Return to

New Threadripper build for engineering calculations



I am currently in the process of building a new home PC, I have not built a PC in a while so looking for some assistance. I already have bought some parts and have some from my current PC, thus the only components I still require are a CPU, CPU cooler, motherboard and memory. I live in the UK and will most likely be using for parts.

I am a mechanical engineer and the PC will be used for undertaking computation fluid mechanics (CFD) calculations, this makes use of parellisation thus a higher core count CPU is beneficial. This system will be under 100% load 24/7 thus not too keen on overclocking anything and will be looking at air cooling for everything. I will be dual booting the system as I have some needs for Windows through university but my primary use will be Linux (Ubuntu), for this I have 2 separate SSDs.

From previous benchmarks with CFD calculations the sweet spot in core-memory channels is 2-1 thus ideally id like to use AMD Zen CPUs due to the 8 memory channels however the wallet does not currently permit this! Therefor I am looking at a Threadripper system, with 4 memory channels the ideal CPU for CFD calculations would be the 1900x with 8 cores. Thus I have went for this CPU, for motherboard I am looking at an Asus Prime X-399-A.

I am mainly looking for assistance in picking memory as I have heard of some issues with certain memory not working with this platform? I would ideally like dual rank DDR4 at about 3000MHz. I have was looking at 4x8Gb Corsair Dominator Platinum DDR4-3000.

Does this sound like a reasonable selection of parts?

Any feedback would be much appreciated.

I have included a link to a parts picker list of the parts I have, only the CPU, motherboard, memory and CPU cooler have still to be bought, I have already bought most other parts and have some from previous PC.

Kind regards



Since you mentioned that CFD calculations have great benefits with additional cores, why are you choosing an 8 core on a platform that goes from 8 to 16 cores (for this generation)?

If you’re at a budgetary limitation, that’s a fair reason, but frankly, I think that if you can get more performance out of more cores, you should do it.

I’m seeing an interesting selection of storage components. Can you explain your logic behind the mismatched HDDs?

I think that aside from that, you’ve got a great choice in parts.



Thanks for the feedback. It’s been found that using more than 2 cores per memory channel has no great benefit and eventually has a negative impact on solution speed thus I’m sticking with 8 cores for this build. Ideally I’d build a dual 16 core Zen system due to its 8 channel memory per socket but the budget doesn’t stretch that far, yet!

The storage is setup as an SSD for boot and 2tb HDD for general storage for both Windows and Linux, then a 4tb HDD to backup both of the 2tb HDDs.




Threadripper systems only have 4 channels. I’m on a 1950x right now.

If you want 8 channels, you’ll need to go with EPYC.

Okay, that makes sense.

Ah, that’s a sane decision.

Alright, now that my questions are answered, I can say, with confidence, that this is a solid build. I have no concerns with it.



Yeah I meant Epyc haha.

Thanks for your feedback!





I recently did something similar to allow Hamiltonian MCMC for Bayesian statistics. I ended up overspeccing (2950x with quad-channel memory) to allow multiple models to be fit simultaneously. I currently use Unraid with the models being fit in docker containers. This also allows me to fire up a VM if necessary.

Still working on a dockerfile to enable GPU compute to be ready for the next version of STAN.

Just a thought, since a 1950x might not be too far away. The RAM gets pricey though.


Solid build… Honestly do you have to roll the corsair case? You can save some pennies on a few things in there and consider an upgraded CPU… IDK Epyc would be well EPIC lol… Its a very solid build. For memory you should consult both your motherboard and people that use that CPU combination for a better set at 32 gb thats prolly samsung B die since you could tighten timings and bump the speed better. Assuming these kinds of calculations are per thread and memory speed bottlenecked primarily right? IDK its the little details. You got a really solid build