Building a Deep
Learning (Dream)
Machine
DIY hardware hacking for the uninitiated
Roelof Pieters
@graphific http://www.graph-technologies.com
27 June 2016
http://graphific.github.io/
How: Machine Thank you gamers!
Why
- Cheaper *
- Can do things you can’t do in the cloud
- Desktop = faster experiment iteration
- Remote still possible
- Customize to your own needs
- Fun!
* in my own particular case
Points to think about
- How many GPUs now and later
- Motherboard (40 lanes / 16x8x8x8 configuration for 4 GPUs)
- Chassis with enough space + air flow
- 4 GPUs: 7 PCIe slots (last GPU can be mounted at the bottom using only one
slot)
- CPU: good enough, as much cores as GPUs, make sure CPU supports 40
PCIe lanes, some new Haswell CPUs only support 32;
- RAM: 2x of total GPU memory
- SSD = nice if data doesn’t fit into GPU+RAM (or hdfs reads from disk): get
larger SSD than your largest dataset
- Mechanical disks: plenty of storage
- PSU: don’t save on efficiency (titanium/platinum)
- Cooling: Water (hard diy) or Fan (easy/cheap but noisy)
I: HARDWARE
How: Machine
How: Machine
How: Machine (2)
II: SOFTWARE
https://github.
com/deeplearningpari
s/dl-machine
And so on...
http://gitxiv.com/
http://www.creativeai.net/
GTX 1080
Drive PX 2
P100
EPGU
PGU
READ MORE
http://graphific.github.io/
https:
//hackaday.
io/project/1207
0-32-tflop-
deep-learning-
gpu-box
http://timdettmers.
com/2015/03/09/dee
p-learning-
hardware-guide/
https://timdettmers.
wordpress.
com/2014/08/14/whi
ch-gpu-for-deep-
learning/

Building a Deep Learning (Dream) Machine