[Veirrei-4]

Hey guys!!
I am relatively new to CFD and have quite bad knowledge in which types of hardware configurations are available and that sets them apart.

I would much appreciate if someone could make a pretty general explanation of what types of configurations are typically used for CFD, how they work and in which situations they are preferably used.

Thanks/ Veirrei-4

[flotus1]

That's quite a broad question.
What level of hardware are we talking about? PC/laptop, workstation or cluster? Or is this kind of disambiguation part of your question?

[Veirrei-4]

A situation I'm considering is performing geometry/mesh/setup on a desktop workstation and then submitting the performing of the simulations to more capable server available at a different department at the company.
In general the meshing of our models result in 5.1 million nodes and 19.6 million elements, a unstructered mesh with three layer prism cells on solid bouandaries.

How should I configure a workstation that will perform these processes of geometery/meshing/setup sufficiently and with relative ease?

[flotus1]

So you are looking for a Workstation mainly for pre- and post-processing.
These are generally lightly threaded applications. So you don't need a CPU with too many cores or even a dual-socket workstation. You are better off with fewer cores (6-8 should be the sweet-spot for now) and higher clock speeds.

Knowing how much money you can spend on the system would help with the decision. Right now you can choose between different offerings from AMD and Intel.
If you want official ECC memory support you have stick with Intel for the next few months. You can use their Xeon E5-1650v4 (6 cores) or E5-1660v4 (8 cores) in this case.
If you don't need ECC support (in my opinion you don't) you can use an I7 CPU like the Core I7-6800k (6 cores). Or you can choose their brand-new "Skylake-X" processors, for example the I7-7820x (8 cores).
AMDs current offerings are a little more on the price/performance side. A Ryzen 5 1600 offers 6 cores with great performance for a reasonable price. The only caveat is the dual-channel memory controller which makes them slower in memory-intensive workloads. Imagine you want to run a small simulation directly on your workstation or the first few simulations of a larger one. This will take longer than with the more expensive Intel processors. This will probably change when AMDs "Threadripper" CPUs are released.

In terms of memory, 32GB might be enough for the models you describe, but you should definitely leave room for an upgrade to 64GB. Running the 20 million cells can even require more than 64GB of RAM, depending on the solver settings. Which memory you need and how to populate it correctly depends on your choice of CPUs.

For the graphics card you don't need anything expensive. Any recent GPU with sufficient memory (>=6GB) should do the trick. For example Nvidia GTX 1060 6GB. AMDs graphics cards like the RX 580 are currently hard to get because of the latest "mining"-boom. From my personal experience, paying extra for the professional line of GPUs ("Quadro" and "Firepro") is not necessary for a CFD workstation. Unless of course you are working directly with the CAD models and need certified drivers for the CAD software.

Then you need an SSD, the size and speed depends on how much money you can spend. Don't go below 250GB. Plus as much HDD storage as you desire.

[donminoti]

Hi all,
I am setting a small cluster of 256 cores ( 8 cores per processor). it will almost run one or two jobs at the same time.

I saw on internet two rules regarding memory requirement.

First one deals with elements number : 2.5 Gb per 1 per one million elements.

Second deals with cpu cores: min pf 4Gb to 8Gb per core.


2Gb per core will be less expensive for me, however if actually 4Gb per core is a must i have to take it into account.

Thank you for advice or feedback experience.

BR

[flotus1]

It is generally a good idea to start your own thread, especially if your topic is not related to the thread you are hijacking

Leaving aside these rules of thumb, your cluster will most likely consist of nodes with two 8-core CPUs. This means 8 memory channels to fill. I really can not recommend using 4GB DIMMs. AFAIK, 4GB DDR4 DIMMs are all single-ranked which yields slightly lower performance. Additionally, these small DIMMs cost more per GB. So 8x8GB are the absolute minimum amount of RAM per node for your cluster.
To comment on these rules of thumb: Depending on the solver settings and the CFD software you use, memory requirements can be higher than 2.5GB per million cells. For example Ansys Fluent coupled solver in double precision is more in the region of 3-4GB per million cells. With these boundary conditions you have enough information to ignore the rule of thumb about memory per core. I think it is mostly useless.

[reza.sayareh]

hi guys , my PC has Intel core i7 6700k with 4 core and 8 threads, i just want to know on the ANSYS fluid setup for number of cores should i choose 4 or 8??if i choose 8 what will happen.

[BlnPhoenix]

Originally Posted by reza.sayareh

hi guys , my PC has Intel core i7 6700k with 4 core and 8 threads, i just want to know on the ANSYS fluid setup for number of cores should i choose 4 or 8??if i choose 8 what will happen.

Try it out and see

My guess is, that it might work, but you will not have much faster perfomance because HT are not real additional cores.

[reza.sayareh]

hi guys , i have an intel core i7 6700k which is 4.2 ghz at stock with 4 cores,i am about to oveclock it to 4.5ghz. i am going to do some simulation on fluent which might take 1 week for each run. since i boost up the speed the voltage will go up and it will generate more heat.i was wondering if it worth over-clocking??

tnx

[flotus1]

No, overclocking the cores alone will not pay off. An 7-6700k will most likely limited by memory.
By the way, the voltage does not increase on its own when you increase the frequency. Whatever you do, you should read a few basic tutorials first.

[reza.sayareh]

Originally Posted by flotus1

No, overclocking the cores alone will not pay off. An 7-6700k will most likely limited by memory.
By the way, the voltage does not increase on its own when you increase the frequency. Whatever you do, you should read a few basic tutorials first.

what i mean to say by increasing the voltage is that , as long as you boost up ur cpu from 4.2 ghz to 4.5 ghz you will need to put more voltage on it so that it will be stable, any ways it seems like oveclocking will not help me .
tnx for the reply

[Offtec]

Flotus, based on your expertise which setup would be a better bang for the buck: i7 7820X or Threadripper 1950X? Running mainly OpenFoam with cases up to 10 mil cells. Would you recommend something else in the 2000EUR range (second hand dual Xeon v2?)?

[flotus1]

Dual Xeon E5 v2 is without a doubt the best bang for the buck with these requirements. This won't change until DDR4 becomes much cheaper.