Edge computing is here! You've probably heard of this latest entry to the long lineage of tech buzzwords like "IoT," "LoRa," and "cloud" before it, but what is “the edge” and why does it matter? The cloud is impressively powerful but all-the-time connection requires power and connectivity that may not be available. Edge computing handles discrete tasks such as determining if someone said "yes" and responds accordingly. The audio analysis is done at the edge rather than on the web. This dramatically reduces costs and complexity while limiting potential data privacy leaks.
In collaboration with Google and Ambiq, SparkFun's Edge Development Board is based around the newest edge technology and is perfect for getting your feet wet with voice and even gesture recognition without relying on the distant services of other companies. The truly special feature is in the utilization of Ambiq Micro's latest Apollo3 Blue microcontroller, whose ultra-efficient ARM Cortex-M4F 48MHz (with 96MHz burst mode) processor, is spec’d to run TensorFlow Lite using only 6uA/MHz. The SparkFun Edge board currently measures ~1.6mA@3V and 48MHz and can run solely on a CR2032 coin cell battery for up to 10 days. Apollo3 Blue sports all the cutting edge features expected of modern microcontrollers including six configurable I2C/SPI masters, two UARTs, one I2C/SPI slave, a 15-channel 14-bit ADC, and a dedicated Bluetooth processor that supports BLE5. On top of all that the Apollo3 Blue has 1MB of flash and 384KB of SRAM memory - plenty for the vast majority of applications.
On the Edge you'll have built-in access to sensors, Bluetooth, I2C expansion, and GPIO inputs/outputs. To support edge computing cases like voice recognition the Edge board features two MEMS microphones, an ST LIS2DH12 3-axis accelerometer on its own I2C bus, and a connector to interface to an OV7670 camera (sold separately & functionality coming soon). As TensorFlow updates their algorithms more and more features will open up for the SparkFun Edge. An onboard Bluetooth antenna gives the Edge out-of-the-box connectivity. Also available on the board is a Qwiic connector to add I2C sensors/devices, four LEDs, and four GPIO pins. To boast the low-power capabilities of the board we've outfitted it with battery operation from the CR2032 coin cell holder. Programming the board is taken care of with an external USB-serial adapter like the Serial Basic Breakout via a serial bootloader, but for more advanced users we've also made available the JTAG programming and debugger port.
As a brave explorer of this new technology, you'll have to use some advanced concepts but don't worry. Between Ambiq Micro's Software Development Kit and our SDK Setup Guide you'll have access to plenty of examples to begin working with your hardware.
Now get out there and make something amazing with the latest machine learning technology at your very own fingertips!
Note: The HM01B0 Himax Camera is NOT included with the SparkFun Edge and will need to be purchased separately.
What It Does
If a board needs code or communicates somehow, you're going to need to know how to program or interface with it. The programming skill is all about communication and code.
Skill Level: Experienced - You will require a firm understanding of programming, the programming toolchain, and may have to make decisions on programming software or language. You may need to decipher a proprietary or specialized communication protocol. A logic analyzer might be necessary.
See all skill levels
If it requires power, you need to know how much, what all the pins do, and how to hook it up. You may need to reference datasheets, schematics, and know the ins and outs of electronics.
Skill Level: Rookie - You may be required to know a bit more about the component, such as orientation, or how to hook it up, in addition to power requirements. You will need to understand polarized components.
See all skill levels
Based on 19 ratings:
3 of 3 found this helpful:
Well, demo sketch on the board only works 5% of the time. Kinda disappointing.
3 of 3 found this helpful:
They dropped the EDGE 2 version shortly after I bought without any notice or explanation. Didn't work anyway. Very disappointed in product and service
By contrast the Adafruit PyBadge is well supported and works. Can use Arduino libs to access Tensorflow Lite and Google says Circuit Python support is coming. And the underlying processor is significantly faster. Costs more than double -$34 but Adafruit is clearly committed to ML so a much better place to invest your money
4 of 4 found this helpful:
The demo of the Voice regognition is not working and no help from the forum.
Without Tensorflow feature working, this board has no interest
I have done some debugging and I think I have found a workaround. Waiting for confirmation. more details here:
1 of 1 found this helpful:
Poorly maintained documentation and out of sync repo / library versions! I originally bought the board back in Nov. 2019 and it took quite a bit of messing with to get it to program properly because the documentation did not match the method that eventually worked, I figured maybe because I had an older board so I bought another one in Jan. 2020 since it states in the documentation that it is supposed to come with the SVL bootloader but even this newer one could only upload using the Ambiq secure bootloader setting. After getting uploading working you quickly find that the examples and the dependency libraries don't work well together, you have to try different versions until you find a combination that might work but more than likely you have to go in and edit a bunch of code to fix the issues, you spend so much time just trying to get something to work just to find out that once you finally get it working the examples are pretty crappy and not even worth the time. Total waste of time and money!!
2 of 2 found this helpful:
Great price. The promise is nice. But when i put battery in it is difficult to make blink the leds. And no much doc about how to use it. I'd like to use Bluetooth but nothing about that.
2 of 2 found this helpful:
I tried the default program (yes, no) and it worked about 3 times out of 200. I even tried building it from scratch via the tensorflow sites tutorial. This product is essentially unusable in its current state.
3 of 3 found this helpful:
Preloaded demo doesn't work well... Getting started is basically impossible due to poor documentation at the moment. Needs a lot of TLC to become a product you can actually ask money for. Very disappointed.
