[psaracoglu]

Hi all,

I'm posting because I see a lot of advice in the forums for those building a cluster / running simulations, but haven't seen anything for pre-post. I want a workstation mainly for meshing StarCCM+ simulations (with meshes up to 100M cells) and for post-processing results. We can run simulations on the cloud, but would like a fast and capable station for mesh generation, checking meshes, checking solutions, and generating scenes. Occasionally I would like to be able to run just a few iterations of a simulation, but don't see a need for more than 8-10 cores and will probably usually be using a single core for meshing.

Our current setup is slow and can not keep up with the memory requirements. Larger models with 50+M cells are really hard to work with and often freeze.

The budget for this would be somewhere around $3k.

For the new station here's what we've been looking at:
RAM: 256GB DDR4
CPUs:

• Intel i9-10900X (19.25MB cache, 4 memory channels)
• AMD Ryzen Threadripper 2920X (32MB cache, 4 memory channels)
• Intel Xeon W-2245 Processor (16.5 MB cache, 4 memory channels, not available for purchase yet)

OS: Linux (CentOS or Ubuntu?)
Graphics: MSI GeForce RTX 2080

What I would really appreciate is some hardware recommendations for our use case - mostly single core meshing, post-processing, and the occasional few iterations of running. Or thoughts on what we've already considered (above). Also, what are the important metrics to watch out for? It seems to me that CPU clock speed, RAM speed and capacity are key.

Thank you in advance!

[flotus1]

A second gen Threadripper CPU would be the wrong choice here: lower single-core performance, 2 NUMA nodes.

Choosing between a Xeon and I9 is up to you. I9 can be overclocked (both CPU and memory), Xeon is completely locked, but offers ECC support.
If overclocking is not for you, then both choices are about the same in terms of performance.

Pretty short comment, I know. I just had to agree with most of the decisions that led you to this point.
I would probably spend way less on a GPU though. For your applications, the RTX series brings nothing to the table compared to the GTX 1000 series. I would probably go as low as an AMD RX 570 8GB.
VRAM is what matters most when it comes to "how big of a model can I fit". I personally don't mind if interactions then happen at 30FPS or 120FPS.

[psaracoglu]

Alex, I just wanted to belatedly say thanks! We went with the i9, and the GeForce RTX 2080. We're building a second pre/post machine now with the Xeon W-2245, and I think I'm going to try out the RTX 2060 at half the cost.

Your comments across various threads have been really helpful!

[zhaobo]

Hi all,

I come up with nearly the same questions. Similarly, since I can run simulations on the cloud, I want a workstation for my mesh generation and post-processing (with meshes up to several hundred million cells, not exceeding one billion).

And it make me confused that what kind of CPU and how much the menmory are needed at least for such amount of grids, as well as the graphics.

Further more, since CPU and graphics may be overcapacity, is it true that the memory size is the most important factor meeded to be considered. And what kind of relationship between mesh and memory size?

Thanks a lot, and sorry for my poor English.

[flotus1]

Not all mesh generation software is the same, and not all types of meshes are the same.
The safest way for you would be to create some smaller meshes with geometries that are still representative for your use-case. And then plot memory consumption vs cell count. This should allow you to extrapolate memory consumption for larger meshes. You need more than one data point for the extrapolation to be reliable, because the relationship between memory consumption and cell count will probably not be perfectly linear.

Edit: if it should turn out that you need more than 256GB of RAM, a Xeon W-2245 or 2255 might be needed. Intel "Prosumer" CPUs like I7-9800x do not support more than 256GB of RAM. AMD 3rd gen Threadripper CPUs technically support more, but with the limitation of currently available UDIMM memory modules -32GB- the limit is the same.

[zhaobo]

Thanks very much for your advice