4GPU PC for Deep Learning - Compatilibity Check

Sep 19, 2018
3
0
10
Please tell me if those components are compatible:

Motherboard: Gigabyte X399 AORUS XTREME
GPU: 4x MSI GTX1080 GeForce GTX 1080 8GB
CPU: AMD Ryzen Threadripper 1900X
Cooler: CPU Arctic Freezer 33 TR Red 120mm (ACFRE00038A)
RAM: HyperX Fury Red 32GB [2x16GB 3466MHz DDR4 CL19 DIMM]
PSU: CORSAIR HX1200I 1200W
case : Kolink Vault (VAULT R)

Thank you!
 
which country is it and budget?
Can you describe major usage? What software are you using?
quad channel ram is crucial for TR performance
CL19 is kind of too large imo
Are you planning on any overclocking? I will probably get a 1300W PSU as well since stock -> 180*5+~100 = 1000W already
The case does not look promising, I will get a full tower with very good air flow.
 
Sep 19, 2018
3
0
10
Thank you for your answer.
Poland. I don't have a budget limit. I just need a computer for 24/24h neural networks training, with 4xGTX1080 and resonable CPU.
ML Frameworks on Linux. Like Tensorflow or PyTorch. They are running CUDA operations on GPU.
I neved had a problem with RAM performance on NN training, because I'm only feeding GPU with images from SSD(or even HD). So RAM is not a bottleneck. GPU is.
I don't think I will do any overlocking. Because I'm afraid that life span of those GPU would be shorter because of that (how I said i plan to use this computer even non-stop).
Thx for the tip about more powerfull PSU than 1200W. Which one would you recommend? Some model name or link pls :)
Why do you mean by CL19 is too large?
 
Description, True Latency (ns), Clock Cycle Time (ns)
3200 MHz CL 16 ,10.08, 0.63
3466 MHz CL 19, 11.02, 0.58

https://www.reddit.com/r/Amd/comments/63gc1s/ram_speed_and_cl_equivalance/

Lower the better, thus I will go with 3200 CL16 instead and could be even cheaper.
 
Any reason not going 1080ti?

1*ll6taa1F3jdjmUY8PahDIw.png


This shows me the scaling of # of gpus is sublinear, it looks roughly 11 for ti, 8 for non-ti, so 3 1080ti may be faster than 4 1080.

pic_disp.php