We need some suggestions about unix workstations. We are planning to buy a new unix workstation with a 64 byte processor from SUN.It is suggested us to buy a workstation with an amd opteron processor. Is there any benchmark test comparing amd opteron, sun sparc IIIi processor with other workstations from IBM or HP. We are planning run our CFD code written in fortran90 on this workstation.

Any suggestion is wellcome

Bumin Ozsoy

SiSoftware Sandra Professional 2005

There can be a significant difference between performance in benchmarks and real world CFD codes. But check out www.spec.org, what you are after are the SPECFP2000 single CPU (http://www.spec.org/cpu2000/results/cfp2000.html) as well as SPECFPrate_2000 multi CPU (http://www.spec.org/cpu2000/results/rfp2000.html) benchmarks. That will tell you that the IBM Power 5 is the fastest workstation CPU, followed by Intel's Itanium II, then AMD Opteron, and then Intel Xeon. On a value for money basis, Opteron is almost certainly the best choice, but for outright performance you would pick the IBM. Bear in mind that we are now entering the multi-core era, so things will be changing a bit.

thank you very much but what is the main difference between the muti-core cpu and two single cpu for the pallel compuationl thank you again.

In a nutshell, chip manufacturing processes have reached the point where it is very, very difficult to increase the frequency further, so now the manufacturers are putting two (or more) cores on the same chip, which in some cases can even slot into the same socket on the motherboard as the single-core CPU. It's still too early to tell exactly where this is going, but the SPECFPrate (multi-CPU) benchmarks give you some idea. The multi-core sustems have the advantage of very fast communications between the two cores (which is very good), but they now have to share the memory bandwidth between them (which is bad). So how well it works will obviously depend on the particular kind of application. Judging by SPEC results, two single-core CPU's are better than one dual-core, but you now have to start taking total system cost into account.

As an example, for a typical CFD workstation, it is quite cost effective to buy a system with dual AMD Opteron sockets (e.g. from Sun or Fujitsu Siemens, one could also put it together yourself), and to fit up to 8 GB of RAM (2 GB RAM modules are a bit pricey, but these boards have 8 slots, and filling them with 1 GB modules is cost-effective). If you now stick two dual core processors in, you effectively have a quad-processor machine with 2 GB of RAM per processor, and the cost should be under about US$ 8000.00. That is a LOT of processing power for the money. A system with 4 separate CPU's will probably be faster, but a lot more expensive.

At about $750 per node you can get 10 headless networked PCs. This is going to be much quicker if all that matters is CFD number crunching and your code can run in parallel.

People buying unix workstations are not buying them to get high performance CFD number crunching. My guess is that lack of parallelisation, high performance graphics and/or compatability is a constraining factor.

Anyway to address Bumins question I would second Charles suggestion and have a browse around www.spec.org. I would also add that in my experience all the unix workstation suppliers that I have dealt with have been happy to benchmark our own codes and/or run more useful CFD benchmarks than the SPEC ones such as the NAS Parallel Benchmarks. On second thoughts there is one exception which is Apple who denied having a technical support department (I assume they must mean presales support) but they certainly have a huge marketing setup.

PS. On third thought! I am not being totally fair to Apple who put me in contact with a local consultancy/shop who did allow me to benchmark Apple hardware.

The whole issue of choosing between a powerful "Unix" workstation and a cluster of cheap PC's is really a different question, and it is mostly determined by the type of code you will be running. If it is a well-parallelised code for which you don't need to pay for parallel licenses (such as an in-house code), then a big PC cluster is the obvious choice. In a commercial environment, where software licenses cost real money, a fast workstation can get more value out of a given amount spent on software. If you have to pay listed software prices, your software spend can easily be several times what you spend on hardware.