Want to build an affordable deep learning box and get all the required software installed? Read on for a proper overview.
By Hui Han Chin, DSO National Laboratories, Singapore.
Deep learning is a technique used to solve complex problems such as natural language processing and image recognition. We are now able to solve these computational problems quickly, thanks to a component called the Graphics Processing Unit (GPU). Originally used to generate high-resolution computer images at fast speeds, the GPU’s computational efficiency makes it ideal for executing deep learning algorithms. Analysis which used to take weeks can now be completed in a few days.
While all modern computers have a GPU, not all GPUs can be programmed for deep learning. For those who do not have a deep learning-enabled GPU, this post provides a step-by-step layman’s tutorial on building your own deep learning box. Our deep learning box is essentially another computer equipped with a deep learning-enabled GPU. The GPU is the main difference between a regular computer and a deep learning box.
First, check if your GPU is listed on this site. If it is, it means you already have a GPU capable of deep learning, and you can start from the second section below on software installation. If your computer does not have a suitable GPU, read on to find out about how you can get the required components for less than $1.5k.
Purchase Hardware
This section lists the main components of your deep learning box. If you have no computing knowledge or you want to save time, you may choose to purchase these in-store and pay a nominal fee for assistance to assemble the components. Prices listed below are in USD as of March 2016. The total cost was $1285, which we funded using prize money from competitions.
Motherboard
Distributes power to the rest of the components and enables communication between them. It is a printed circuit board.
Central Processing Unit (CPU)
Performs general computational operations, like a brain. It is a chip on the motherboard.
$410 (bundle) – Asus B150i Pro Gaming ITX (Motherboard) + Intel i5 6600k (CPU)
Random-Access Memory (RAM)
Enables quick retrieval of information, like working memory. It is another fixture on the motherboard.
$115 – 16GB (2x8GB), Corair Vengeance DDR4 DRAM 2400MHz
Graphics Processing Unit (GPU)
Performs intensive computations for deep learning. Our chosen GPU, the GTX 970, has a small form factor so that we could fit it into a portable case. An older GPU, the GTX 770, was able to train a deep learning model with 5120 images on 20 iterations within 33 seconds. Hence, our machine is expected to perform even faster.
$415 – 4GB, Gigabyte GTX970 ITX GDDR5
Power Supply Unit (PSU)
Converts electricity from the mains to power the machine.
$75 – 550W, Cooler Master G550M
Hard Disk Drive (HDD)
Enables lasting storage of information, like long-term memory.
$160 – 4TB, Western Digital Blue 5400 RPM
CPU Cooler
Fans the CPU to prevent it from overheating.
$35 – Cooler Master GeminII M4
Case
Protects and holds all the components. Make sure that it has a USB port.
$50 – Cooler Master Elite 110
Assembly
$25 – in-store service to secure and wire the components.
After the box has been assembled, connect it to a standalone monitor, keyboard and mouse.