Here is an interesting summary article about the history of bit field assignments for number representation. Includes a brief history of the newer 8 and 16 bit representations, like Block Float and AdaptivFloat. I did not realize this is an active area of development for DDN (Deep Neural Networks). Michael

https://www.techspot.com/article/2630-n ... -hardware/

Glad to see it includes the Posit/Unum ideas (at the end, not sure anyone has implemented the ideas yet)
https://posithub.org/about

I came across an article the other day: Mathematical model and implementation of rational processing about calculation and representation of fractional numbers as floating point values with additional pointers that indicate bits of the mantissa that repeat indefinitely.
If an operation runs out of precision, the result is just a regular floating-point number without any pointers.