Machine Learning Build

[Encoder]

Budget (including currency): INR 180,000

Country: India

Games, programs or workloads that it will be used for: Training Machine Learning & AI Algorithms

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): Completely new PC build. Already have a good enough Keyboard, Mouse and Monitor that serves the purpose.

 

Hello Linus Tech Tips community,

 

I am planning to build a PC to train my Machine Learning and AI algorithms and some occasional gaming. I used to train those off of my Omen Laptop (which actually does a decent job) with a i7-9750H CPU, GTX 1650Ti graphics card (Cuda cores are a important factor for me) and a measly 8GB DDR4 RAM earlier. But now I am at a point where I need some more juice out of my system. Also the thermals are a mess on heavy and sustained runs. Some of my iterations might take more than 24 hours to run.

But I do not want to pay AWS or Azure the charges they are quoting for my use case. So I have decided to build myself a dedicated PC for this purpose.

Apart from doing programming and training algorithms, I like to game occasionally. I mostly play racing games like F1, Moto GP and NFS.

 

After about 2-3 weeks of research and checking availability in my country, I have narrowed down to these Components

PCPartPicker Part List

Type	Item	Price
CPU	AMD Ryzen 7 3700X 3.6 GHz 8-Core Processor	$274.49 @ Amazon
CPU Cooler	Corsair H100i RGB PLATINUM 75 CFM Liquid CPU Cooler	$139.99 @ Best Buy
Motherboard	Asus TUF GAMING X570-PLUS (WI-FI) ATX AM4 Motherboard	$189.99 @ Best Buy
Memory	G.Skill Trident Z Neo 32 GB (2 x 16 GB) DDR4-3600 CL16 Memory	$178.99 @ Newegg
Memory	G.Skill Trident Z Neo 32 GB (2 x 16 GB) DDR4-3600 CL16 Memory	$178.99 @ Newegg
Storage	Crucial P1 500 GB M.2-2280 NVME Solid State Drive	$62.99 @ B&H
Storage	Crucial MX500 1 TB 2.5" Solid State Drive	$110.20 @ Newegg
Storage	Seagate Barracuda Compute 2 TB 3.5" 7200RPM Internal Hard Drive	$54.99 @ Newegg
Video Card	Gigabyte GeForce RTX 2070 SUPER 8 GB GAMING OC Video Card
Case	Corsair iCUE 220T RGB Airflow ATX Mid Tower Case	$99.99 @ Best Buy
Power Supply	Corsair RMx (2018) 850 W 80+ Gold Certified Fully Modular ATX Power Supply	$164.99 @ Best Buy
	Prices include shipping, taxes, rebates, and discounts
	Total	$1455.61
	Generated by PCPartPicker 2020-06-18 09:31 EDT-0400

I am also considering future upgradability in case I need a dual GPU or a more powerful CPU someday. I am also open to about +10% flexibility on my budget.

 

What do you experienced guys think? Any suggestions for any better or efficient components?

Thanks in advance for all your help. 

[Radium_Angel]

15 minutes ago, Encoder said:

Thanks in advance for all your help

What are you using (OS and programs) for your ML?

[gloop]

I'd recommend getting a d15/dark rock pro 4 for reliability ofd this is a work machine. If you do switch the cooler, you will need to change the ram sticks to an LPX kit (get 2 of these of course).

[Encoder]

37 minutes ago, Radium_Angel said:

What are you using (OS and programs) for your ML?

OS: Windows 10

Database: MySQL and MongoDB

Programming:- Python 3.7 with ML/DL libraries like Tensorflow, Keras, Scikit-Learn, XGBoost. Graping tools like Seaborn and plotly.

In addition to them I also have some custom libraries that I have developed to help out with some of my usual stuffs. However those are super inefficient comparatively speaking.

So a generally snappy CPU and RAM are basically an essential.

Also I majorly work on Image and Natural Language Processing. So VRAM is also critical.

[Encoder]

15 minutes ago, gloop said:

I'd recommend getting a d15/dark rock pro 4 for reliability ofd this is a work machine. If you do switch the cooler, you will need to change the ram sticks to an LPX kit (get 2 of these of course).

I am planning to keep this thing running 24 hours for multiple days at a time. So, will an air-cooler be sufficient for such sustained loads?

