[Drunken_SheeP]

Hei,

For my institute we are looking to invest some money into a workstation that will allow me to run my CFD simulations faster. I'm running FENSAP-ICE, which is an ANSYS product used for icing simulations. Pretty much similar to Fluent. I'll all pre- and post-processing on another workstation.

Currently I'm running on a workstation with a Xeon E5-1630 v3 @ 3.70GHz, 4 cores, 10M cache and 128GB RAM. An average simulation with below 100k cells takes right now 2-3 hours which is way to long for me.

Currently there are two options. First is to upgrade the current workstation and and replace the current CPU with more cores, e.g. with a Xeon Phi E5-2683 v4 @ 2.1GHz, 16 cores, 40M cache. This would be the cheapest options, and therefore maybe the most interesting one. Is there any rule of thumb way to estimate how much faster the computations will run with more cores but lower frequency?

The second option is to buy another workstation. We do have some funding for this, maybe 6-7000EUR. Any suggestions at what kind of solutions to look for? I have found an interesting setup using the Xeon Phi 7210 @ 1.3GHz, 64 cores, 32M cache. I can use up to 80 cores (5 tasks à 16 cores) so maybe more cores would give me the best performance?

[flotus1]

Originally Posted by Drunken_SheeP

[...]
Currently there are two options. First is to upgrade the current workstation and and replace the current CPU with more cores, e.g. with a Xeon Phi E5-2683 v4 @ 2.1GHz, 16 cores, 40M cache. This would be the cheapest options, and therefore maybe the most interesting one. Is there any rule of thumb way to estimate how much faster the computations will run with more cores but lower frequency?

The second option is to buy another workstation. We do have some funding for this, maybe 6-7000EUR. Any suggestions at what kind of solutions to look for? I have found an interesting setup using the Xeon Phi 7210 @ 1.3GHz, 64 cores, 32M cache. I can use up to 80 cores (5 tasks à 16 cores) so maybe more cores would give me the best performance?

The first option you are looking for seems to be a normal Xeon processor, not a Xeon Phi.
There is no general rule of thumb how much faster your solution times could be or if they are faster at all with a slower processor that has more cores. First thing you have to check if there is a bios version for your mainboard that supports v4 CPUs.
Your example consists of very few cells. It is possible that the solution times are much faster with your new processor because most of the simulation will fit into the L3 cache. But it could also be that the strong scaling is so bad on many cores that your solution times are actually higher.
A safer option would be a CPU with a relatively high clock speed and 8-12 Cores.
Xeon E5-1660 v4 or Xeon E5-2687W v4 are some of the options.
To get at least an idea how the strong scaling might be on 8+ cores you could remove some memory from your current workstation to get a dual-channel memory configuration. Then you could check the strong scaling with your current CPU going from 1 to 4 cores. If the scaling is ok there is a good chance the simulation will scale well on 8+ cores with quad-channel memory.

Your second option is even more risky. The KNL Xeon Phi processors can run normal code without any further adjustments. But because the CPU architecture is quite different from normal processors you will not see any benefit. The slow clock speed will result in much longer simulation times compared to your current workstation. Using this CPU effectively requires at least recompiling the application, but changes to the code are necessary to unleash its full potential.
I don't know if Ansys is working on architecture-specific optimizations for your software and how long it will take. I can not recommend this option unless Ansys guarantees you these architecture-specific optimizations and backs up their claims with benchmarks that are similar to the simulations you run.

[Drunken_SheeP]

Thank you very much. Maybe I was a bit too quick with my considerations. I'll try to look into this a bit more in detail. Any recommended reads on the topic?

[flotus1]

Concerning KNL Xeon Phi:
https://colfaxresearch.com/how-knl/