[laxwendrofzx9r]

Hello, I would like to know your opinion. I can buy a used dual 2x e5 e2660 Xeon workstation for CFD purpose, however is possible to upgreade for better e5 e2690 processors. My question is could you know more or less a performance gain between dual e2660 vs e2690? It is important for me because upgrade could costs additionally around 35-45% of e2660 workstation end price.

[CapSizer]

It would be best if you could actually test it with the software that you are going to run, but of course that is very difficult to arrange. In the absence of such information, you can look at the SpecFPrate (www.spec.org) numbers, which are about 440 for the 2660 and 490 for the 2690. So the difference looks like it could on average be about 11%, which scarcely justifies the difference in price. In more detail, the times for the leslie3d test (a research CFD code) are 1106 and 1080 respectively, which is far too small a difference to justify the expensive upgrade. Others on this forum have emphasised the extent to which CFD performance tends to be determined by memory bandwidth, and this seems to be an example of that characteristic. So if you have the option of tweaking the configuration of this workstation, perhaps you should make sure that you are getting the fastest available memory rather than the faster CPU's.

[rmh26]

The E5 series is still pretty new how are you finding a used dual core system so soon?

[laxwendrofzx9r]

Originally Posted by CapSizer

It would be best if you could actually test it with the software that you are going to run, but of course that is very difficult to arrange. In the absence of such information, you can look at the SpecFPrate (www.spec.org) numbers, which are about 440 for the 2660 and 490 for the 2690. So the difference looks like it could on average be about 11%, which scarcely justifies the difference in price. In more detail, the times for the leslie3d test (a research CFD code) are 1106 and 1080 respectively, which is far too small a difference to justify the expensive upgrade. Others on this forum have emphasised the extent to which CFD performance tends to be determined by memory bandwidth, and this seems to be an example of that characteristic. So if you have the option of tweaking the configuration of this workstation, perhaps you should make sure that you are getting the fastest available memory rather than the faster CPU's.

Great piece of information but maybe do you know where can I find any Ansys-Fluent or OpenFoam benchmark numbers for new Intel architecture?


Sorry for another question, maybe is trivial but is the dual xeon e2660 system faster as 2 nodes single Intel 3960X workstation?

[laxwendrofzx9r]

Originally Posted by rmh26

The E5 series is still pretty new how are you finding a used dual core system so soon?

You can easly find any used dual e2660 workstation for sale

[CapSizer]

Originally Posted by laxwendrofzx9r

Great piece of information but maybe do you know where I can find any Ansys-Fluent or OpenFoam benchmark numbers for new Intel architecture?


Sorry for another question, maybe is trivial but is the dual xeon e2660 system faster as 2 nodes single Intel 3960X workstation?

http://www.ansys.com/Support/Platfor...ent+Benchmarks , see if you can find a relevant benchmark there. The single board, dual socket or two single socket nodes question has come up several times in the past on this forum, without really coming to a firm conclusion. In the "old" days of dual-socket workstations that still worked with a northbridge and a front side bus, it was not unusual to find that the cluster of single socket machines was faster. These days with Hypertransport or QPI links and NUMA, I don't think it is so simple. Personally, I like the convenience and effectiveness of a dual socket workstation, and I've had one or more of them on my desk for nearly all of the last 14 years.

It gets tricky when you bring price into the equation. You may very well be able to buy three or more of the fast single socket machines for the price of a single dual socket workstation, in which case you can probably expect the small cluster to be quicker, if somewhat more of a hassle to deal with.

If you are interested in quite a novel approach, take a look at this system which packs several motherboards into a single chassis:
http://limulus.basement-supercomputing.com/

[JBeilke]

Originally Posted by laxwendrofzx9r

Sorry for another question, maybe is trivial but is the dual xeon e2660 system faster as 2 nodes single Intel 3960X workstation?

I would suspect that the XEON machine might be slower.