The EXCEL file contains visualizations of all currently analyzed pfloat types. You might also check the pdf files in this directory for details.

For CNN simulations, it can be beneficial to have consecutive 16->32->64-bit types, where each 'higher type' has 2x the dynamic range and more than 2x the number of mantissa bits.

For approximate computing, it is relevant that the full dynamic range is available when calculating (coarsely) with lower bit numbers. The pfloat[x]d types all have the same full dynamic range as double for that reason.