Build your deep learning box: wiki thread

(Sanyam Bhutani) #122

@Moondra I would recommend using a single 2080Ti and not 2xGPU to avoid the hassles of cooling. You’ll have to configure all of the complications of maybe liquid cooling both of them.

I’m not against that, but it seemed to be pretty intimidating to me, so I decided to stick to a single GPU.

Also-I’m not sure of the current scenario but Intel offers some speedup because of the MKL scernios so it might be a better pick than AMD (please verify)

For the Mobo-make sure each GPU gets plenty PCIE lines. Note: M.2 drives use PCIE lines, also you would want to make sure that you’ll have enough RAM for your future needs (64 or 128GB)



@init_27 Thanks for the advice. I have been looking at the complications of cooling and it seems I need to get the cards which have a special type of fan attached. I think it was called “blower fans”

I will look into the MKL related speedups.

As for the PCIE lines, I am looking into that as well.
I willl most likely go with multiple GPUs and let you know how it goes with cooling.

Thank you so much for your advice.


In the first lecture of Part 2 last year (2018), Jeremy gave these suggestions for choosing a GPU:

  • GTX 1080ti is the price performance sweet spot for today (Mar 2018)
  • GTX 1070 8GB is also fine
  • Used cards are fine
  • Dual GPUs let you train while experimenting

I also read Tim Dettmers’ blog on picking a GPU (updated Nov 2018), and the tl;dr version was:

  • Best GPU overall : RTX 2070
  • GPUs to avoid : Any Tesla card; any Quadro card; any Founders Edition card; Titan V, Titan XP
  • Cost-efficient but expensive : RTX 2070
  • Cost-efficient and cheap : GTX Titan (Pascal) from eBay, GTX 1060 (6GB), GTX 1050 Ti (4GB)
  • I have little money : GTX Titan (Pascal) from eBay, or GTX 1060 (6GB), or GTX 1050 Ti (4GB)
  • I have almost no money : GTX 1050 Ti (4GB); CPU (prototyping) + AWS/TPU (training); or Colab.

Although additionally from the performance per dollar graph, the top three are the RTX 2070, GTX 1060, and GTX 1070.

The common factor seems to be that the GTX 1070 is a good option, but it’s surprising to me that this would still be the case a year later. Is there a better choice today to balance performance and price?

Thanks for any suggestions! :grinning:

Update: buried in the comments on Tim Dettmers’ blog is a recommendation for the RTX 2060 as a [relatively] budget option right now.

(Michael) #125

Thanks for this post!

Here is also the link for the updated blog post of Tim Dettmer.
It seems if you can afford it the RTX 2070 is the way to go.

I am also currently thinking about getting my own DL machine. I wonder if I have to be careful with Windows Gaming components as I maybe run into troubles with drivers or is this not a issue at all?
I am thinking of getting a Windows Gaming PC and use it with Ubuntu for DL. The setup which I currently have in mind is: Nvidia GeForce RTX 2070 8GB GPU, ASUS TUF Z390-Pro Gaming motherboard, Intel Core i5 9600K CPU, 16GB RAM.

Do you think this is a good option?
I am happy for suggestions, links, etc.!

(adrian) #126

Windows / Ubuntu dual boot not a problem either can use totally separate drives and select boot drive from bios on startup or have Ubuntu grub loader set boot order win/Ubuntu. With second option can either install windows or Ubuntu first, can’t remember which worked better for me. I just use first option and select from bios.

(urmas pitsi) #127

I would suggest to start right away with 32GB RAM.