1. 16
  1. 3

    Also mentioned at the bottom of this is that there is an official IEEE 754 standard for decimal floating-point, available in IBM’s POWER chips for several years now (e.g. from a PDF about POWER9). I could not tell exactly when this page was originally authored (it is 5 years old at least), but a claim today that “[IEEE 754 dfp] has not found much acceptance” might not be as strong as it once was?

    1. 3

      I’m not sure why you would use this.

      Numerics are like cryptography; you want to use a library written by experts who live and breathe the domain.

      IEEE 754 decimal floating point has at least 3 software implementations and some hardware support.
      Intel’s reference implementation implementation isn’t listed but also looks pretty good.

      DEC64 is unspecified, doesn’t support rounding modes, has buggy implementations of mathematical functions, and industry support is nonexistent.

        1. 2

          This looks like a numeric type that would be fun to code up a module for in Verilog! The handling of NaN values thankfully doesn’t look to be any more complicated than normal IEEE 754 floating-point, so that’s good.

          One interesting thing I saw just from glancing over the page is that the value 0 has 255 different representations – all of which are considered equivalent. Having not built an ALU working with DEC64’s, I can’t say yet whether or not that will be an issue in practice.

          Stories with similar links:

          1. DEC64: Decimal Floating Point via inactive-user 8 years ago | 6 points | 3 comments