All very good questions. Which really have to be answered by real world testing.
On the subject of performance, there are really two points - Microcontrollers are now a lot more powerful than they were, although clock speed isn't everything, an esp32s3 runs at 230Mhz and the RP2350(Rp Pico 2) can easily be clocked past 300 and have two cores.
- Also, MicroPython now supports two extra code emitters, Native and Viper, which create native code. They are already used on the spi display driver nodes and layout engine.
So I'm not sure it's as much a toy are you might think it is. OTOH, if this gets take up, I'd look at a lower level language.
Also, with AI somewhat more reliable now, making C/C++ addins is a much more reasonable proposition. Let me get this one tested and being used by folks and I'll think about how to create high speed paths :-)
At the end of the day, I'm trying to aim for use cases that aren't trying to squeeze the last drop of performance, but not toy use cases either.
These questions have been niggling me all weekend, and I've been running a number of benchmarks to answer them. I'm not finished yet, but so far the viper code generator is around 10x slower than C code, plus or minus, which is much slower, but not as much as some people seem to think. The specifics of running an ADC at 20KHz are that the Pi Pico's can do it, the Pico2 at 28ks/s and the RISV 38ks/s. At stock clock speeds. It's prompted me to write a block input node that would run dma and emit sample frames for batch processing.
It's worth remembering that for stuff that really takes some cycles, you can drop down to machine code, like this fft from Peter Hinch. https://github.com/peterhinch/micropython-fourier