Edge Impulse brings TinyML to millions of Arduino developers

Running machine learning (ML) models on microcontrollers is one of the most exciting developments of the past years, allowing small battery-powered devices to detect complex motions, recognize sounds, or find anomalies in sensor data. To make building and deploying these models accessible to every embedded developer we’re launching first class support for the Arduino Nano 33 BLE Sense and other 32-bit Arduino boards in Edge Impulse.

This is a companion discussion topic for the original entry at https://www.edgeimpulse.com/blog/edge-impulse-brings-ml-to-arduino/


is there some research paper that we can cite when we are using your tool for ML on IoT in research papers?

Furthermore, I’m asking if I can use the accuracy, memory and latency datas obtained running some different configurations on edge impulse for some papers on ML on IoT…

Thank you in advance for the awesome tool


Hi @wallax, not any papers to reference, but yes, you can use the measurements freely!

And if you have interesting research we’re always happy to publish on our channels too :slight_smile:

ok thanks,

I’ll write something, of course I will provide you referements if the articles will be accepted!



Hi everyone,

Just out of curiosity, are there other future boards which are planned to be added to Edge Impulse? I’m particularly interested in the Kendryte K210 chip as it seems to provide a good amount of processing power and built-in signal processing capabilities.

Keep up the great work!

Hi @Martin_08, preferably not too many future boards as the data forwarder and the C++ library export should allow you to target virtually any platform under the sun.

Naturally we will continue to add targets, but this is always done together with the silicon / hw vendor, and should bring something unique to the ecosystem that’s hard to do without full end-to-end integration of a board - e.g. work we’re doing with Nordic right now to add sampling over BLE and do over-the-air updates to your board straight from the Studio.

Side note, major pain point with RISC-V targets right now is the lack of something akin to CMSIS-DSP/CMSIS-NN (like in the Arm world) which has optimized code already for typical DSP/ML operations like FFTs and vector operations - so lot of custom work for these boards to make them run fast. But you can just run the C++ Library export there now - performance won’t be the greatest, but the chip seems plenty fast already. And I think for vision models you can always take the TFLite model from Dashboard and it looks like Kendryte has a way of pushing that into the target right now.