On Mon, Mar 13, 2023 at 3:27 PM John Cowan <[email protected]> wrote: > Currently, SRFI 231 requires an 8-bit floating point storage class. > However: > > 1) There is no hardware support anywhere for such a thing. > > 2) There is no standard of any kind, either IEEE or otherwise, for such a > thing either. > > 3) The sample implementation doesn't provide them. > > 4) I researched various articles that discuss them, and basically some are > in favor of 4-bit exponent and 3-bit mantissa, and others the opposite > way. Furthermore, there are different plausible values for the exponent > bias: 1 is a common value, making 1/8 the minimum positive float and 15360 > the maximum, but -2 is also a possibility, in which case all representable > values are integral and the maximum positive float is 122880. >
I've used and in fact implemented my own 8-bit floating points before. At least for my use case I found a 5-bit exponent and 2-bit mantissa worked best. I think with so few distinct values magnitude matters much more than precision so this seems a good default, but you can find every variation imaginable. -- Alex
