Hi All,

My company needs to make some new CFD hardware purchases. We are wondering whether to wait for the new Nehalem Xeon workstation chip which I gather is due in the first quarter of 2009 or to buy an older CPU design now. This is the workstation version of the technology that Intel call i7core in their desktops.

I've seen a few early benchmarks of the Nehalem Xeon chips for database (SAP) tasks. Has anyone seen a realistic CFD or finite element benchmark for these new chips yet?

I've been told that in this case it may be worth paying the price premium to get the new technology soon after release,due to the much improved memory controller, etc. However, I need evidence. Any thoughts or pointers to appropriate CFD benchmark tests would be welcome.

With thanks and best regards, andy2o

(Waiting until the end of March for the release, then finding out it was not an improvement for CFD tasks would be very annoying!)

On old Intel CPU's the memory speed is the largest bottleneck when running CFD simulations. The new Core i7 processor, or Nehalem as it was called before intel named it i7, has a new memory bus architecture that will speed up memory access significantly. This is especially important on multi-core and multi-cpu machines. Take a look on these memory tests for example:

http://www.tomshardware.com/reviews/...m,2057-19.html

I have not seen any CFD benchmarks for Core i7 / Nehalem CPUs yet. If someone has please let us know. The benefit will most likely not be as big as the memory tests above show, but I am hoping it will give something that is significantly better than the old Intel CPU's.

seems that we need a section on CFD's hardware in new forum, is it?

Dividing up a forum into small topics is a very efficient way to kill a forum. A forum needs a certain critical mass of users in order to keep traffic alive. Hardware is a general topic that I think belongs in a genereal main forum.

i mean split with new forum tools, not current setting;

in which all members can monitor new posts at each sub-section, of course such section can be of low traffic but usually of more interest and worth to keep its posts up for more time.

(of course in this setting lot of splitting is infeasible) I quit know what you want to imply by "killing of forum" (of course it is matter at an immediate glance) but in this case i talk with idea and do not think like you

Andy2O,

I can't find a reference to Nehalem Xeon processors. Are you sure they are already around? The inlet website only mentions Nehalems for single processor boards. The Xeon families are Xeon 3000, 5000, and 7000. Which Xeon processor indicator is used for the Nehalem?

Bart

Xeon chips based on Nehalem wil not be available before Q1 2009:

http://www.xbitlabs.com/news/cpu/dis...Next_Year.html

So no benchmarks for CFD available?

Bart

Nehalem will be out in Jan 2009.

Silicon Graphics, Inc. - (SGI) is commited to Intel platforms and making Linux Clusters scale and perform better.

I am understanding that the performance for CFD on Nehalem will depend on the codes/projects that you are running.

SGI has many benchmarks available for Harpertown (previous to Nehalem) for various codes. The performance difference of Nehalem will vary--there are some cache limitations on Nehalem that could be a 'gotcha' on cache-sensitive codes, but if memory is the issue for you with what you are currently running, it could be a boost for you.

If you are interested in knowing more about SGI clusters with Windows or Linux, you can contact me offline. zavoral@sgi.com

We have benchmarks to show with most commericially available codes, but if you are running home-baked, we can offer a benchmark for you if you'll consider SGI as your vendor.

We have factory integrated clusters that work out of the box and are very affordably priced (Intel/Linux afterall) and world class engineering/expertise to back it up.

Thanks.

Shelly

Next week, I will receive a Core i7 920 with 12Gig DDR3 ram. I will be able to run a benchmark with CFX and give you the comparison with a Xeon X5365 with 8Gig DDR2 ram. I think it will give you an idea of the increase in performance.

I ran the benchmark yesterday on my new computer. Basically, it's a CFX simulation with approximatively 1.1*10^6 nodes. It ran for 15 iterations with a fixed timestep.

First computer: Xeon X5365 8 gig ddr2 ram Run mode: serial run time: 50:55 Run mode: MPICH2 local parallel (4 proc.) run time: 32:56

Second computer Core i7 920 12 gig ddr3 ram Run mode: serial run time: 30:47

I was not able to run on the parallel run mode on the second computer because it seems that actually there is an incompatibility with the X58 chipset. It seems to be link with the "x86info.exe" file as related by:

http://www1.ansys.com/customer/wb/fo...?TID=2859&PN=1

I will try to fix this problem

Fred

Hi, are there any news about the i7 scores? Are You able to fix problem with x86info? Am I right that the single core i7 was a bit faster than 4 coreso of 5365?? Luk

I expect to fix the x86info.exe soon (i.e. within 2 weeks). After that I will be able to give you the results for the benchmark on 4 cores with the i7.

I read that CFX is not able to take advantage of hyperthreading. However I will give you results with hyperthreadung enabled and disabled.

Fred

Ok, thanks a lot.

Luk

Or (in the other way): do You suspect, that it is a case of replacement of x86info.exe file or new verion or any ansys-cpeficied modifications of the ansys (or other) files?

Luk

[PattiMichelle]

I always thought an unaddressed question was the hypertransport thing - Intel bailed on hypertransport a long time ago and it's now AMD's bailywick. I wonder how that affects doing parallel calculations on a single node?

[Luk_Fiz]

Hi all,
Since the forum was reformatted I cannot contact Fred. Meantime, the i7 has become very atractive proposition.
I renew my question: does anyone know about problems of running ansys cfx 10 or 11 on i7?

Luk

[pete]

The i7 architeture has removed the frontside bus and replaced it with QuickPath, which is similar to AMD's Hypertransport.

[Jackill]

Hello
for use CFX 11 on Nehalem you need x86info file from 12 version

i have some results for Nehalem
unfortunately for this test i use only small model (it use about 2 gb ram)
now i waiting test with more RAM space

testmodel1 CFX11sp1
8 core Intel QuadCore Xeon 5345 2,33Ghz RAM 16 Gb 2610 s
8 core Intel DualCore Xeon 5160 3,0Ghz RAM 16Gb 1380 s
8 core Intel QuadCore Xeon 5472 3,0Ghz RAM 32Gb 1620 s



AMD SHANGHAI – 1690 s

Intel Nehalem - 873 s




testmodel1 CFX12 p12

AMD SHANGHAI – 1559 s

Intel Nehalem - 770 s

summary 1620 (Xeon 5472 ) /873(Nehalem) = 1.85 times faster

[Luk_Fiz]

Thanks You very much for these results. I have following remarks:
- is the only thing necessary to run i7 with CFX 11 is to copy x86info file? Can it work also with CFX10?
- impressive speed up of i7 relatively to older Xeons!! Can You say which version of i7 You used (920, 940...)? It seems that in CFX, i7 is much faster than older Xeons than in spec.org tests (where is notices about 1.25 speed up i7 965 vs. 2x5472)

Thanks You very much.

Lukasz

[Jackill]

Yes,
you need only copy this file x86info, it also work with 11 version,
but i don't tested with 10vers

- it was i7 940 with 2.93Ghz , at the moment i think that for big tasks speed up will not so high, simply our vendor try to find more ddr3 memory for our testing