If you want to get to grips with Deep Learning at the Edge, NVIDIA's Jetson Nano is an excellent choice.
It combines a powerful, fast Quad Core Arm processor with a powerful NVIDIA GPU capable of just under 0.5 Teraflops.
Better yet, it has 4 GigaBytes of RAM. That's enough to do serious work with TensorFlow or PyTorch, both of which are supported on the platform.
And it costs around $100 in the USA, or around £120 with shipping in the UK!
There are instructions for setup on the NVIDIA Jetson Nano website, but I found them a bit daunting, and they miss out on a couple of bits of advice I wish I'd had when I started.
Before you startThe first time you boot your Nano you'll want
- A micro-SD card. I recommend at least 32 GB; 64 GB would be better.
- An HDMI monitor. It must be HDMI. You cannot use a DVI monitor with a DVI-to-HDMI adaptor of the sort you might use with a Raspberry Pi or a workstation.
- A USB mouse.
- A USB keyboard.
- An ethernet cable.
- A suitable 5V power supply.
A 5V 2.5A supply with a micro USB connector will work. A 4A supply with a 5.5mm outer diameter (OD) x 2.1 mm inner diameter connector is better, for reasons I've set out below.
If you decide to go with a 4A supply you'll also need a Jumper.
Out of the box your Nano will use 2A of current. However, your mouse and keyboard may take the current draw over 2A. If the voltage drops your Nano will shut itself down. This happened to me on my first attempt, and it's very disconcerting.
A 2.5A supply should avoid that problem, but there's another.
The Nano has two power modes. The default mode only uses 2 amps but this restricts the CPU to single core operation. The max power configuration will give you full access to the Nano's processing power, but the Nano will then draw more than 2A of current.
If you decide to stick with a 2.5A power supply and you're based in North America you can use this one.
Alternatively you can use the official Raspberry Pi power supply for the Pi model 3B, since that also provides 2.5A.
I've recommended two 4A supplies below. One comes from North America, and the other is available from a UK supplier.
- In North America: Adafruit 5V 4A Switching PSU
- In the UK: 5V 4A 4000mA AC-DC Switching Adaptor Power Supply
If you're going to power the Nano with a 4A supply you'll also need a jumper to tell the Nano to power itself from its power Jack socket instead of the micro-USB socket.
The NVIDIA website has details, but I suggest you visit the excellent JetsonHacks website which has more information on power options and shows you just where to install the Jumper.
Prepare the SD card
You'll need to download the latest NVIDIA Jetson Nano image NVIDIA Jetson Nano image and write it onto the SDS card using an SD card writer and appropriate software on your computer.
If you've prepared an SD card for a Raspberry Pi the process is very similar.
The way you do it depends on the Operating System you're using. The NVIDIA website provides instructions for WIndows, MAC and Linux, and theyr'e pretty good.
Tomorrow I'll go through the remainder of the setup process and describe how to train your first Deep Learning model on the Nano. Later I'll be posting more details of my Nano experiments.
To stay up to date, you can follow me (@rareblog) on Twitter.