Best EPYC for CFX and Fluent (running 1 x 32 or 2 x 16 cores)
 

Dear all,
I am trying to figure out the best option for running ANSYS CFX / Fluent tasks at 1 x 32 or 2 x 16 cores.
Considering keeping some room for meshing or post-processing, dual-socket, 24 cores 3rd generation EPYC seems to be the best option for me.


Can you please help me with this?
Is 74F3 worth it (when compared to 7443)?
And what about 73F3? Why is this processor even more expensive than 74F3? I know, the base frequency is higher. But is assume with the Turbo this is irrelevant when comparing the same count of active cores?

I haven't looked that closely at AMDs current lineup, since most of the CPUs were not available anyway. You are right, pricing the 24-core "F" variant lower than the 16-core is weird.
From a technical standpoint, there is no obvious reason why the 74F3 should be worse. Maybe AMD had a very specific target market in mind for the 16-core, with software that has a license bracket at 32 cores total. Who knows, even reviewers could not make any sense of it.
So if you can, definitely get 2x 74F3 instead of the 73F3.
As for the price difference to the 7443... it's roughly 900$ more per CPU going by list prices. Factoring in the rest of the workstation, license costs, and the value of development time in general, I would consider that a good investment. Those huge L3 caches are really tempting.

Dear all,
I noticed the latest Xeons have 8 x DDR4 3200 MHz, too.
Thus, in terms of memory subsystem performance they should be very much comparable.


Do you have any tips what might be favorable for 1 x 32 or 2 x 16 cores (ANSYS CFX or Fluent)?
I am considering EPYC 74F3, Xeon Gold 6354 or Xeon Platinum 8358 (all three options as dual-socket).

For software like Ansys CFX or Fluent, always buy dual-CPU if you can afford it.
Whether the latest Intel or AMD server CPUs are the better choice... Intel did make some improvements with their 3rd-gen scaleble architecture. Especially in the memory subsystem department. But if a system with 2x Epyc 74F3 is within your budget, just get that. It is the fastest CPU money can buy right now for CFD software with per-core license fees.

Ok - thank you for the recommendation.

what would be the best (value for money) motherboard supporting such dual AMD system?

I think we are slightly out of value for money territory with these "F" SKUs.
In the system builder market, there are only two options anyway: Supermicro H12DSI and Gigabyte MZ72-HBO.
Level1Techs did a review on the latter recently: https://www.youtube.com/watch?v=6v6VdVXoNOs
That's looking pretty promising. As usual, there are different revisions of this board floating around. Rev. 3.0 is what you want for Epyc Milan.

Dear,
Kindly check this out:
https://www.spec.org/cpu2017/results/cpu2017.html
you might check 73F3 v.s 74F3 results. I believe 50% more cores count on 74F3 would only improve 25% scale up in calculation it means 73F3 cores are more powerful than 74F3's and of course the price is higher. Even 75F3 with double core counts of 73F3 only improves 36% in calculation. So the main aim for 73F3 might be super efficient cluster nodes connected with low latency fabrics like InfiniBand for CFD cases. as the market is special the price would also be. Some CFD Multi-physics programs like ESI Procast recommends core count to the integer power of 2 only or you would encounter scales up problem. My experience shows one or two cores should be off calculation for OS tasks. but unfortunately AMD does not have 18 cores CPU at all. Does any one have similar experience?

Regarding the 74F3 vs 73F3:
The crucial question is whether 73F3 running 16 cores is faster then 74F3 running 16 cores. Personally, I cannot see any reason for this.
So, if the scheduler can distribute the load properly (i.e. 32 cores as 16 + 16, not 24 +8), then I do not see any reason to prefer 73F3. At least for a dual-socket CFD machine.

Yup. As long as you distribute the threads evenly among all 8 CCDs, there really should not be any performance difference. One of the big advantages of Zen3 over Zen2 is that one CCD equals one CCX, with only one slice of L3 cache.
My best guess for the price difference between these CPUs: some software has license cost tied to the number of physical cores present in your system, or number of physical cores per socket. Unlike commercial CFD/FEA packages, where you pay for the number of threads you are actually using. So there might be some edge-cases where paying more for the fastest possible 16-core CPU is cheaper in total than paying less for a CPU with 24 cores.

Carter83, I believe, asked if dual CPU is better than a single CPU with the same # of cores. Dual CPU is better because of the more memory lines and memory bandwidth.
My question is if you have a single CPU with 8/16(cores/subcores) then what is better for parallel processing in CFD - selecting 1x16 or 2x8 option (providing the licensing supports 16 threads)?

If you have 8 physical cores, than using more than 8 cores for simulation is an overkill - it is slower than using only 8. Thus, both options (1x16 or 2x8) are wrong.


P.s. If you are talking about ANSYS, than these two options are not equivalent (in terms of licences), anyway. For 2x8 you need two solvers, which is much more expensive.