CFD Online Discussion Forums

CFD Online Discussion Forums (https://www.cfd-online.com/Forums/)
-   Hardware (https://www.cfd-online.com/Forums/hardware/)
-   -   Desktop - CFD and Machine Learning Workstation (https://www.cfd-online.com/Forums/hardware/231219-desktop-cfd-machine-learning-workstation.html)

Sebs October 27, 2020 08:06

Desktop - CFD and Machine Learning Workstation
 
Hi,
I am a doctoral student in a lab, that in the past has done work mostly in numerical simulations (Matlab/Fortran/C/Python).
Recently, we have been shifting more and more in the direction of machine learning, so I build some computers with just regular hardware (Ryzen 3700x and some beefy GPUs for tensorflow), which work well. Now we also want to do more in the direction of CFD - namely StarCCM, Fluent and OpenFoam. We have the budget to build a workstation that will mainly be used for CFD, but might sometimes also be used to run some ML and or our Matlab models.
I am fairly well versed in hardware, but am not so familiar with the Epyc CPUs and what the best bang for the buck would be for the following requirements:

Budget:USD5-10k
Usage: Mainly OpenFoam and Fluent (maybe also StarCCM and ML and Matlab simulations)
Special considerations: The workstation needs to be in the office, since we don't have a server closet available. So it should not sound like a jet engine.

Any help would be appreciated.

flotus1 October 28, 2020 07:57

From AMDs Epyc lineup, I would recommend either the 24-core 7352 or the 32-core 7452. Depends on your budget, 5-10k is quite a large range.
As always, paired with 16x16GB DDR4-3200 reg ECC memory, or more depending on your requirements. Preferably dual-rank instead of single rank.
One possible advantage of 2nd gen Epyc for your non-CFD applications might be support for PCIe 4.0, given the right motherboard. Nvidias latest GPUs support that. You probably know best whether PCIe bandwidth matters for your applications.

On the other side of the fence, Intel is marketing its latest Xeon CPUs with a feature called DL boost: https://www.intel.com/content/www/us...ing-boost.html
It's basically another AVX-512 instruction targeted at deep learning algorithms. Whether you can make any use of it, I don't know. I have zero experience with machine learning.

Sebs October 28, 2020 08:39

Honestly, this will be a ML machine as a secondary priority. CPU power does not really matter much for that anyways - GPUs are orders of magnitude faster. So ideally I would find a board, that could accept at least two PCIe x16 graphics cards (and I would choose a power supply that's dimensioned big enough).

So let's say I have a budget of 6k to start, with the main priority CFD, what would be a recommended hardware combination?

Oh, and I saw you are from Germany - I am as well - just studying in the US right now - fortunately the hardware prices here are much cheaper than back in Europe :D

flotus1 October 28, 2020 09:47

Something along those lines: https://geizhals.eu/?cat=WL-1781998
Storage is intentionally left blank, since I can't know what you need. The only mass storage is a small SSD for the OS.

Sebs November 3, 2020 19:54

Ok,

how about this kind of configuration, as a high end system?

https://imgur.com/sNMep6m

Anything that looks out of place?

flotus1 November 4, 2020 04:30

Oh, look how cheap hardware is in the US compared to Europe :p

H11DSi, two problems:
Only revision 2 boards can take Epyc Rome CPUs. Not all shops keep up the distinction between revisions.
It's only PCIe 3.0

The memory you picked is single rank (1Rx4). If you have pay that much anyway, you should be picky about that detail.

Sebs November 4, 2020 10:08

Yeah, I was surprised, that the server components are equally expensive here. All pc components seem to be very expensive due to COVID atm. 'Asche auf mein Haupt'

Which RAM would you suggest, that is dual rank? I am having difficulty finding any on Newegg.
These are the only ones I can find Click

And regarding the board: there do not seem to be many options out there. The Gigabyte board does not seem to be easily available in the US?!

flotus1 November 4, 2020 12:06

Browsing Newegg from within the EU would require a VPN, so I can't really help you with that.
What you can do is search for the model numbers directly, and see what comes up.
https://geizhals.eu/?cat=ramddr3&sor..._RDIMM+mit+ECC
E.g. type M393A2K43DB3-CWE into Neweggs search bar.

