Standard extended format for binary floating-point numbers?

[Dennis Brzezinski]

In an earlier memo Bill Janssen asks:Can anyone tell me if there is a standard,
>or an emerging standard, or a de facto standard, for floating point 
>representations longer than 64 bits?  

The IEEE-754 does not "fully" define a higher precision than double, but rather
defines the minimum acceptable extended double format via >= sizing.  The
two most popular extensions are the 128 bit quad precision and the 80bit 
extended double.  Both of these make "sense" in that the 128bit uses two double
words in memory or two adjacent 64-bit registers, and  the 80bit, in addition 
to meeting the mimimum(79bit)  requirements, has a 64bit significand and hence
the hardware can be shared for 64bit integer operations. 64-bit multipliers
and adders map well into efficient circuits and topologies.

HP has defined in its PA-RISC architecture a quad precision floating point
format that is within the open ended extended double precision defined in the
IEEE 754-1985 standard.  This quad precision format is defined in 128 bits, 
which for PA-RISC represents two 64bit double words in memory.  Quad precision
numbers must be aligned on double word boundaries.

PA-RISC QUAD PRECISION BINARY FLOATING POINT:
---------double word 1---------------|--------double word 2----------|
sign  exponent     high fraction     |          low fraction
 1      15               48          |             64

  Note that the integer bit is "implicit" or hidden as in the single and
double precision definitions in the IEEE 754. This yields a 113 bit significand
(a 112 bit mantissa or fraction).  Presently quad floating point math is 
handled through software routines.  The SPARC architecture from SUN has the 
same definition for the quadword format.

 The most popular extended format is, by far, the 80bit extended double used
by Intel in its x86 processor family.  It uses an explicit integer bit for:

----halfword ---|--------double word ----------|
sign    exponent     significand
 1       15           1. 63bit fraction


Dennis Brzezinski
HP Labs  Palo Alto
A private posting.