TensorFlow can now run on £12 edge hardware

"Cortex M4 processor is extremely low power, using less than 1 MW in many cases"

Talk about low standards. Data centers run into megawatts but I expect more power efficiency from embedded processors. (Why do articles so frequently mix up milli (m) and mega (M) like this? I get that it's just capitalization but they teach this in science classes, right?)

"the model takes up only 20KB of flash storage space, the footprint of the TensorFlow Lite code just 25KB, and the app itself only 30KB of RAM."

Those guys on google and outside google surely worked hard to achieve this level of benchmark!

Would be interested to hear how the most recent TFlite stacks up against ArmNN running the same model. ArmNN will run TF, TFlite, or caffe models and is built (by Arm) for speed on Arm hardware

Actually you don't need the $15 Sparkfun Edge as STM32F103 boards can also be used [0], and you can have them for $3 (probably even less).

[0] https://github.com/google/stm32_bare_lib

If you're interested in bringing deep learning to edge hardware, I'd suggest looking at XNOR nets:

https://allenai.org/plato/xnornet/

Here's Pete demonstrating / talking about it too: https://youtu.be/bDZ2q6OktQI?t=6922

The dev board: https://www.sparkfun.com/products/15170

How long would it take one of these boards to run e.g. Google's Inception v4 network?

In other words, how would this hardware compare in these benchmarks: https://www.tensorflow.org/lite/performance/benchmarks

Title should really say TensorFlow Lite.

Training a model on £12 hardware would be really something.