Has anybody got any good suggestions regarding CFD-suitable workstations in the US$ 10,000 to 15,000 bracket? PC's are starting to get right up there now in terms of clock speed, but am I right in thinking that the Sun-SPARCs and DEC-Alphas (and similar) of this world are still superior, in terms of having much more cache memory, and in general having much superior memory <--> processor communication?

Hi,

my experience is that a at least 2 PC's (sometimes the factor is even only 1.5) equal the speed of one high-end workstation (with alpha-chip). So regarding the price-performance ratio PC's are a serious alternative as long as your code can run in parallel mode. (Getting the same performance for only half the price) If one want to stuck on workstations alpha-based machines seem to be the fastet.

Tim

I have done some recent benchmark tests if anyone is interested. Times are in seconds for the same model, using one CPU.

Intel P3 (400 MHz): 690 HP 9000/785/J5600 (500 MHz): 252 DEC AlphaServer DS20 (ev6, 500 MHz): 242

BTW on a different problem, SGI's latest offering (R12, 300MHz) came in at about half the speed of the HP and and Sun's (Ultra2, 300MHz) about a third the speed.

So it looks like Compaq (DEC) and HP have got their act together and you'd need a 1.5GHz Intel to catch up.

If your code is capable of running in parallel, I would strongly recommend building a PC cluster running LINUX (a so-called Beowulf cluster). I'm building one a present out of 8x800 MHz PC's. I expect an equivalent performance of around 5000MHz. This has cost just over 10000 UKP, so I'm getting super computer perfomance at a fraction of the cost (plus a load of distributed memory so I can run really large meshes). There's lots of info on Beowulfs on the net, start at the site of the original Beowulf at http://www.beowulf.org/

(1). Based on my experience, there are always different models available for the same brand of the computer. (2). The trend has been using one brand of computer in one large company, this will eliminate a lot of commercial software compatibility problem (different versions for different brand of computers). (3). It seems to me that the difference between two computers from two different companies is not the main issue, unless you are pushing to the limit of the speed. In this case, it will always cost you extra. (4). As a commercial cfd code platform, I have used HP workstations in one company and Sun workstations in another. Both can handle the popular commercial cfd codes. (5). For the graphic applications, SGI used to be the best choice because of its graphic hardware speed and available software. But, based on my experience, there was one available in one company, and in the other company, SGI with multi-processors is being used as mainframe computer. (6). In terms of the speed, as I said before, it depends on the particular model in the series. The high end of workstation in the HP line was fairly fast (relative to the rest of the computers in the office). But there was only one in the office (it has 2 cpu and 1 Gbytes of RAM, in 96). So, it must be expensive. (7). At that time, most of the HP in the office can't even process 3-D results of Fluent code. Most have only 128mega bytes of RAM. (8). The same is true for the company using Sun workstations. A year ago, I had to use the mainframe to run the cfd code. The Sun workstation I was using was simply to small and too slow. Since then, it has been upgraded to Sun/Ultra60. Now I can run the commercial cfd codes on my workstation, without worrying about the RAM and Hard disk memory and the machine speed. (9). So, based on my experience, both high end of HP and Sun/ultra can handle the commercial cfd codes smoothly (network is a separate issue, if your workstation is on a large network, its up-and-down will definitely make your workstation life very miserable. ) On the other hand, the low end of workstations are definitely not suitable for commercial cfd applications. (10). The reason behind is that general commercial cfd codes tend to be large (require a lot of RAM memory). And their heavy usage of GUI and pre- and post-processing graphics, demand the use of workstation type of computers. (11). If you are not using commercial cfd codes, then PC will be fairly good for cfd research. (12). So, my suggestion is: look at the high end of the workstation first. And then make the decision in terms of the cost. Not all models of workstation are suitable for commercial cfd applications.

Charles

I think you already have some useful comments, but for what it's worth I think I'll add a few of my own.

If what you are most interested in is flat-out processing speed for a single processor, within your specified budget constraint, then I'm sure you'll find Steve's benchmark timings very useful.

However if your objective is best performance/cost, I would recommend consideration of using Linux PC's, following Tim and Rich's comments. This recommendation brings in a few other issues though as well. With a Beowulf-class machine, you can get the entire development environment for free, which may be appropriate to factor in to your budget if you are developing your own code. However if you plan to run commercial code, you have to find out if the code is supported on Linux - some aren't. Lastly Beowulf machines offer less communication bandwidth than SMP architectures; if you have big problems that have to run in parallel which require a lot of communication then you might end up being disappointed by the performance. I'll stick my neck out though and say that I think most parallel implementations of pressure-based and "density-based" CFD codes are ok in this respect.

When trying to make this kind of decision it's often difficult to move away from qualitative descriptions like "big" and "a lot of" to the more quantitative information which is needed. However I hope my comments have been useful.

Brian.