[fedez91]

Hi everyone,

I am going to buy a new Workstation to run CFD simulations with OpenFOAM. They will be big 3D meshed with DES or LES, so probably very small cells. My budget is 10,000 U$, and it is probably gonna be a Dell Workstation.

I am thinking in a Dual Core Xeon, and about 64 GB of RAM. Basically I have 3 questions:

1- There are SOOO MANY Xeon options that I am confused. Should I look for as many cores as possible? or should I make a balance between cores and base frequency (N_Cores * Base Freq)? Based on what I read in the past, memory bandwidth is extremely important, but some of the processors don't provide that information (e.g: https://ark.intel.com/products/12048...Cache-2_40-GHz). Is there something else that I should be looking at regarding the processor?

2- How many memory should I be looking for? I honestly don't know how to correlate mesh size with RAM requirements.

3- Is there anything else I should pay attention to? I think that the Quadro4000 that come's with the workstation is gonna be enough for postprocessing, but I don't really know.

Any help would be really appreciated, and if you have a specific configuration that you would recommend that would be really welcome.

Thanks!!

[flotus1]

1) High core count CPUs are usually a waste of money for CFD due to memory bandwidth limitations. For OpenFOAM you should be aiming for a core count somewhere between 12-18. Lower core counts -with higher frequencies- are beneficial for lightly-threaded workloads often encountered in pre- and post-processing. Higher core counts can increase solver performance a bit. 6140 or the cheaper 6130 seem like good options.
All Skylake-SP Xeon processors have 6 memory channels, the only difference is the maximum supported memory speed. Bronze: DDR4-2133; silver to gold 5100(most of them): DDR4-2400; gold 6100 to platinum: DDR4-2666. So pick one of the Xeon gold 6100 SKUs for your workstation. An attempt to make some sense of Intels lineup: https://www.pugetsystems.com/labs/hp...rs-Guide-1077/

2) For a workstation with two Skylake-SP Xeons, you want 12 identical DIMMs for best performance. So either 96GB or 192GB. Run a smaller representative case on the hardware you already have to get an idea how much memory per million cells your solver settings require.

3) This should rather be a Quadro P4000. Using mostly OpenFOAM and Paraview on Linux, you could also use a GTX card like a 1060 6G or a 1070 to save some money. But Dell won't sell them in a workstation. A P4000 should be more than enough.

"4)" With transient simulations and high cell counts in mind, some flash-based storage with at least 2TB would be a great addition to a workstation. SATA or NVMe depends on your budget. Use it as intermediate storage for your current projects, then move them to spinning disks when you are done.

"5)" If you care about getting the most bang for your buck, a workstation with two 16-core AMD Epyc CPUs would be better. See the sticky thread in this sub-forum. But Dell only sells servers with Epyc CPUs, not workstations.

[fedez91]

Hi Alex!

Thank you very much for your answer. Really good advice!!

Regarding the processor, you mean 12-18 cores per processor right? That would be 24-36 cores total.

I think I have reduced my options to Xeon 6140 (18C-2.3GHz/3.7GHz) and Xeon 6136 (12C-3.0GHz/3.7GHz). Is there any reason why you would choose any of them?

Best,

[flotus1]

Yes, I meant cores per CPU.
Deciding between the 6140 and 6136 is exactly the tradeoff you have to make: higher total performance vs. higher performance per core. Since you don't have license costs per core and high cell counts, I would probably choose 6140 when given these two options. The fact that both have identical single-core turbo frequencies makes the decision slightly easier. But again, what I would actually choose myself is AMD Epyc 7301/7351 because it is both cheaper and faster.

[JBeilke]

One drawback of the Xeons compared to the Epycs might be, that no Noctua coolers are available at the moment. So building a quiet machine can be a problem.

[fedez91]

I will try to see if I can make my own Workstation with epycs. I've been doing some RAM test for the solver I am going to use (twoPhaseEulerFoam), which stores a lot of variables compared with other models. 3M mesh uses 8.4 GB of RAM, so that would mean that with 96 GB would be a little bit above 30M cells. Does that seems reasonable?

Thanks!!

[flotus1]

3GB per million cells sounds reasonable. If you ask if 30M cells total sound reasonable for LES...it is on the lower end of cell counts we usually have in our LES. But this certainly depends on the application. But with a budget of 10000$ I would definitely buy more than 96GB RAM. If DELL charges too much for more memory, you can always buy a minimal configuration and add cheaper RAM yourself. DDR4-2666 RDIMM starts at around 11$/GB these days

[HappyS5]

Hello,

Based on COMSOL recommendations of 8 GB per core, I designed a CFD workstation that has worked well for me so far. I work with OpenFOAM. Tell me what you think. I designed both, but selected the SuperMicro(Yellow) to be built.

[HappyS5]

Originally Posted by HappyS5

Hello,

Based on COMSOL recommendations of 8 GB per core, I designed a CFD workstation that has worked well for me so far. I work with OpenFOAM. Tell me what you think. I designed both, but selected the SuperMicro(Yellow) to be built.

The above workstation can properly resolve the Coefficient of Drag and Strouhal number for flow around a circular cylinder in 24 hours.

[flotus1]

Wait a minute... somebody sold you a workstation based around 2x Xeon E5-2650v2 for well over 5000$. In 2021? Or is there something I am missing?

[HappyS5]

Originally Posted by flotus1

Wait a minute... somebody sold you a workstation based around 2x Xeon E5-2650v2 for well over 5000$. In 2021? Or is there something I am missing?

Parts alone cost over $6000. I designed the workstation myself about 5 years ago.