the first bullet point mentions significant energy efficiency improvements, but was there anywhere that the math behind this efficiency was mentioned? might've missed it...

I believe this is just from more efficient parallelism reducing the number of ops we need to preform so it takes less energy.

third bullet point -- to gain that insight, i assume tools equivalent to Intel's vtune exist?

What does it take to get a language like Spatial actually used in practice, at places like Google (E.g. for the edge TPU)?

(Sorry, I meant *Chisel for the edge TPU…)