Building a Deep Learning Computer

Source: Deep Learning on Medium

Building a Deep Learning Machine

Why should you do it?

  • Better understanding of the end to end training process.
  • Pays for itself. Assuming that you train for 240 hours a month, renting an amazon p3 instance($3/hour) would cost you 240 * 3 = $720. Going by this estimate you will break even in 4 months.
p3 instance prices as of 11/07/19
  • Do not have to remember to turn off you instance once you are done with the training and save your work.
  • Ability to take advantage of price drops on components.
  • It is fun understanding how different components fit together.
  • Of the shelf GPU desktop builds are available but are a more expensive and come with pre-installed software and have limited scope for expansion.

Where do you start?

  • Reach out to friends or social media who have previously built their own machines
  • Read online blog posts (I have included links in the Useful Resources section)
  • Subscribe to Youtube channels (I have included links in the Useful Resources section)

How do you choose your components?

Here are the main building blocks for a desktop are –

  • Case
  • Power supply
  • Motherboard
  • CPU
  • Memory
  • Storage
  • GPU
  • Cooler

Depending on your training needs, you can start off with a single GPU. Choose the motherboard and case such that there is room to add additional GPUs if you start working with bigger models and data sets.

As far as CPUs go, you have 2 companies AMD and Intel. In the past 1 year AMD has been coming up with better CPUs and hence I chose AMD. Since GPUs produce a lot of heat while running it is important to have a cooler. I chose the conventional fan cooler over the liquid cooler as I felt that this would require less maintenance.

Most hardware websites allow you to compare different components and builds and it has been very helpful for me in choosing the components.

Here are the specs for my desktop build –

+------------------+-----------------------------------------------+
| Component | Description |
+------------------+-----------------------------------------------+
| GPU | Nvidia EVGA GeForce RTX 2080 Ti |
| CPU | AMD 2nd Gen Ryzen 2920X 12-Core 24-Thread |
| RAM | G.SKILL Ripjaws V Series 32 GB DDR4 SDRAM |
| Motherboard | ASRock X399 Taichi |
| Power | Seasonic PRIME Ultra 850W 80+ Titanium |
| Storage | WD Black 500GB SSD |
| Case | Fractal Design Mesify - C |
| CPU Cooler | Noctua NH-U14S |
| Thermal Compound | Arctic silver high density thermal compound |
+------------------+-----------------------------------------------+

I followed the steps in this video to assemble all the components. Once you have the hardware setup and you are able to get to the BIOS screen of the motherboard your next step is to install software which I go over in the next blog post.

Final hardware build

Useful Resources