we found out that the difference in Fluent between AMD and Intel is not all that big (as mentioned previously in this thread).

Paul:

If you have the time to build your own system I would opt for that. Workstations usually have a very high cost associated with just the word "workstation". You can also buy a cheap system from any of the big companies (dell, Hp, etc,) and modify it yourself, since they usually tend to have really good prices, but lacking some critical component. They then hope that the customer will upgrade the critical component to a premium price. --> buy the critical component from a cheap external source and you have a cheap system

I mostly agree with Charles...

Except that if you are paying for your Fluent License then you really shouldn't skimp on CPU power to save a few $. The computer workstation costs a small portion of the total CFD cost (10-15%) so doubling up on workstation cost to get a 20% improvement in performance is actually cost effective.

Otherwise if you are not paying for your Fluent licenses then you really don't have an excuse not to spend more on your workstation

I would suggest a 64bit machine, I understand that solving chemistry is best done in double precision. A 32bit workstation is significantly slowed down when running in double precision mode where as a 64bit workstation is ideally suited to such a task, and suffers minimal slowdown over single precision.

If you are doing complex CAD to mesh operations then don't skimp on the Graphics. One possibility is a NVIDA Quadro FX 4000 256Mb. Take note of Charles' comment about Linux.

Dual core machines (2 Processors on one chip) are not suitable as they have a performance reduction over single core processors. This is due to the need to share resources such as memory access. They also suffer from increased heat production and so are only available at reduced clock speeds compared to single core chips. Remember that the most expensive part is the license and a dual core processor will require two licenses. The speedup when using a Dual Core processor vs. a Single Core processor is of the order of 180% and not 200%. See http://Fluent.com "Server Application Single-to-Dual Core Scaling 1-4P Servers"

Dual machines (2 CPUs on the motherboard) offer reduced maintenance (e.g. 1 computer vs. two) and with the AMD Opteron architecture there is little to no performance reduction, as they are not sharing resources as in the case of the Intel Xeon architecture. AMD Opteron 852 is 9-87% faster than the Itanium 2 1.6GHz and the AMD Opteron 248 is faster (15-20%) than the Xeon 3.2GHz 2MB L3 cache. See http://www.fluent.com/software/fluen...ch/fullres.htm for more details.

Memory is a difficult to spec, a large grid could be anything from 3 to 9 million cells in size. Requiring from 6 to 18 Gb of memory using the segregated solver and 12 to 32 Gb using the coupled implicit solver (values estimated from current use). But in this case you would most likely need a parallel license to get reasonable run times, so you can spread the memory requirement over more than one machine (unless you are using multi processor/multi core machines)

Note that a 32bit machine is limited to 4Gb of memory and memory bandwidth has a direct performance implication when dealing with large amounts of data such as is the case in any CFD solution. The AMD Opteron architecture has proven to have excellent memory bandwidth, better than what is found in the Intel Pentium/Xeon architecture. Also the Opteron does not suffer significant degradation in bandwidth when running two processes simultaneously. Due to the impact that memory bandwidth has on CFD system performance it is highly recommended that the fastest memory available be used (DDR400).

DDR memory is currently the most common type of memory available. DDR2 is a revision of DDR that will offer higher speeds in the future but currently has little to no performance advantage over regular DDR. There is also a trade-off between memory size and speed as the faster memory cards (400+) are not readily available in the larger sizes.

Get two HDD and run them in Raid. Well worth the investment. I still don't know if a really fast SCSI interface is worth the price.

Hyper-Threading = BAD, this will require two licenses to run on a single processor, until the licensing recognizes Hyper-Threading as a single CPU.

I have used a Fijitsu Siemens Celsius Workstation Running Red Hat with great success.

Note that this information comes from a recent survey that I did to try choose the best workstation for our CFD work. There may be errors and miscalculations, so do your own homework.

Gus makes an important point about relative costs of software and hardware. To a first approximation, if paying commercial license fees, the ratio of hardware:software:labour cost is something like 10%:30%:60% , so saving on hardware is false economy. I have done some benchmarking, and running 2 processes parallel in double precision produced the following times for a given number of iterations:

Dual CPU Xeon 3.6 (106 seconds), Dual CPU Opteron 252 (92 seconds), Dual core 3.0 GHz P4 830 D (113 seconds).

So the advantage is with the dual CPU Opteron. It is not a huge lead, but it is worth having. All systems used the same Fedora Core 4 64-bit Linux and 64-bit Fluent. The Opteron's lead over Xeon is about 15%, and not the 33% one would expect based on SPECfpRATE. My guess is that a version of the code exclusively optimised for Opteron would stretch its lead closer to 30%, but with a commercial solver that's not really an option.

Double precision slowdown in 64-bit mode is marginally better than in 32-bit mode, but the difference is suprisingly small (Speed ratio goes from 71% [32-bit mode] to 75% [64-bit mode] for Xeon. For Opteron it is around 79% in 64-bit mode). The slowdown is really just due to shifting more data to and from the memory. The Intel systems suffer from a bit of bandwidth contention, but there is actually a lot of bandwidth available.

About the benchmark; When we benchmarked Fluent we found that there was a performance relation that was time dependant when comparing intel and amd. This was some time ago and the tested cpus were a amd 3500+ and a 3.6 GHz intel. The amd system was faster initially but as the calculations leveled out the intel system gained on the amd system.. I have no idea why this occured. Anyway it would be interesting to see if the same happens with your systems.

Maybe a much faster disk system on the AMD, which gives a quicker start-up? Alternative would be an AMD system that reduced clock speed when it heated up, although that's more of an Intel trick.

Hi

Size of RAM does matter, so, minimum RAM ought to be 2GB, however, HDD has nothing to do with speed; that depends on how much work you want to do. Secondly, I'll suggest highest possible L2 cache since CFD is matrix based work and matrices need higher L2 cache for higher speeds. Another thing, its better to have dual core and run parallels, however, upto 1M grid density, single processor does a good enough job.

Now most important thing is your budget. Because the best are still Alpha Processors, much faster, much better. Then the OS you use.

It's never like if you use everything of top most ratings it will work faster, it might be slower than a normal PC. Most important thing is synchronization of components, that's why get a better company's system. Dell is good, HP or Compaq better than that.