[songyi719]

Hi, I am going to build my first pc, and I am trying to make it CFD-able for my personal project.


I am considering about option of buying cheap Epyc or Threadripper from chinese guys in ebay and build single socket workstation.



Those CPUs have hell lot of cores, but I know that memory bandwidth is always the bottleneck, and it is normally decided by memory channel.



Lot of people says 2~4 cores per channel is good, but is it enough with only 4 memory channels for 16 cores?



Also, would it be better choice to just wait for more DDR5 rams with higher bandwidth so I can build similar performing system with less budget?


What is best way to find out if my configuration will get bottlenecked by memory before actually building it?

[flotus1]

As a hobbyist, you don't need to overthink this issue.
Cores per memory channel is fairly important when operating expensive solvers with per-core licenses. In order to maximize how much you get out of the limited licenses.

For your personal projects, I would assume you are using some free or open source software like OpenFOAM.
Here "cores per memory channel" is just a question of diminishing returns vs hardware costs. It doesn't really matter if you only get a 1.5x speedup when going from 16 to 32 cores. That's still faster at no additional software costs.
And for the hardware, you need to factor in the whole system. For AMD Epyc CPUs, motherboards in particular can be expensive. Factoring in the total cost for the system, the choice between a used 16-core and a used 32-core CPU is likely a no-brainer.

DDR5 alternatives: Apart from the mainstream dual-channel platforms that are already available, there are no interesting releases in the future in terms of price-to-performance.
And these platforms are also the better choice compared to all "HEDT" platforms like AMD Threadripper or Intel X299.
Not necessarily for peak solver performance. But since you will also spend quite some time with lightly-threaded tasks during pre- and post-processing, the superior single-core performance of the latest mainstream CPUs should be factored in.

In short: for a single-socket workstation, I would either recommend a used 2nd/3rd gen Epyc CPU for maximum solver performance. Or one of the dual-channel DDR5 desktop CPUs (Ryzen 7000, Intel 13th gen).
Another data point that could help make the decision, is how much memory you need. DDR5 is still expensive, and you can't go beyond 128GB currently. 64GB really if you want fast DDR5.
But that's all theoretical without some approximate budget

[songyi719]

Originally Posted by flotus1

As a hobbyist, you don't need to overthink this issue.
Cores per memory channel is fairly important when operating expensive solvers with per-core licenses. In order to maximize how much you get out of the limited licenses.

For your personal projects, I would assume you are using some free or open source software like OpenFOAM.
Here "cores per memory channel" is just a question of diminishing returns vs hardware costs. It doesn't really matter if you only get a 1.5x speedup when going from 16 to 32 cores. That's still faster at no additional software costs.
And for the hardware, you need to factor in the whole system. For AMD Epyc CPUs, motherboards in particular can be expensive. Factoring in the total cost for the system, the choice between a used 16-core and a used 32-core CPU is likely a no-brainer.

DDR5 alternatives: Apart from the mainstream dual-channel platforms that are already available, there are no interesting releases in the future in terms of price-to-performance.
And these platforms are also the better choice compared to all "HEDT" platforms like AMD Threadripper or Intel X299.
Not necessarily for peak solver performance. But since you will also spend quite some time with lightly-threaded tasks during pre- and post-processing, the superior single-core performance of the latest mainstream CPUs should be factored in.

In short: for a single-socket workstation, I would either recommend a used 2nd/3rd gen Epyc CPU for maximum solver performance. Or one of the dual-channel DDR5 desktop CPUs (Ryzen 7000, Intel 13th gen).
Another data point that could help make the decision, is how much memory you need. DDR5 is still expensive, and you can't go beyond 128GB currently. 64GB really if you want fast DDR5.
But that's all theoretical without some approximate budget

Thanks for great answer.

By the way, It was very surprising that commercial DDR5 system is better than HEDT config. Should I understand that 2 channel DDR5 is better than 4 channel DDR4, but not yet against octachannel?

Also, I saw your thread about cfd hardware, and it was amazing. You wrote that more than 8 cores in dual channel is waste of money for cfd. Does it also apply to DDR5? Or would higher memory speed of DDR5 be able to sustain their bandwidth? I am curious which cpu will perform btw Amd and Intel.

I am estimating my budget in very broad range. Lower end is 1000$, and upper end is 2500$. But everybody wants to save money, and since this isn't my CFD-only workstation but will serve many roles, (you know, gaming and daily things) I guess I will build cheaper system if it doesn't show significant difference.

[flotus1]

By the way, It was very surprising that commercial DDR5 system is better than HEDT config. Should I understand that 2 channel DDR5 is better than 4 channel DDR4, but not yet against octachannel?

It's a bit of a tradeoff. Currently available HEDT systems with their quad-channel DDR4 memory controllers, might be ever so slightly faster than the latest desktop platforms with their dual-channel DDR5. But only for the solver phase.
On the other hand, the latest-gen desktop platforms offer vastly superior single-core performance. Which will benefit you in interactive tasks. The kind of stuff where you sit in front of the computer, eagerly waiting for it to finish. Geometry preparation, grid generation etc.
And you already mentioned that this is an "all-purpose" workstation. Higher single-core performance is always preferred IMO.
WIth the used Epyc Rome or Milan CPUs, you need to know where your preferences are. They are significantly faster for solving CFD models, but lack behind in everything else. It's not unusable, I'm still on 1st gen Epyc Naples for my personal workstation. But you have to be aware of this tradeoff.

You wrote that more than 8 cores in dual channel is waste of money for cfd. Does it also apply to DDR5

To some extent, I still stand by that assessment.
For Intel, Things are easy. You don't get more than 8 performance cores anyway.
AMD is more nuanced. By going from 8 to 12 cores on Ryzen 7000, you get twice the amount of L3 cache, which is always nice. And if you want to overclock memory (i.e. use anything faster than DDR5-5200) having 2 CCDs (12 cores and up) instead of just one CCD (8 cores and below) can be beneficial. The interconnect between a single CCD and I/O-die is not fast enough to saturate higher memory bandwidth. Considering how much motherboards and RAM for these CPUs cost, the 12-core 7900X is worth it if you also want to use faster memory.