[Rec]

Hi, everybody!

There is:

Motherboard: Supermicro H11DSi
CPU: 2 x AMD EPYC 7351 - 2.4 GHz - 16-core
16 x Memory: RDIMM Crucial 8GB CT8G4RFD8266 (Specs: DDR4 PC4-21300 • CL=19 • Dual Ranked • x8 based • Registered • ECC • DDR4-2666 • 1.2V)
SSD: Samsung V-NAND SSD 970 PRO M.2 - 512 Gb
OS: Windows 10 Pro x64
SOFT: ANSYS CFX 2017(in future 2020 R2)

need to increase the speed of calculation.

Need to buy a similar server configuration?
I want to buy a new generation 2 CPU x AMD EPYC 7352 (48 cores), is it possible to correctly distribute the calculation?
What adout Gigabyte C621-SD8 with 8 channel memory architecture (4 channel per CPU)?
Recommended to buy InfiniBand or use Ethernet (1 Gbit)?
If InfiniBand, then what model and speed?


Thank you.

[flotus1]

How much more performance do you need?
That upgrade will get you an extra 30-40% at best. You decide if that's worth the investment.

Need to buy a similar server configuration?

I don't understand that question

is it possible to correctly distribute the calculation?

Same here, not quite sure what you mean. Are you having trouble fully utilizing CPU resources on your current workstation? Ansys CFX should run pretty well on 1st gen Epyc.

What adout Gigabyte C621-SD8 with 8 channel memory architecture (4 channel per CPU)?

That's a motherboard for Intel CPUs. And even if you decide to use Intel CPUs, don't get this motherboard. The CPUs for this socket have 6 memory channels, the motherboard only enables 4 of them.

Recommended to buy InfiniBand or use Ethernet (1 Gbit)?

Based on that question, I deduct that you want to keep your old workstation, and get another one to run distributed parallel across both machines.
Since it's "free", you can just try Gigabit Ethernet first. If scaling across nodes with your simulations is bad, you can still invest in a faster node interconnect.
I.e. run your simulation on the slower machine only, then run the same case distributed across both machines, with twice the thread count. If solver run times are cut in half: good. If not: consider faster node interconnects.

Buyers beware: 3rd gen Epyc will be announced soon™, definitely this year. Actual availability of CPUs and boards might be a different story, given AMDs track record in that department.

[Rec]

flotus1, thanks for your reply! Many of your answers helped me!

Originally Posted by flotus1

I don't understand that question

If i bild a new server, it will be more productive than the old one. How to balance the load of calculations if the speed of calculations is different.
Is it mandatory to build a server with the same configuration when they are working on the same calculation.

[flotus1]

I see...
Heterogeneous clusters are a thing. So you don't necessarily need the exact same hardware on all nodes.
There are ways to leverage the full potential of all nodes, even if they differ in compute capabilities. For MPI+domain decomposition (which is what Ansys CFX uses), there are at least two straightforward ways I can think of: 1) assign larger domains to the cores on the faster nodes. Or 2) assign more domains to the faster node. The latter would be really easy for you, since your new, faster node has more cores anyway. Assuming you go with the 24-core Epyc 7352 CPUs.