performance concerns

[gnotu]

I created a couple synthetic benchmarks and also measured the mlcommons MLPerf Tiny benchmark on a RP2350.
The results are here.
I found that models compiled using pico-tflmicro were up to 23x slower than reported results on comparable platforms.
Before I dive into sorting it out, I thought I'd post the issue here.
The 23x faster results were running on a 160MHz NUCLEO-U575ZI-Q (also Cortex-M33) with models compiled using X-CUBE-AI.
A few possibilities are:

1. I'm doing something wrong.
2. X-CUBE-AI really is that much better.
3. The memory bandwidth on the Nucleo board is a lot higher

I don't have access to the Nucleo platform to make side-by-side measurements.
I'll report back here if I find the root cause.
But if someone has a clue, I'd appreciate it if you shared it.

Thanks!

--Mike