Workstation for Deep Learning

[aeduG]

Hello,

I would like to assemble a system for deep learning, but I am not an expert when it comes to selecting the optimal parts.

As you can see below I wish to use 4 GPUs. I want to avoid possible bottlenecks.

I hope to find someone in this forum who has experience with these kind of builds.

 

Here is my tentative selection of parts:

 

CPU: Intel Core i9-7900X X-Series (LGA 2066, 3.3GHz, Unlocked)

SSD: Samsung 960 PRO (512GB, M.2 2280)

POWER: Corsair HX1200

HDD: WD Black (6000GB, 3.5", Desktop, Workstation)

COOLING: NZXT Kraken x62 inkl. AM4

CASE: Corsair Obsidian 750D

BOARD: MSI X299 XPOWER GAMING AC (LGA 2066, Intel X299, E-ATX)

RAM: 2x CORSAIR Vengeance LPX 4 x 16 GB (total 128 GB)

GPU: 4x ASUS TURBO-GTX1080TI-11G

 

Alternative AMD Build:

 

Board: ASUS ROG ZENITH EXTREME (TR4, AMD X399, E-ATX)

CPU: AMD Threadripper 1920X (TR4, 3.50GHz, Unlocked)

COOLING: Noctua NH-U14S TR4-SP3 (16.50cm)

 

Edited September 26, 2017 by aeduG
Changes for AMD build

[yaya1]

I don't really know what is best to be used for deep learning, but I'd go for AMD Ryzen Threadripper for the higher core count. But as I said, I don't really know too much about deep learning and therefor am not sure if cores or clocks are "better".

[Mr alex]

Only 4? :0

[TVwazhere]

1 minute ago, Jonas_2909 said:

I don't really know what is best to be used for deep learning, but I'd go for AMD Ryzen Threadripper for the higher core count. But as I said, I don't really know too much about deep learning and therefor am not sure if cores or clocks are "better".

Threadripper also has 64 pcie lanes to allow for all 4 cards to bun at PCI-E 3.0x16 speeds (also not sure how important that is to deep learning) I think it's shown that 3.0x8 isnt fully utilized in video games, not sure if that holds true for this application

[aeduG]

2 minutes ago, Jonas_2909 said:

I don't really know what is best to be used for deep learning, but I'd go for AMD Ryzen Threadripper for the higher core count. But as I said, I don't really know too much about deep learning and therefor am not sure if cores or clocks are "better".

ok I will have a look at ryzen

[aeduG]

1 minute ago, Mr alex said:

Only 4? :0

I think this is the max I can fit into the case/board/budget

[aeduG]

16 minutes ago, TVwazhere said:

Threadripper also has 64 pcie lanes to allow for all 4 cards to bun at PCI-E 3.0x16 speeds (also not sure how important that is to deep learning) I think it's shown that 3.0x8 isnt fully utilized in video games, not sure if that holds true for this application

I looked into threadripper and I can consider it but what cooling solution do you suggest? Water or air?

Edited September 26, 2017 by aeduG
type

[TVwazhere]

Just now, aeduG said:

I look into threadripper and I can consider it but what cooling solution do you suggest? Water or air?

If most of your load for deep learning will be on the GPU's you can go air coolers. But either one, make sure you get a cooler that's TR4 compatible, and look for ones that are specifically designed for them (Noctua for example released a new line specifically for this CPU)

[cynexit]

31 minutes ago, aeduG said:

COOLING: NZXT Kraken x62 inkl. AM4

Don't. If this is a workstation build you don't want to risk something like a leak or a failing pump. Get a good and solid air cooler, it'll have the same cooling capacity but two points of failure less. You'll have to set all fans to 80%+ anyways to keep the GPUs cool so noise isn't an issue anyway. This will also have the added benefit of more airflow around the socket so the VRMs will stay cooler.

 

 

Regarding storage: If you really want to do deep learning you'll have to deal with many TBs of data. If you intend to stream the data from a Cassandra cluster you might want to invest into a 10gb nic. If you plan to house everything locally I would highly advise to build a big RAID array or IO will be your main bottleneck. I would suggest to just skip the SSD and get three more drives, try to bump them all up to 8TB and get something like WD Red Pro or Gold.

 

Besides that try to cram as much RAM in it as you can.

 

Personally I've only used Xeons, but I heard a lot of positive feedback for threadripper so you might want to consider switching plattforms and bump up the core count for the same price.

[aeduG]

5 minutes ago, TVwazhere said:

If most of your load for deep learning will be on the GPU's you can go air coolers. But either one, make sure you get a cooler that's TR4 compatible, and look for ones that are specifically designed for them (Noctua for example released a new line specifically for this CPU)

Ok I see. Maybe this will do just fine: Noctua NH-U14S TR4-SP3

 

 

[TVwazhere]

2 minutes ago, aeduG said:

Ok I see. Maybe this will do just fine: Noctua NH-U14S TR4-SP3

A lot of the PC reviewers use that cooler because (a it came with a lot of the TR4 reviewer kits) and B because it's really good and works as well as most of the watercoolers

[aeduG]

@Mr alex, @TVwazhere, @cynexit Thank you all for the recommendations. This helped a lot and I know what to buy now.