New approach could sink floating point computation (Posits)

https://www.nextplatform.com/2019/07/08/new-approach-could-sink-floating-point-computation/

• float was designed in 1985 by IEEE.
• Performance and accuracy is not the best
• John Gustafson reworked the “float” idea (his 3d iteration) -> Posits
• Difference: denser representation of real numbers.
• Instead of fixed-size exponent and mantissa -> variadic representation
• Additional “regime” part, of ANY length

  

• Coded as “total number of bits + max number of bits for exponent”: posit< bits, es >
• Priority: sign -> regime -> exponent (up to es bits, =0 if no bits available) -> fraction (=1.0 if no bits available)
• Advantages:
  • +∞ = -∞
  • only one NAN (1/∞)
  • Higher precision (32bit posits can replace 64bit float in almost all use cases!)
  • Higher performance

    

• IMHO
  • Seems to be a better variant then IEEE float from 1985
  • Not sure that it can’t be improved
  • Very difficult to support in hardware -> I’d RFC it first