https://www.st.com/content/st_com/en/campaigns/nanoedgeaistudio.html
Part4 - Development Boards:
Which development boards are supported by NanoEdge AI?
> For the Trial version: all STM32 demo boards listed in the Studio (not the generic ones) (see doc)
For the Paid version: all STM32 dev boards
What was the hardware used in the webinar?
> STEVAL STWINKT1B : https://www.st.com/en/evaluation-tools/steval-stwinkt1b.html
3 speed USB fan (https://www.amazon.fr/RATEL-Ventilateur-ventilateur-Utiliser-personnel/dp/B07R5HKSYL/)
Does It work also with stm32f429I discovery?
> Works with any STM32 dev board listed in the studio
May NanoEdge AI be used on a STM32L4 board? E.g., the B-L475E-IOT01A kit?
> Works with any STM32 dev board listed in the studio
Sensors:
What sensor do work with NanoEdge AI?
> Any sensor, but bear in mind that MCUs are resource-constrained by nature.
So HD image recognition, or video analysis may not be a good fit
How do I integrate non-ST sensors?
> Same as ST sensors
we are not limited to ST sensors in any way, the only hardware limitation is to use a ST Cortex M4 MCU
May I also monitor an external accelerometer that is not included in the board?
> Yes, any sensor is supported (combine with questions above)
If the sensor in the application is replaced due to malfunction, is it necessary to repeat the entire process or just the embedded learning phase?
> Embedded learning should only be fine, but hard to tell without further info. If not sufficient, make a new lib.
Can we use NanoEdge AI studio with BlueNRG family?
> See above
STM32:
With which STM32 MCUs does NanoEdge AI work, and does it require any specific AI HW block?
> See above. No particular AI hardware required
I have a CAN bus interface with my STM32 application to output sensor data, can I use this interface to record with?
> Yes, any data source can be used, although some inputs may require reformatting (see doc)
What is the expected CPU usage?
How many MCU performance is using for AI library & data processing?
> Impossible to tell. depends on input data size, use case type, sensor type, model selected, and 1000 other things. Typically, on a M4 @ 80Mhz with a 3-axis sensor and buffer size 256, processing speeds are under 15-20 ms.
Is it possible to log data from ADC channels?
> Any data source can be used, although some inputs may require reformatting (see doc)
Will the libraries work with custom boards using STM controllers or only with STM boards?
> Only STM32 boards, unless a pro license is bought, and royalties paid. Then any board with ST cortex M4 MCU.
Are there STM32 products particularly suitable for this application e.g., devices with AI hardware accelerators?
> NEAI library this is generic solution suitable for any STM32 Cortex-M core: M0, M0+, M3, M4, M7, M33
Is this working also on stm32L series that has low operating frequency?
> Any STM32 board, although available RAM and processing speed can be too restrictive depending on the use case
How much processor time does the anomaly algorithm take in your example?
> Impossible to tell. depends on input data size, use case type, sensor type, model selected, and 1000 other things. Typically, on a M4 @ 80Mhz with a 3-axis sensor and buffer size 256, processing speeds are under 15-20 ms.
Is the learned model saved in non-volatile memory?
> Not by default, but it is easy to achieve
Can you save the learning parameters to flash, so that you can load them on reboot?
> Yes, it must be implemented by the user (combine with question above)
What is the effect of reduced ram amount?
> It restricts the search space during benchmark in the Studio, so it is possible that a more suitable library (better performances) is excluded because of ram constraints
Also, the amount of ram available effectively limits the size of input signals that can be acquired (this has nothing to do with NEAI though).
