Maximum Model Size for STM32F303ZE in Machine Learning Projects?

[shreyaschandran]

I've successfully implemented a small machine learning model on the STM32F303ZE, and it's working well so far. However, I’m curious about the maximum size of the model that this microcontroller can handle. Has anyone experimented with larger models on the STM32F303ZE? What are the limitations in terms of model size, and at what point should I consider switching to another board?

[Julian E. · Accepted solution]

Hello @shreyaschandran ,

Sorry for the late answer.

Basically, there are 2 factors that may force you to change board:

• You don't have enough RAM/Flash to use the model you want
• You consider the inference time being too big and you want to get a more powerful MCU.

I would say that you can stay on your current board as long as it is ok for you and when you encounter an issue, you can look at a board that satisfy your needs based on the model you are trying to implement.

There are multiple tools within ST that can help you that:

• In ST Model Zoo, you have a selection of models that you can download, retrain or use directly. In the tables per usecase in the readme, you can find model and their details to have an idea of performances and sizes (for example: https://github.com/STMicroelectronics/stm32ai-modelzoo/blob/main/object_detection/pretrained_models/st_ssd_mobilenet_v1/README.md)
• You can also benchmark your models with ST Edge AI Developer Cloud, find more information here: https://wiki.stmicroelectronics.cn/stm32mcu/wiki/AI:Getting_started_with_STM32Cube.AI_Developer_Cloud

Best regards,

Julian 

[shreyaschandran]

Thanks, this helps a lot.