Great points. I agree - the performance hit is simply the cost of being accurate and having predictable behaviour.

I'm not suggesting we replace all our current HW with chips that implement posits (let's fix branch prediction first!!). More that FP should be opt-in for most HLLs.