I just received a MAX78000FTHR which is Maxim's new low-power artificial intelligence development board. It's a pretty feature packed board in the Adafruit Feather form factor for $30 https://www.newark.com/maxim-integrated-products/max78000fthr/dev-board-32bit-arm-cortex-m4f/dp/09AJ3133?st=max78000 .
I'm really interested in how well the low power CNN accelerator works and seeing how low power it actually is. The board has video and audio capability with a VGA camera and a digital microphone - it's incredible what they have on this board at such a low cost.
The board comes pre-programmed with a keyword spotting demo that you can try out-of-the-box.
Quick Start
Apply power to the MAX78000FTHR using the USB cable. The pre-programmed 'Audio Keyword Spotting' demo will begin to execute. The RGB LED (D2) will turn on green, indicating that the demo is running. The on-board microphone starts listening for the keyword GO. When the keyword GO is detected, RGB LED (D2) will turn on yellow. In this mode, when one of nine keywords is detected, the RGB LED (D1) starts to blink blue one to nine times based on the number detected by the convolutional neural network. The STOP command exits number keyword detection, and the RGB LED (D2) turns on green again, and RGB LED (D1) turns off.
The keyword spotting worked pretty well, except it had a problem with the number four the first time that I said it. It classified ten as unknown because that was not part of its training set. I'm impressed that it did so well considering that it did not train with my voice and the onboard hardware.
This board certainly looks promising. It will be interesting to try out the camera with video classification. And maybe try some power profiling.
Model training
The primary issue that I can see for me in using this board is the model training. I need to research it more, but currently there only appears to be local based training (on your own hardware). And it appears that other than simple models, this only works well with Linux based CUDA hardware which I don't have. Here is a link to their training repository: https://github.com/MaximIntegratedAI/ai8x-training .
And a brief excerpt:
I guess I've been spoiled by using Cloud based training using Edge Impulse and SensiML. Hopefully, something like that will be available for the MAX78000 in the future.
For now, I guess I'll play with the hardware using the existing examples and maybe try a few simple models...
Top Comments