On a slight tangent, I do wonder if the Raspberry Pi is the right dev board for running compute-intensive AI models. It is cheap but the real hero of the Pi is the Broadcom SoC. Where as there are other dev boards out there - lots of them in fact - which would have a faster CPU, dedicated ethernet bus (even gigabit ethernet - which would be better for clustering too), etc.

I do love my Pis (I have a few of them) but if you're having to overclock the thing and then strap on a CPU fan to make the Pi even viable for running compute-intensive AI models then perhaps the exercise might have been better suited for running on another piece of hardware instead?