Machine Learning Workstation Build

[jackjcoop]

Hi Everyone!

 

Sorry if this post has been done already a million times (if it has) I'm new to the forums and not entirely sure!

 

Today, I need help building a machine learning workstation (and my first pc build). I currently use a cloud Tesla K80 for my development work, and looking to transition to a physical PC (for various reasons).

 

I've loosely scouted out a build on PC part picker (https://ca.pcpartpicker.com/list/tjPTdX) and have a budget of roughly $6000 (Canadian ruppies). Ideally, I want the best price to performance and have no need for Quadro GPU. I'm open to any suggestions/parts/hacks you might have as well!

 

While I'm new to building PC's I do have quite a lot of experience with ML as well as creating a company in the field. So, if you have any questions about the subject feel free to let me know!

[Crunchy Dragon]

This might just be me, but I feel like a single higher end Quadro would be a better choice than dual SLI 2080Ti.

[jackjcoop]

6 minutes ago, Crunchy Dragon said:

This might just be me, but I feel like a single higher end Quadro would be a better choice than dual SLI 2080Ti.

Hi @Crunchy Dragon !

 

Quadro's don't see a real improvement when training ML, however, Tesla's do. However, they are out of my price range and I have no need for ECC memory.

 

On the other hand, I am considering the RTX Titan over the dual 2080ti

[Crunchy Dragon]

2 minutes ago, jackjcoop said:

Hi @Crunchy Dragon !

 

Quadro's don't see a real improvement when training ML, however, Tesla's do. However, they are out of my price range and I have no need for ECC memory.

 

On the other hand, I am considering the RTX Titan over the dual 2080ti

I'm not sure how well ML scales across multiple GPUs(haven't looked into it much), but generally one powerful GPU is better than multiple weaker GPUs.

[jackjcoop]

8 minutes ago, Crunchy Dragon said:

I'm not sure how well ML scales across multiple GPUs(haven't looked into it much), but generally one powerful GPU is better than multiple weaker GPUs.

That's what I've seen too in some testing (only a 40-60% performance boost in training rather than a 2x).

 

I just found an interesting bench-marking article showing that even with that loss, 2x2080ti's outperforms 1xTitan.

 

Article: https://www.pugetsystems.com/labs/hpc/RTX-Titan-TensorFlow-performance-with-1-2-GPUs-Comparison-with-GTX-1080Ti-RTX-2070-2080-2080Ti-and-Titan-V-1345/

 

One requirement I forgot to mention was the need for 3-4 PCI slots to adequately space the GPUs since have seen a lot of heating issues expressed when running dual setup.