Hello,

asking around the following is the (statistical) voice:

XEON (Woodcrest, Clovertown) performs better for serial jobs.

Opteron (say Socket F) performs ahead of XEON for parallel runs esp. on multiple-core systems, say beyond 8x cores.

Any experience in regard, confirmation, objection, perhaps useful hints out there ?

Xeon up to 8 cores (2 clovertowns per motheboard) will beat the Opteron like a red headed step child. It would still beat the Opteron at higher core counts but Xeons max out at 8 per cores per motherboard.

Clusters are generally much cheaper eg. For a cheap cluster go with the quad core Core2Quads e.g. the Q6600 in January 2007.

Can you substantiate your findings. As I did hear what I reported, from more then one source (which is no substantiaton by itself).

Google

Have you looked at the SPEC scores? Not that SPEC is the last word on what you can expect in terms of performance in CFD, but it does give you some pointers. Judging by that, it's far from clearcut that the quad core Xeons scale that well.

Hello Charles,

a link wrt intended SPEC scores?

thank you, flex

http://www.spec.org/cpu2000/results/rfp2000.html

Hello Charles,

thanks for the interesting link. Quite some up and down, even for the same processor, depending on system supplier and configuration. Surprising perhaps, that they don't really uttilize parallelized major industrial codes in the benchmark packages.

Seeems as this leaves us to hear-say to some extend, still. What I don't really figure is the 'parallel/auto parallel = no' setting under the 2000 / 2006 cpu floating point benchmarks.

flex

I think you need to be very careful about making decisions based purely on SpecCPU. It tells you some things, but definitely not the whole story. For example, the Itanium II processor gets very good scores, but when I ran Fluent on Itanium, I was very disappointed, it was a lot slower than on Xeon CPU's! But SpecFPrate tells you something about how the performance scales, and it definitely doesn't like the quad-core Xeon. My guess is that in the real world, where memory access is so important, the quad core will be even worse, but we need somebody with access to the necessary hardware and software to test a real CFD code on these quad-core processors. When I last looked, Fluent hadn't published such results yet. I think at the moment, if you want an 8-way SMP CFD machine, it has to be a quad-socket dual-core Opteron, but if you want 4-way it could be dual-socket dual-core AMD or Intel. A single socket machine needs to to be Intel though.

Charles,

thanks again, for leaning out. Perhaps the XEON quad-cores gain ground with [eventually] faster FSB clock speed in '07 ?

regards, flex

Anybody's guess, it needs to be tested ... But I think they need more than a fast FSB for the multi-socket SMP systems.

thanks again

have a good day,

flex