[SOLVED] PC build for Machine Learning

[AL2K]

I'm thinking of build a PC for Machine learning.

My first build is:
CPU - AMD Threadripper 1900X 3.8 GHz 8-Core Processor
CPU Cooler - Corsair H100i PRO 75 CFM Liquid CPU Cooler
Motherboard - Gigabyte X399 AORUS PRO ATX TR4 Motherboard
Memory - 2 x Corsair Vengeance LPX 32 GB (2 x 16 GB) DDR4-3200 Memory
Storage - Samsung 970 Evo Plus 1 TB M.2-2280 NVME Solid State Drive
GPU - Asus GeForce RTX 2070 SUPER 8 GB Turbo EVO
Case - NZXT H500i ATX Mid Tower Case
PSU - Corsair RMx (2018) 850 W 80+ Gold Certified Fully Modular ATX Power Supply

I'm also planning to add the second GPU after some time and a bit worried about the cooling.
Will it be needed to add extra cooling in case of using 2 GPU? Isn't the current build prone to overheating problem?

I have not investigated CPUs a lot so far, so - what would you recommend as an alternative or better CPU option for this build?

Will be glad to know about your thoughts and recommendation!

 

[kanewolf]

I am not a fan of using two 2-DIMM RAM kits. A 4-DIMM RAM kit is guaranteed to work as a set.
How big are the data sets you are training on? A single 1TB disk may not be enough space for your test data sets.

 

[AL2K]

I usually work with 10-200 GB size datasets. They might be slighly big in the future. I thought of adding extra 2-5 TB HD as a storage and decided not to do it for the starter build (cause I can add extra HD without any problems at any time when it is needed).

Thank you for the recommendation about RAM, I will keep that in mind!

 

[null_dev]

Mind AI/ML are dataset/database zombies, so no matter how tough processors hands you have with limited data. A good process power helps of course, but a good database helps MUCH MUCH better.

Having 4 fast 250GB storage is better than having 1TB super fast(for sake of database data grid/clustering), but might not be an appreciated setup.

Happy programming

 

[AL2K]

Mind AI/ML are dataset/database zombies, so no matter how tough processors hands you have with limited data. A good process power helps of course, but a good database helps MUCH MUCH better.

Having 4 fast 250GB storage is better than having 1TB super fast(for sake of database data grid/clustering), but might not be an appreciated setup.

Happy programming

I mainly work with image data and don't use any databases.
Could you explain more detailed about "Having 4 fast 250GB storage is better than having 1TB super fast(for sake of database data grid/clustering)" or give links on some resources where I can read about it?

 

[null_dev]

I mainly work with image data and don't use any databases.

So that's different, you need more process power.

Could you explain more detailed about...

That's logical, having 4 dedicated HW for reading/writing brings more performance over one big one(if you assume same read/write performance), but as mentioned it has its CONS and PROS, you may get performance, but mind the power draw, setup, management, where these are vice-versa for other side.
But also mind, having a grid data storage, needs it dedicated usage. For example, if you go write just linear, so the performance might be the same(even single-driver better), so please note both software and hardware MUST be sync/compatible with each other.

For example in SQL stuff, you have to know how to perfectly split data(a table for example) vertically, or horizontally to achieve best performance at peak sessions. You must have 90% focus over SW after you decided the target HW, to build/develop it completely sync/fit with your hw.

If you are going for study/education, so just take it easy and go for a solid, single storage/process-power, but in term of "enterprise", so have a dedicated time studying both sw and hw architectures, to come up with a great setup.

Just as same as storage, process could be better to be clustered, but again, it DEPENDS!

Happy programming

 

[AL2K]

So that's different, you need more process power.



That's logical, having 4 dedicated HW for reading/writing brings more performance over one big one(if you assume same read/write performance), but as mentioned it has its CONS and PROS, you may get performance, but mind the power draw, setup, management, where these are vice-versa for other side.
But also mind, having a grid data storage, needs it dedicated usage. For example, if you go write just linear, so the performance might be the same(even single-driver better), so please note both software and hardware MUST be sync/compatible with each other.

For example in SQL stuff, you have to know how to perfectly split data(a table for example) vertically, or horizontally to achieve best performance at peak sessions. You must have 90% focus over SW after you decided the target HW, to build/develop it completely sync/fit with your hw.

If you are going for study/education, so just take it easy and go for a solid, single storage/process-power, but in term of "enterprise", so have a dedicated time studying both sw and hw architectures, to come up with a great setup.

Just as same as storage, process could be better to be clustered, but again, it DEPENDS!

Happy programming

Thank you!