HackSpace – September 2019

FIELD TEST

VERDICT

For low-

power AI with

Bluetooth

connectivity,

the hardware’s

here, but you’ll

need to see

if there are

the models

you need.

8 / 10

This can make sense, particularly if you’re

running a number of devices in a small area

and power consumption is an issue, or if your

devices only need occasional connectivity.

You need to be realistic with your expectations

with the SparkFun Edge. If you’re looking for

high-precision matching of complex input – such

as recognising people from images – then you’re

probably going to struggle. However, if you’re

looking for something to run on very low power and

react when it recognises one of a small number

of conditions, then this might well be the board

for you. At the moment, the lack of pre-prepared

models means that it’s not really suitable for casual

uses; however, given that it’s an official board by

TensorFlow, we expect that there’s likely to be

more in the future, so keep an eye out to see what

models currently exist before making a purchase.

The hardware on this board is very good

for a very specific set of circumstances – AI

without mains power and needing Bluetooth

connectivity. While these are quite specific,

they’re not all that rare. This lets you bolt on

neural networks to remote sensor deployments.

For those conditions, this board stands alone.

This refers only to programming the device.
Programs take models (which are the matching
engines of neural networks) and tell the board
what to do depending on how the model reacts
to the input. Creating models is a separate
task from programming – it exists somewhere
between art and science. Essentially, it comes
down to throwing a lot of example data at a neural
network and attempting to ‘train’ it to understand
particular patterns. In many cases, it’s possible to
bypass this, and use pre-created models similar
to how you might use a library in programming.
Once you’ve got your code into the device, you
need a way to interact with it. This board is designed
by SparkFun, and the team at TensorFlow, to run
neural networks. These models are first trained to
recognise if an input falls into a particular category
or not; for example, if a particular sound is a word.
On the Edge, there are two microphones, four GPIO
pins (that can run SPI or I^2 C), a Qwiic connector,
and a camera connector. As yet, there’s no camera
module available for the camera connector.
This is all controlled by an Ambiq ARM Cortex
M4F processor with Bluetooth. The Cortex-M4F core
is one of the more powerful microcontrollers around,
and this one runs at 48MHz (with a 96MHz burst
mode). What makes this stand out is that it does this
while drawing under 2 mA of current at 3.3 V – great
if you’re running on battery (there’s a CR2023 holder
on the back), solar, or other limited power supply.

CONNECTION CONUNDRUMS
Perhaps the biggest limitation for this board as an
edge device is the connectivity. Bluetooth makes
sense from a power and cost perspective, but it
makes your overall system setup a bit more complex
as it’ll need something to pair with to send data into
the world.

The hardware is very good

for a very specific set of

circumstances – AI without

mains power and needing

Bluetooth connectivity

”

”

Above

There’s a battery

holder for

off-grid running

down to throwing a lot of example data at a neural
network and attempting to ‘train’ it to understand
particular patterns. In many cases, it’s possible to

Above