andyo@glasperl.masscomp (Andy Oram) (11/28/90)
I would like to find out what binary floating-point formats are currently in
use, besides the familiar IEEE 754 and DEC PDP/VAX. If you know of such
existing systems, please tell me:
1. The bit formats themselves.
2. The machines built with them. (Just hosts. I realize that
any attached processors by the same manufacturer or by VAR's will use
the same formats.)
3. A brief assessment of why the formats were chosen, and whether they
offer any advantages over IEEE 754. (Simplicity, range, and speed are
all valid advantages).
4. Whether you expect the formats to be used in future architectures.
Please e-mail responses. Nobody but me could be cookoo enough to be
interested in this issue.
-------------------------------------------------------------------------------
Andrew Oram Concurrent Computer Corporation
Analog Communications Routes: One Technology Way
Westford, Mass. 01886
(508) 392-2865
Digital Communications Routes: andyo@westford.ccur.com
{harvard,uunet,petsd}!masscomp!andyo
(I don't represent Concurrent; this message represents my own opinion)
Automatic UN*X documentation utility:
cref | compress | crypt | man
-------------------------------------------------------------------------------andyo@glasperl.masscomp (Andy Oram) (11/28/90)
My mailer is garbling my From: line, as a kind respondent just pointed out to
me. Please send your responses to andyo@westford.ccur.com as shown in my
signature.
-------------------------------------------------------------------------------
Andrew Oram Concurrent Computer Corporation
Analog Communications Routes: One Technology Way
Westford, Mass. 01886
(508) 392-2865
Digital Communications Routes: andyo@westford.ccur.com
{harvard,uunet,petsd}!masscomp!andyo
(I don't represent Concurrent; this message represents my own opinion)
Automatic UN*X documentation utility:
cref | compress | crypt | man
-------------------------------------------------------------------------------richg@locus.com (Rich Greenberg) (11/29/90)
In article <ANDYO.90Nov27184117@glasperl.masscomp> andyo@glasperl.masscomp (Andy Oram) writes: >I would like to find out what binary floating-point formats are currently in >use, besides the familiar IEEE 754 and DEC PDP/VAX. If you know of such >existing systems, please tell me: The format used by the IBM 360/370/390 series is as follows: (I am not familiar with IEEE 754 or DEC PDP/VAX so I don't know if they are like this one, but knowing IBM I doubt it.) The first byte: leftmost bit = sign of mantissa portion right 7 bits = power of 16 the mantissa is raised to plus 127 (or mabe 128, not sure and the book is at work). Rest of number is the binary mantissa, can be: 3 bytes (short float, about 7 decimal digits precision) 7 bytes (long float, about 15 decimal digits precision) 14 bytes (extended float, about 30 decimal digits precision) The extended format actually has a second exponent in there that wastes a byte. Total is 16 bytes. Layout: emmm mmmm emmm mmmm as 4 32 bit words. Note that because the mantissa is raised to a power of 16 (not 2), that a normalized mantissa can have up to 3 zero bits on the left end. Rich
lusol@vax1.cc.lehigh.edu (12/01/90)
Floating point information for current CDC CYBERs running NOS/VE: 64-bit single precision, accurate to 15 digits, with a magnitude ranging from -5.2E+1232 to +5.2E+1232. Smallest negative number apporoximately -4.8E-1234, smallest positive number approximately 4.8E-1234. One coefficient sign bit, a 15-bit signed exponent biased by 4000(16), and a 48 bit coefficient. Two's complement. 128-bit double precision, ranging from -5.221944407065762533458763552E+1232 to 5.221944407065762533458763552E+1232, accurate to about 29 digits. Remaining 48 bits of the 96-bit coefficient stored in the next word. Integers on a CYBER are signed 64-bit quantities, ranging from -9223372036854775808 to 9223372036854775807. Two's compliment. More info in the CDC CYBER 180 state Hardware reference Manual. Steve Stephen.O.Lidie@CDC1.CC.Lehigh.EDU