Path: utzoo!utgpu!news-server.csri.toronto.edu!mailrus!wuarchive!udel!udccvax1!mccalpin
From: mccalpin@vax1.acs.udel.EDU (John D Mccalpin)
Newsgroups: comp.arch
Subject: Re: RS6000 Multiply/Accumulate instruction
Message-ID: <5913@udccvax1.acs.udel.EDU>
Date: 22 Mar 90 22:33:00 GMT
References: <1872@awdprime.UUCP>
Reply-To: mccalpin@vax1.acs.udel.EDU (John D Mccalpin)
Organization: College of Marine Studies, Univ. of Delaware
Lines: 48

I wrote:
---------------------------------------------------------
   Errors in solution of the 1000x1000 system of equations
           from the LINPACK benchmark suite
machine                 precision               RMS error
---------------------------------------------------------
ETA-10                   64-bit                 1.3e-08
Cray X/MP                64-bit                 2.5e-11
IBM 3090                 64-bit                 5.8e-12
IBM RS/6000              64-bit                 1.2e-12  <-- new ibm
IEEE (Sun-3)             64-bit                 2.3e-13  <-- ieee
VAX "D"-format           64-bit                 7.2e-14
---------------------------------------------------------

In article <1872@awdprime.UUCP> ...@cs.utexas.edu:ibmaus!auschs!troyhix.austin.ibm.com!troy writes:
>    Needless to say this has caused quite a stir here. :-(

I have said before, and I will say again, that I do not believe that
the results that I have shown imply that there is anything *wrong* with
the IBM machine.  The question of error in this case came up because
of the unusual add/multiply instruction of the RS/6000 and I simply
wanted to see how much difference it made in this test case.  It looks
like IBM has a surprisingly strong product here....

The actual accumulation of roundoff error is an inherently data-dependent
process, and the above results are not strong enough to imply that
the IBM results are "bad".  The one data point above does not say
anything useful about whether the IBM floating-point is better or
worse than the IEEE floating-point, it *does* says that there is a
non-trivial difference (~2 bits) in the RMS error of the solution 
of this system of equations.  Other matrices will generate other results,
some in the opposite direction (with the IBM more accurate than IEEE).

>I brought this to the attention of our resident Floating-point math
>guru, Dr. Peter Markstein of IBM Yorktown research, who was one of the
>original 801 (first RISC) team member.

Glad to get some attention!

[... example programs and explanation deleted ...]
>                              Peter Markstein
>uunet!cs.utexas.edu!ibmaus!auschs!mrkstn.austin.ibm.com!mrkstn
>       or         ...@cs.utexas.edu:ibmaus!auschs!mrkstn.austin.ibm.com!mrkstn
>troy@troyhix.austin.ibm.com
-- 
John D. McCalpin                               mccalpin@vax1.acs.udel.edu
Assistant Professor                            mccalpin@delocn.udel.edu
College of Marine Studies, U. Del.             mccalpin@scri1.scri.fsu.edu