This is a headless box running Ubuntu 16.04 that was built to train Deep Learning model architectures, e.g., in TensorFlow and Torch. I based my build on this one by FourMoBro: https://pcpartpicker.com/b/NW4qqs
This was my first build so the large, two-partition case made assembly much easier. It further enabled us to tidy away the cables in the back partition alongside the power supply, creating a pleasing aesthetic in the windowed, front partition.
Another component that was tremendously helpful given that it was my first build was the motherboard. It provides feedback via multiple LEDs that allowed us to debug rapidly and precisely.
Our models are memory-intensive so I filled all eight RAM slots with 16GB cards for a total of 128 GB).
For storage, a 512 GB SSD was sufficient for our day-to-day modelling purposes. I bought a cheap 3TB HDD for local storage of larger data sets that we on occasion (say, monthly) sample from.
The estimated maximum wattage is 882W so I added a 100W buffer to be safe, purchasing a 1000W power supply.
To read more detail on our experience putting the box together, our summer intern, Max, published a blog post here: https://insights.untapt.com/building-the-magical-data-science-box-e933656d625e