Scientific computing workstation build. Budget $12k-18k.

[femtoseconds]

Hey everyone, long time lurker here.

 

I would like to ask for some help with a new workstation build for scientific computing (both CPU and GPU heavy computations). This build is going to be mainly used for FDTD calculations (CPU heavy tasks) and some deep learning with TensorFlow (GPU heavy tasks) at my university. I have some experience with simple builds, but I feel that this is a bit over my head. My current budget at the moment is around $12k-18k.

 

I did some digging around to find the requirements for such system. Apparently, the total memory bandwidth or overall data transfer rate between the RAM and CPU is crucial for CPU heavy tasks on systems with many cores.

https://kb.lumerical.com/en/index.html?installation_and_setup_hardware-specifications-and-be.html

Another article suggests determining the maximum amount of memory for typical design problems first and designing the system from there.

https://www.lumerical.com/support/whitepaper/parallel_processing_fdtd_solutions.html

 

Following the advice above, I evaluated that our memory need for the time being is around 128-256GB (with ability to expand it later).

• Memory

  • 128-256 GB DDR4 2400MHz

• CPU/Motherboard
  • Currently my biggest struggle. I am looking at Intel Xeon's for a possibly 20-30 core solution.
    • How do I determine the total memory bandwidth between RAM and CPU to choose the optimum CPU for fastest calculations?
    • Should I get one processor with more cores or should I go for a double-processor setup with lower number of cores per processor?

• GPU - NVIDIA Tesla's are better suited for GPU computing, but since CPU computations are more important for me, I opted out for a GTX 1080 Ti

  • ASUS ROG STRIX GeForce GTX 1080 TI 11GB $800 (https://www.amazon.com/STRIX-GeForce-Gaming-Graphics-ROG-STRIX-GTX1080TI-11G-GAMING/dp/B06XY25VTC)

• Storage

  • Samsung 850 EVO 2TB $700 (https://www.amazon.com/dp/B010QD6W9I/ref=twister_B00PRDMHLU?_encoding=UTF8&psc=1)

  • Seagate BarraCuda Pro 10TB $500 (https://www.amazon.com/Seagate-BarraCuda-3-5-Inch-Internal-ST10000DM0004/dp/B01IA9H22Q)

 

All comments and suggestions are welcome.

[Zando_]

Can you wait? If EPYC has dual socket boards, you could get an insane core count, and it'll support ECC RAM if you need that. 

[jcw150]

Personally if I was spending that much money on a PC I would go to a professional, not a forum.

[honor]

You might want to consider getting professional counsel on this, since I thinkthis goes way beyond most peoples expertise on this forum. Especially when you are willing to spend huge sums like that.   Maybe getting a prebuild server with good customer support might be the way to go (sorry if this isn't helpful, but I wouldn't want to give bad advice)

[Moress]

If you need it ASAP then a Xeon setup will be very good as the use of ECC ram is important for critical calculations that you cant have ram errors destroy.
Other than that I think this question is out of the league of most of us here. 

[Nicnac]

I read somewhere that a good rule of thumb seems to be getting at least the amount of RAM that equals your VRAM for neural networks and stuff. might be complete bs though....

I'd second epyc, but Ryzen already supports ecc ram as well.

GPU wise I'd say you go either 1080ti or full professional Quadro cards.