How to build a workstation with high end AMD Epyc, 4x RTX 3090, 256GB RAM?

[bedilbek]

Budget (including currency): $20,000

Country:  Uzbekistan

Games, programs or workloads that it will be used for: Deep Learning training and inference

Other details (existing parts lists, whether any peripherals are needed, what you're upgrading from, when you're going to buy, what resolution and refresh rate you want to play at, etc): 
Currently we have GeForce RTX 3090 Graphics Card, 10th Generation Intel Core i9-10850K Processor, 64 GB RAM, 1 TB SSD and it is not enough for our current workload and we want even more computational power, so we were thinking of upgrading to 4x RTX 3090 with server grade level high end AMD Epyc 

 

[bedilbek]

Can anyone help guys?

[Stormseeker9]

9 minutes ago, bedilbek said:

Can anyone help guys?

Wouldn’t perhaps those RTX quadro cards be better suited for your application? IIRC they were good for these type of things (deep/machine learning)

[BiotechBen]

Quadro cards are going to be much better suited for these types of workloads. "A" series cards have higher tensor counts and more specialized optimizations.

[igormp]

14 hours ago, bedilbek said:

Budget (including currency): $20,000

Country:  Uzbekistan

Games, programs or workloads that it will be used for: Deep Learning training and inference

Other details (existing parts lists, whether any peripherals are needed, what you're upgrading from, when you're going to buy, what resolution and refresh rate you want to play at, etc): 
Currently we have GeForce RTX 3090 Graphics Card, 10th Generation Intel Core i9-10850K Processor, 64 GB RAM, 1 TB SSD and it is not enough for our current workload and we want even more computational power, so we were thinking of upgrading to 4x RTX 3090 with server grade level high end AMD Epyc 

 

At this level, you're better off purchasing a pre built from some vendor. If you can afford to import from the US, then lambda labs is a great option: https://lambdalabs.com/products/blade

 

Otherwise, there are some options in the EU:

https://www.aime.info/#products-server

https://deeplearningbox.com/

 

14 hours ago, Stormseeker9 said:

Wouldn’t perhaps those RTX quadro cards be better suited for your application? IIRC they were good for these type of things (deep/machine learning)

14 hours ago, BiotechBen said:

Quadro cards are going to be much better suited for these types of workloads. "A" series cards have higher tensor counts and more specialized optimizations.

For ML there's no difference between the A series and regular Geforce ones, other than the VRAM and price.

Also, Quadro were meant for workstations, whereas the Teslas were the ones that went into servers (however, all of those fall under the "A" branding now).

 

Drivers won't make a difference, and they're based off the same chip (GA102), unless you're talking about that GA100 based parts, which are indeed better and way more expensive.