I know it is a really conflicting topic among builders. So, pardon my inquisitiveness, I am really new to this. 

[Radium_Angel]

3 hours ago, Encoder said:

Tensorflow

The workd I've done with RNN/ML under linux, really is dependent on GFX card more than CPU, so in that respects, I think you've got both covered well in your list.

[igormp]

15 hours ago, Encoder said:

 

Memory	G.Skill Trident Z Neo 32 GB (2 x 16 GB) DDR4-3600 CL16 Memory	$178.99 @ Newegg
Memory	G.Skill Trident Z Neo 32 GB (2 x 16 GB) DDR4-3600 CL16 Memory	$178.99 @ Newegg

If I were you, I'd try to go with 2x 32GB sticks. I know that corsair has some LPX ones. It'd allow you to easily upgrade to 128gb if needed.

 

Another thing is, isn't better to go for a 2060 Super then buy another one later on? It has the same 8gb of VRAM, which is a really bad bottleneck for the 2070S IMO. Also, if it's possible, I believe that you should wait for the new Ampere cards since those will have a higher number of tensor cores.

 

14 hours ago, Encoder said:

I am planning to keep this thing running 24 hours for multiple days at a time. So, will an air-cooler be sufficient for such sustained loads?

I know it is a really conflicting topic among builders. So, pardon my inquisitiveness, I am really new to this. 

IMHO, an air cooler is more reliable since in worst case scenario only the fan will fail, but the heatsink should still be able to handle things with some mild thermal throttle.

[Encoder]

On 6/19/2020 at 10:47 AM, igormp said:

If I were you, I'd try to go with 2x 32GB sticks. I know that corsair has some LPX ones. It'd allow you to easily upgrade to 128gb if needed.

I am planning on doing this. But due to this pandemic situation, as of now there is no stock of 32 gb sticks from any reputable manufacturer in my country. So I have selected 4x16Gb sticks. But the original plan is to wait for 32 Gb Sticks to come back in stock.

 

On 6/19/2020 at 10:47 AM, igormp said:

Another thing is, isn't better to go for a 2060 Super then buy another one later on? It has the same 8gb of VRAM, which is a really bad bottleneck for the 2070S IMO. Also, if it's possible, I believe that you should wait for the new Ampere cards since those will have a higher number of tensor cores.

I believe 2060 Super has lesser cuda and tensor cores as compared to 2070 Super. Also I plan on adding another 2080Ti or some other new cards having better VRAM, tensor cores and cuda cores 1 year down the line. That is why for the time being I am considering the 2070S.

 

On 6/19/2020 at 10:47 AM, igormp said:

IMHO, an air cooler is more reliable since in worst case scenario only the fan will fail, but the heatsink should still be able to handle things with some mild thermal throttle.

Oh... I didn't know that. Awesome. I'll switch to a good air cooler. Probably the Dark Rock Pro as @gloop mentioned earlier.

[Encoder]

Also what do you guys think, Should I go for the X570 or switch to a pocket friendly B550 version considering the latest release?

Will it affect my particular use case?

[igormp]

13 minutes ago, Encoder said:

Also what do you guys think, Should I go for the X570 or switch to a pocket friendly B550 version considering the latest release?

Will it affect my particular use case?

I'd say that a B550 is good enough, since it already has multi GPU support anyway. Not sure if you'd need any feature found on x570.

[koptr]

51 minutes ago, Encoder said:

I am planning on doing this. But due to this pandemic situation, as of now there is no stock of 32 gb sticks from any reputable manufacturer in my country. So I have selected 4x16Gb sticks. But the original plan is to wait for 32 Gb Sticks to come back in stock.

 

I believe 2060 Super has lesser cuda and tensor cores as compared to 2070 Super. Also I plan on adding another 2080Ti or some other new cards having better VRAM, tensor cores and cuda cores 1 year down the line. That is why for the time being I am considering the 2070S.

 

Oh... I didn't know that. Awesome. I'll switch to a good air cooler. Probably the Dark Rock Pro as @gloop mentioned earlier.

I don't know much about SLI but can you run two different GPUs in it?

[Encoder]

16 minutes ago, koptr said:

I don't know much about SLI but can you run two different GPUs in it?

I am not going to pair them in SLI.

The programs I will be using (Tensorflow and Keras) can handle the memory and process sharing between the GPUs without a SLI link between them.