The AMD Epyc motherboard situation for system builders has always been like that. You only have these two options - Supermicro H11DSi or Gigabyte MZ72-HB0. In fact, there was only one option up until a few weeks ago. I have read from other people that the MZ72-HB0 is still not available in the US.
That sums up AMD Epyc nicely. The CPUs are good, but the ecosystem around them is still a joke compared to Intel.

I don't know your time-frame for the purchase. If you can still wait a bit, maybe the Gigabyte board becomes available in your region. What is definitely going to happen within the next few weeks is the "launch" of Zen3-based Epyc Milan CPUs. Maybe this time there will be some new motherboards along with it. Fingers crossed...
If you can't wait, H11DSi with PCIe 3.0 is your only option right now.

Sebs November 4, 2020 12:50

I actually found out the Gigabyte board for sale here in the US from a different vendor. So it would be an option if it has advantages.

As for the wait. We could potentially wait a bit. Would buying a Milan instead of a Rome Zen CPU/CPUs make sense as well, or is Rome the best bang for the buck?

flotus1 November 4, 2020 14:12

The main advantage of that Gigabyte board is PCIe 4.0.
It also has better power delivery, but that won't make a difference with the CPUs and RAM you intend to use.

Hard to tell what price/performance of Epyc Milan might look like, there were no leaks about CPUs and prices yet. Judging by Zen3 for desktops, AMD might increase their prices quite substantially.

Sebs November 4, 2020 18:57

I think, that I can get my hands on the Gigabyte board. The fast PCIe will be important, if we ever decide to drop some GPUs in there for ML.

How loud will this machine be? Not louder than a regular PC, I assume? My prof asked me about it, since he only knows the fans of real servers - and we will have to keep this in our office.

flotus1 November 5, 2020 03:39

Without any powerful GPUs, this PC can be relatively quiet. It depends a bit on which options for fan control you have with that Gigabyte board. If it should be as bad as Supermicros fan control, there are always options to work around it.
Depending on which GPUs and how many of them you add, things will get a bit louder. But as far as air cooling is concerned, this case with 6x140mm Noctua fans is about as good as it gets.

Speaking of GPUs: the board has onboard graphics, but you should at least add some graphics card in the 100$-200$ range for anything 3D. I kind of forgot that in the configuration I posted.

Sebs November 5, 2020 16:39

Thanks,
yeah, we have multiple older workstation GPUs around. Those should be good enough for now. We also will probably use this machine headless most of the time.

Just a side question. Which OS is recommended for a headless machine for Fluent and/or Openfoam? Windows server or should I use a Unix based OS?

flotus1 November 5, 2020 16:54

I would highly recommend Linux. All software mentioned so far supports it, and OpenFOAM doesn't run natively on Windows.

Sebs November 5, 2020 16:55

That's what I was leaning towards to anyway. Thanks!
It should be pretty easy to setup a scheduler, so that multiple people can just send jobs to the server, right? So far we have only been using CFD software independently on our workstations.

flotus1 November 5, 2020 17:01

Up until this point, I successfully managed to avoid setting up a queuing system myself. So unfortunately, I can't share any insight.

Sebs November 6, 2020 17:21

Ok, I will figure the scheduler out.

I am having two systems together now with prices. I want to give my prof two options at different price points.
Anything that can be improved? Click me!

Sebs November 6, 2020 17:32

I just realized that the board is wrong in the configuration, so neglect that.

flotus1 November 6, 2020 18:04

Noctua NH-U12S will do the job, but the NH-U14S has more surface area and quieter fans.
Other than that, I see nothing wrong with the build.

Sebs November 7, 2020 11:25

I Actually reached out to some HPC suppliers and it looks like getting a pre built system would be actually cheaper than building it myself, since they get deals on the server hardware.
This was one suggestion, that I worked out with them. The only concern I have is noise level, but I was told "it is not very loud". Not sure if I believe that with a server PSU and fans.
CLick

As an alternative, I will probably try to work with them in sourcing exactly the parts I want (Non server chassis and fans and PSU) and even then come out ahead vs. building it myself. They can also source the Gigabyte board in the US, but delivery would be beginning of December.


All times are GMT -4. The time now is 15:24.