microcontroller port feasibility

[street-grease-coder]

is feasible to port this to an esp32 or similar (e.g. micropython) microcontroller environment?

[stubbb]

I would not ever attempt to use this on anything below 2 GB of RAM.

[street-grease-coder]

would it feasible in your opinion to port it to e.g. int8 format (if I read correctly Tensorflow lite has some functionality built-in to allow easy transformation of weight matrices), so it requires only a fraction of RAM, with minimal performance losses?

[stubbb]

Well, I reach 8GB of RAM to get that 10 fps at the accuracy I need while smoking 30 Watts at the most power efficient AI board on the market. Even if it was technologically possible, you would end up with a very high processing time and very low accuracy.

The processing would likely be x100 longer at 2 Watts, and if you get down from 8 GB of RAM to, say 256 MB sequential, you are looking at another x100 processing time increase. Thats 1 frame every 100 seconds if it is even technologically possible.