4 of 5 found this helpful:
Overall I think this Edge board is good, but "Steep Learning Curve" is the essence of my experience. I decided to really invest the time and learn how to use this board and the capabilities of this nifty Ambiq Apollo3 Blue core & now that I've worked through most of my issues and can easily and reliably build & boatload Ambiq, SparkFun, and TensorFlow Lite for MCU examples to the board, I'm starting to see the promise for this board in all sorts of small low-power, wirelessly connected projects. I think the BLE capabilities are pretty compelling & It really is a nice little product once you get through the learning pains (which I think most people will just give up on, unfortunately).
However, this board is not for the faint of heart, I have spent weeks (yes, weeks!) getting to this point where I'm a bit more comfortable with it and modifying example projects to do my own thing with it.
A couple of notes:
Built-in "Yes/No" speech demo does not work great. Same results when building directly from the TensorFlow tutorial from source-code. I'm not sure if this is a imitation of the MEMS mics, the limitation of the inference model that can be implemented in the Apollo3, or PWB/manufacturing quality;
I had a similar experience with the TensorFlow Magic Wand demo ... buggy and unreliable results (and poorly documented). This leads me to think it is not a great choice for Machine Learning inference ... but I'm really still investigating this.
The documentation and troubleshooting for this board needs lots of work, lots of references and old tool-chains, etc. It took me a while to eventually find the "LTS" Github repo, which is much cleaned up ... but I found this quite by accident.
I had lots of problems initially getting the serial UART boot loader working. The documentation focuses on the CH340 series boards & this caused my numerous headaches (and almost threw out my SF Edge and CH340 board and quit!). Numerous driver problems in MacOSX & Linux (Ubu 18.04LTS), tried compiling from the sources, etc, etc. So frustrating! I eventually found an FTDI-based USB UART breakouts (with 3.3v/5v selection) that works great! and this solved lots of my headaches. However, I still don't think the button bootload process was explained clearly in the original SF tutorial, and endless searches through google, ending up on random forums, trying different baud rates, etc. ... so again - lot's of frustration until I figured out how it works & it is so seamless now.
I wish there was a good bare-metal BLE example in your repo, instead of one based on FreeRTOS. So, now I have to learn a bit about FreeRTOS too (which may not be a bad thing). There may be a good reason for this (task scheduling is crucial to get BLE to work?), but as I'm still new to BLE too, I just don't know & I couldn't find any docs that really break this down for you other than just trying to start from one of the example projects (which I'm glad you guys provide!)
Please update your tutorial (on your learn section for this product) and please steer people to better documentation and support & AWAY from the CH340 and towards a suitable FTDI product instead!
2 of 3 found this helpful:
So I was sitting at my desk, repeatedly saying "yes" in the vain hope that the yellow LED would blink (it did, 3 times out of 50) when my wife came in looking very puzzled and wondering just what I was up to. So I told her and gave her the board so that she could try it. With 1 success out of 20 tries, she gave this review its title.
If this is as good as it gets with a professionally developed and supplied application, then we should plan on ignoring the deep learning features and just use it as a low power Arduino replacement.
1 of 3 found this helpful:
This chip is literally the future of AI, it lets you run vocal inference on a chip which isn't even that expensive. Amazing!
The SparkFun Edge Micro Speech recognition demo works after adjusting the sensitivity of the app. The SparkFun demo by Nathan Seidle and Pete Warden showed terrible results for recognizing "yes" and "no".
After changing the call (in micro_speech/main_functions.cc around line 125) to:
static RecognizeCommands static_recognizer(error_reporter, 1000, 100, 500, 2);
I now get very good results with "yes" and "no".
The demo code is from https://github.com/tensorflow/tflite-micro
I highly recommend getting the O'Reilly book TinyML (by Pete Warden) that goes over the installation and development of code for ML on the SparkFun Edge.
The Apollo3 chip is going to be worth the time investment to understand and master for ML at the low power micro level. The SparkFun Edge is a nice board to get started with this effort but could use a few more GPIO pins.
SparkFun should update the demo app with the code change above to make the demo work at an acceptable level.
The ML has lots of room to be improved. Maybe there is bugs in the firmware?
I will work it out the bugs and upload my own model to it, but if you don't have the skills to read the code and create your own Tensorflow lite model, then I don't think this device is for you.
Great little innovation, but functionality? Near zero, as many other customers describe. What annoys me is also that you have to take the battery out. I'm just not playing with toys all day, so ON/OFF switch is just basics like water.
Ran a workshop with these and they were nothing but trouble. The "reset/button 14, let go of reset" process is silly and shouldn't be necessary.
JTAG connector's pin assignment is inverted !!! Please do NOT implement the pin header on this board when you use J-link debugger.
Thanks for the review, but the connector is used from the other side of the board. If you flip the board over, everything is in the correct order. :-)
This is a really great tool for using embedded machine learning. The stock speech recognition demo works as advertised, and is tremendous fun to play with. It's very capable hardware.
I think this is an abandoned product. Bought it with a camera. So far, I haven't been able to make it work. To upload the sketch hold 14 down.
Had some startup issues with the battery holder not making contact on two boards, resolved by multiple insertions/removals. Initial operation has the yes/no accuracy and sensitivity are pretty low. Have not explored the toolchain yet, that sounds like a very interesting area to poke at.