Cancel
Showing results for 
Search instead for 
Did you mean: 

Number of CPUs vs Performance

Genius
Genius

I was wondering if someone has done a study of efficacy of CPUs. I am grateful to have 24 cores, and it seems like using all 24 is about the same speed as using 20, for example. Is there a upper bound above which diminishing returns are seen in the 100k nodes (plate model) size range?

9 REPLIES

Re: Number of CPUs vs Performance

Creator
Creator
Hi,
I plan to invest in a 8 0r 12-core machine, so I'm interested in your question.
Have you read this document? http://community.plm.automation.siemens.com/t5/Femap-Knowledge-Base/FS16-SDA-1-Femap-and-NX-Nastran-...
May be helpful to set your machine properly.

FL

Re: Number of CPUs vs Performance

Solution Partner Experimenter Solution Partner Experimenter
Solution Partner Experimenter

Femap is mostly single threaded, so won't take much advantage of multi-core today, although that may change in the future.  NX Nastran can take advantage of many cores if you set that up.  But it is very solution dependent.  Frequency response scales very well with number of CPU's.  Linear statics is good up to somewhere around 4-8 cores.  NX Nastran likes high speed bandwidth, so setting up the scratch drive with an SSD, or even better, a PCIe based SSD can have a huge effect on NX Nastran performance.

Re: Number of CPUs vs Performance

Genius
Genius

I can confirm that with 8 cores I see no slowdown as compared to 24 cores for my 100k nodes typical plate model.

Re: Number of CPUs vs Performance

Phenom
Phenom

Hi,

 

We have recently purchased a new 20 core machine, and I can report the following brief summary:

 

For Advanced Non-linear explicit simulation (30k nodes), maximum performance was reached using 4 cores.  There was a performance penalty when increasing the core count beyond 8.  

 

I also ran some other test models using a 230k nodes dense cube analysis (still using the standard SOL101 sparse solver, even though iterative is much better for this specific matrix style).  The optimum core count was around 8.  Performance improvements were still occuring all the way up to 20 cores, however, the percent advatange above 8 cores was not enough to justify using the extra cores for so little benefit.

 

Using GPU with an nvidia K4200 provides a performance penalty (for such a dense matrix example), so I switched that off. 

 

All of the test analyses ran entirely in memory, including sizeable SMEM, and all scratch set to a ramdisk.

 

My conclusion was that if I want the best productivity for a single user machine, I would choose a computer with a single cpu of 4 physical cores, with the fastest GHz rating.  Probably an Intel i7-4790K (4.0GHz).  This provides the best combination of Femap modelling performance (this cpu has excellent single thread rating), plus sufficient cores to run standard and advanced non-linear solvers at optimum parallelisation, without spending a fortune on the cpu & computer.  Even though our machines here are all Dell Precision dual cpu with numerous cores, the multi-threading advantages are not big enough beyond 4 cores to make buying that style of computer worthwhile for an individual using Femap.

 

This means that the optimum Femap with NX Nastran single user computer looks more like a top-end gaming machine than an engineering workstation.  Save money on the dual cpu architecture and premium Xeon cpus and spend it on memory, graphics and (preferably PCI) SSD's.

Re: Number of CPUs vs Performance

Genius
Genius
I would add that the ability to run multiple multi-core jobs at once can be
a huge advantage, so having more than 4 cores is still very useful.

Re: Number of CPUs vs Performance

Creator
Creator
Thank you EndZ, very useful (and money-saving) report Smiley Happy
It will drastically modify the config I was looking at (it was a dual Xeon E5-2620 2.4GHz 6 core with 32Go RAM).

Re: Number of CPUs vs Performance

Solution Partner Experimenter Solution Partner Experimenter
Solution Partner Experimenter

On running multiple jobs, each with multiple cores, keep in mind you will very soon saturate the memory bandwidth on a standard workstation, which will slow down the NX Nastran performance.  Fast memory and I/O Bandwidth is key to good NX Nastran solve times.

Re: Number of CPUs vs Performance

Genius
Genius
Good point. I see no slowdown. I think my machine is really badass. Aside
from the 250 GB HD. It's an SSD but is full of old jobs I can't delete.

Re: Number of CPUs vs Performance

Phenom
Phenom
On our last "big" computer (Dell T7500, 96 GB DDR3 RAM, 8 core W5590 and PCI SSD's) we would not run multiple analyses at the same time, as the total runtime would be noticeably penalised. Our best new machine (128 GB DDR4, 2 x 10 core E2640 V4 and a few nvMe SSD's), seems to run 2 x 4 core jobs and 1 x 8 core job simultaneously without any noticeable penalty. However, it is very rare for us to do that much simultaneous analysis - the single thread Femap performance of an i7 4GHz cpu provides more scope for time saving than all those Xeon 2.4GHz cores!