Engineering desktop dual xeon build - need some advise

[kearkg123]

hey all,
im helping a friend build a workstation so he can run transport models as he's recently decided to start up his own consultancy.
apologies if the info here is second hand, but i've gathered as much info from him as i can.
he is used to using a lenovo p920 which is configured to have:
2 x Intel Xeon Silver 4110s , 1TB SSD , Quadro P400 , 32GB RAM (2 x 16GB)

the application he runs are transport modelling apps such as PTV Visum/Vissim and Bentley Cube.

i found the application requirements for PTV which is here:
System Requirements | PTV Group

i imagine Bentley would be similar since they are both transport modelling applications.
he says that on the p920 he runs up to 4 instances of the software to run models , but for big models he can only do one because of lack of computing power. so if he can run 4 big models in parallel, it would be a good outcome for him (of course, budget and system specs pending).

he says he would like more cores, more threads , so overall a more powerful machine.
for storage, he needs 1 x SSD to run his modelling on, and say another hard drive just for storage.

so im looking to put together a dual xeon system for him with similar components or better that are in the p920.

i have come up with the following build which comes up to about 7.5K AUD.


Supermicro X11DAi-N Xeon Scalable Production Motherboard
Phanteks Enthoo Pro 2 D-RGB Tempered Glass Full Tower Black SSI-EEB case
2 x Noctua NH-U12S DX-3647
64GB Kingston ecc ddr4 2666 KTL-TS426 (2 x 32GB)
Corsair Platinum hx850 850w modular psu
quadro p400
samsung 970 evo plus 1tb m.2 nvme pcie ssd
seagate ironwolf 4tb nas hard drive (st4000vn008)
tp-link tl-wn881nd n300 wifi adaptor
2 x Xeon gold 5218

some questions that I have are:

would these components work well together?
do i need ECC memory?
are there any other mobo options ?as it seems like the supermicro i have listed is hard for me to find here in aus.


thanks in advance.

 

[alex_stief]

Based on the system requirements you linked, a dual-socket workstation with two low-end server CPUs seems like just about the worst choice for this kind of software. Disclaimer: I don't know anything about this software in particular, I am simply judging by their own recommended specs.
They state that more than 8 cores won't help much, and instead recommend CPUs with the highest possible frequency. Even going as far as recommending overclocking, which is pretty rare for official hardware recommendations. Their rason being that some of the workloads are not parallelized at all. And from my experience, with hardware recommendations like these, it is safe to assume that the software is not NUMA-aware/agnostic. I.e. you will shoot yourself in the foot by going dual-CPU.

It sounds like a latest-gen mainstream CPU (Intel Core 11th gen, AMD Ryzen 5000) would be the faster and cheaper choice.

Edit: after reading the whole thing, things change a bit.
For running multiple instances of this kind of software, you can use a dual-socket setup. Provided you pin each instance to one NUMA node.
There are a few things to be aware of: the CPUs you picked have a 6-channel memory controller. We don't know how sensitive the software is towards memory bandwidth, but I would say better safe than sorry. So for a dual-socket machine, you need 12 DIMMs total.
Since your friend currently can't run more than one instance of larger models, you should find out what the limiting factor is. If it is memory capacity, this is a good opportunity to increase that aswell.
And yes, for this kind of setup, you need reg ECC memory.

 

[EffrafaxOfWug]

On the face of it, this sounds like the perfect situation for a workstation-class processor like a threadripper; if you can hold out for the Zen 3 threadrippers you'll get the likely fastest-outside-of-Ryzen-5000 single-threaded performance* as well as high core count and memory bandwidth if you need it (up to 64 cores on a single socket). The current Zen 2-based threadrippers are already very good, but the Zen 3 generation should be ~20% faster on every core.

There's also the Xeon workstation chips which also clock very highly in turbo mode which might be more accessible from your OEM, as well as supporting ECC RDIMMs more easily. They tend to cost a fair whack though.

Like alex says though, this is only based on vague advice - we only know of one of your programs that's got some limits with single-threaded speed (a good general rule of thumb for most applications TBH - you'll almost always run in to a single thread wall at some point) and you're not sure what the other constraints are. Many times I've seen people saying something is slow due to "lack of computing power" when they're nowhere near maxing out the CPUs; slowness can be caused by hundreds of different things - amount of memory and IO bottlenecks being two of the most common.

Are you in a position where you have a 10th/11th gen Intel or a Ryzen 5000 lying around you can test running the software on to see what difference there is in performance compared to the existing setup? As far as CPUs go those are the single-thread kings at present.

* For most software anyway - it's quite common for some software to be heavily optimised for one architecture over others so it's always best to be able to try a real-world test with the application and data you'll be using; performance in BungholioBench2000 might be identical to or almost completely divorced from each other on Intel vs. AMD for example.