As a matter of fact Los Alamos National Laboratory recently decided to build its new Linux cluster for the advanced simulation and computing (ASCI) program using AMD Opteron processors. More information about this can be found here: http://www.lanl.gov/orgs/pa/newsbull...5/text02.shtml

If you are not familiar with what the AMD64 platform is, then please have a look at this: http://www.amd.com/us-en/Processors/...8_9331,00.html

Boxx Technologies (http://www.boxxtech.com) have already started shipping high-end AMD Opteron workstations, but several other manufacturers also offer similar and more reasonable solutions and they can be located on this page: http://www2.amd.com/us-en/sbl/front/...l?redir=CPOP07

Windows 64-bit for AMD64 is available as a beta version for developers today, but several Linux 64-bit versions (SuSE, Turbo Linux, etc) and compilers (see below) already exist with full AMD64 support. All current 32-bit x86 versions of Windows and Linux will ofcourse work perfectly with AMD64, but then without the benefits of the 64-bit processor architecture.

I've collected some links to pages about AMD64 that I hope you all will find interesting:

Some developer information for AMD64: http://developer.amd.com & http://www.devx.com/amd/Door/16009 & http://www.amd64.org & http://www.migratec.com

Some compilers for AMD64: http://www.pgroup.com/AMD64 & http://www.absoft.com/opteron.html & http://www.nag.co.uk/nagware/np/NP_trial.asp

Some benchmarks of the AMD Opteron processor: http://www.digit-life.com/articles2/...-opteron2.html & http://www.caselab.okstate.edu/research/benchmark.html

Some AMD64 technology in-depth reviews: http://www.aceshardware.com/read.jsp?id=55000251 & http://www.digit-life.com/articles2/...ily/index.html

Actually, a decent Pentium 4 beats it to the ground. In the future I would be grateful if you could please post the propaganda into the "Announcements" section.

That's ridiculus, please document this if you can

All benchmarks I've read so far (mostly 32-bit unfortunately) the AMD Opteron cpu have a good performance lead over the fastest available P4/Xeon cpu's, just have a look here: http://www.amd.com/us-en/Processors/...6_8800,00.html

The Opteron cpu also scales much better in SMP systems compared to P4/Xeon, just read this: http://www.amd.com/us-en/Processors/...6_8799,00.html

Here are some other benchmarks as well: http://www.caselab.okstate.edu/research/benchmark.html & http://www.digit-life.com/articles2/...-opteron2.html & http://www.ensight.com/products/performance76.html & http://www.amdzone.com/articleview.cfm?ArticleID=1328

As you can see, the AMD Opteron cpu performs very well also in 32-bit. 64-bit mode should give it a performance boost because of the increased number of registers available and the higher memory bandwidth.

One of the issues of more interest is : will the 64-bit processor be able to handle more RAM on a PC, than the current 2GB limit?, which currently limits how big models can be run on a PC, say in fluent...

Well, that's just the thing! It does not matter how AMD64 performs against a Pentium (well, it does, but only a bit ) What we all want from it is the possibility of addressing more RAM and running big applications. However, that depends on the operating system, compilers and tons of other things... I'd like to hear from anyone who's running 64-bit Linux and using gcc (C++). Any data is welcome, from "don't bother" to "it's great!".

Any takers?

On Windows XP 64bit edition, applications can address up to 1 terabyte of memory (the whole system can use 256TB totally). 64bit Linux, as far as I know, allows even more memory for applications.

The nice thing is that this works out of the box. Just recompiling an application to 64bit mode and you can use more memory immediately. It isn't as kludgy as the paged addressing of the Xeon (which has pages of 4GB)

I presume you are some sort of salesman? and I feel compelled to provide some sort of balance like Hrvoje

* today, a cluster built of 2.4GHz Pentium 4 is a signficantly better option in terms of price/performance unless your algorithms force you to use expensive interconnects rather then gigabit ethernet (most CFD that needs largish clusters do not).

* the environment (OS support and compilers) for the Opteron is flaky at the moment. This will of course improve with time but if you need the cluster to do work today it is important because fixing it is largely outside a users control.

* the dual/quad/oct motherboards for the Opteron are what we need for CFD (if your OS can drive them properly and most cannot today) and can give a linear increase of speed with numbers of processors. Dual Xeons are a waste of time/money in this respect so use single P4 nodes.

* 64 bit hype seems to have a lot in common with the millenium bug hype. The microprocessors in most unix workstations have been 64 bit for many years (the 7 year old Sun I have recently end-of-lifed has a 64 bit microprocessor) and it had little impact on performance. What matters most to big CFD simulations is memory handling.

If you look at the floating point part of the CPU2000 benchmarks at www.spec.org, the Itanium-2 blows away everything else in sheer performance per processor. On a performance per unit price basis, the Itanium-2 is currently not as competitive as a Pentium 4. Also, most applications have not yet been ported to Itanium-2. I don't know how it compares to an Opteron. Thank goodness for the competition between Intel and AMD, though. It gives us consumers better CPUs at a lower price.

Right, but I'm awaiting more third-party benchmarks, such as CFD codes and the like. I'm a tad sceptical to spec-tests, as they are often tweaked, and bear little resemblance to real-world performance.

Time will, as always, tell.

I like the Itanium, because of the large amount of registers, expilictly parallel instruction set, and 64 addressing. Because of EPIC, the Itanium also handles nearly ordered data much better (1,2,3,5,4,6,7,11, etc...)

Even on clusters, I have run in to addressing problems. If you have any gather scatter operations, your master node can address (but not use) a huge amount of memory. 64 bit address makes these problems go away. I have seen the problem on 32 bit IBM SP's, and Linux Clusters.

I wish Intel would lower the price on the Itanium!

Tony

The main difference between these platforms is that AMD took the x86 architecture and extended it to 64-bits to make the AMD64 architecture, supporting both x86-32 and x86-64 code at full speed & performance. You can read more about this here: http://www.devx.com/amd/Article/16101

The Opteron processor have 64-bit registers and address space as well, but also offers higher front side bus, I/O & memory bandwidth than Intels processors. A comparison of Opteron and Itanium2 based workstations can be found here: http://www.amd.com/us-en/assets/cont...omparisons.pdf

If you're interested in clusters, you should definately also have a look at AMD's presentation at ClusterWorld 2003: http://www.amd.com/us-en/assets/cont...dpresFINAL.pdf

Today, a cluster built of Opteron processors is a signficantly better option in terms of price/performance.