What commercial MCU should I choose to achieve on-device learning with MCU?

[haoliu001]

Dear sir or madam,

Do you have any suggestions what MCU should we use to achieve on-device learning? If I directly use NUCLEO-N657X0-Q, is that would work?

Do you have any manual?

Thanks in advance

cc @JacobSzwejbka

[lucylq]

Hi @haoliu001, thanks for your interest!

You might like to check out https://pytorch.org/executorch/stable/executorch-arm-delegate-tutorial.html

cc @digantdesai

[digantdesai]

Thanks @lucylq.

STM32N657X0 has Cortex-M55. @lucylq pointed to the end-to-end tutorial which let's you run simple models using ExecuTorch on Coretex-M55 using Arm FVP Simulator, without any NPU.

That said, when you say on-device learning do you mean training on Cortex-M55?

I would say, ExecuTorch may already support capturing, serializing, and running backwards graph with updating weight. If that is true, @JacobSzwejbka can confirm, then we should be able to do the same on Cortex-M55, given we can run backward() on a network with all ops in fp32. I haven't tried anything close to this but curious to know about your plans.

[JacobSzwejbka]

ET Training currently does not have robust support for backend acceleration/ delegation. @digantdesai would xnnpack run on this cortex? What about the portable ops?

I don't have great clarity on when we want to use the ARM delegate for instance.

[JacobSzwejbka]

@haoliu001 you can check out the readme in extension/training for a walkthrough of ET's current training capabilities.

[digantdesai]

XNNPACK, at least today, will not run on Cortex-M. Portable-ops will.