Building your own Deep Learning Workstation

Hardware

I live in Germany. All my parts were purchased from Mindfactory. Building a PC for the first time can be overwhelming and a daunting task. Spoiler Alert: It might be easier than you think.

Parts	Company	Price (Euros)
Mother Board	Gigabyte Z390 AORUS PRO	170.81
Processor	Intel Core i9-9900 3.1 GHz 8-Core Processor	346.89
Cooler	be quiet! Dark Rock Pro 4 50.5 CFM CPU Cooler	67.22
PSU	be quiet! Straight Power 11 750 W 80+ Gold	115.85
RAM	G.Skill Aegis 16 GB (4 x 8 GB) DDR4-3000 CL16	97.5
Storage	Crucial P1 1 TB M.2-2280 NVME Solid State Drive	94.00
Case	be quiet! Silent Base 601 ATX Mid Tower Case	106.68
Gold Service	Mindfactory	10
GPU	8GB Gigabyte GeForce RTX 3070	649
Grand Total		1647.95

Motivation

I wanted to build a PC that could be used for Machine Learning. I also wanted to have complete control over hardware for upgradability and predictable hardware performance. That’s what I want you to believe, the truth is that I want to play computer games ;).

Basic setup

Use PC PartPicker. This helps you to choose your parts wisely, you will be warned that you are building a PC with incompatible parts.

Motherboard

Choose your motherboard first. Make sure that it fits your requirements (You will need to decide if you want to go for an AMD or an Intel Motherboard). I needed a motherboard with at least 4 RAM slots and at least 2 PCIe lanes for GPUs.

Just buy any decent motherboard in the market. There are a lot of companies that sell motherboards that can be overclocked. You might not need it for deep learning.

Deep Learning on an overclocked machine does not lead to performance improvements.

CPU

Intel CPU i9 comes with hyperthreading. (Basically two threads per CPU core instead of one per core). This is useful if you want to process data in parallel. I also got a great deal on Intel CPU because I know someone working there.

RAM

I chose 16 GB RAM to begin with. I intend to expand to 32GB later. That’s the reason why I needed a Motherboard with 4 RAM slots. In case you intend to get 24GB GPU (RTX 3090) make sure you have at least 32GB RAM.

Just buy the cheapest DDR4 RAM available out there with a clock frequency greater than 2400.

Cooling

You need to choose if you need Liquid Cool or Air Cool. I would recommend that you Air cool your PC. Some Advantages:

They do not leak
Easier to upgrade
Much Cheaper

I needed a single GPU for my needs (games / deep learning experiments). At worst, I would use 2 GPUs.

NVMe

I needed hot storage for reading data quickly. NVMe SSDs are perfect for this. Some motherboards do not have this option. I went for the cheaper NVMe from Crucial instead of Samsung. There is a minor difference in performance according to the benchmark tests.

GPU

Right now there is a shortage of GPUs in the Market. The new RTX 30xx series is cheaper and faster than the RTX 20xx series.

I am currently testing 8GB Gigabyte GeForce RTX 3070. It has 3 fans which keeps the GPU temperature manageable. Games run flawlessly. I have no complaints related to performance of my ML experiments.

Make sure you install your GPU in the first PCIe slot to get the best performance.

Power Supply

PC PartPicker will help you with that based on the configuration you have chosen. Suppose my configuration required 600W. I would keep an extra headroom for another 100-150W (Making it a total of 750W).

Storage / Cabinet and Other Parts

Make sure you already have an External Monitor, Mouse, Keyboard. You might also need to buy a good power strip for powering up your monitors and PSU. Keep screwdrivers handy.

Assembling

I would first check if all parts are working before installing the part to the Cabinet. This way you can check for defects and order replacements quickly if required.

Motherboard lies on its box
Install RAM
Install NVMe
Install CPU
Install Cooler
Install GPU
Connect PSU to the Motherboard
Connect Cabinet Power to the Motherboard
Connect Keyboard, Monitor to the Motherboard

Once you have connected everything so far, switch on the PSU and press the power button. If things are set up correctly, you will see a BIOS screen and Congratulations!. Now all you need to do is to move everything into the cabinet. Be gentle / accurate while plugging in different components, Try not bending the pin(s) like I did.

Final Thoughts

If you are doing this for the first time make sure you have a friend who has already done this. Be prepared to spend at least half a day if you have everything right. You can save yourself the pain of assembling by requesting Mindfactory or your local PC store will do it for you (Approx 100-200€). I also sold my Macbook Air for 262€.