SPSS apparently only now did a full verification

[uunet!cwi.nl!Dik.Winter]

 > Could you post to numeric-interestavalidgh.com any details you may have of
 > this?   The whole issue of 486-Pentium comparisons is fuzzed up by the
 > fact that the transcendentals on Pentium are generally more accurate -
 > except where they access those two missing table entries in the divider.

We have the original article from spss no longer online, but there is a
follow-up, also from spss, which I include here (reformatted because of
long lines in the original).

dik
--
dik t. winter, cwi, kruislaan 413, 1098 sj  amsterdam, nederland, +31205924098
home: bovenover 215, 1025 jn  amsterdam, nederland; e-mail: dikacwi.nl
--
Newsgroups: comp.soft-sys.spss
Path: cwi.nl!sun4nl!EU.net!news.kreonet.re.kr!overload.lbl.gov!lll-winken.llnl.gov!sol.ctr.columbia.edu!howland.reston.ans.net!cs.utexas.edu!news.unt.edu!hermes.oc.com!internet.spss.com!news
From: peckaspss.com
Subject: Pentium Analysis, part II
Message-ID: <D0GHFo.GMsaspss.com>
Sender: newsaspss.com
Organization: SPSS Inc
X-Newsreader: AIR News 3.X (SPRY, Inc.)
Date: Wed, 7 Dec 1994 19:42:58 GMT
Lines: 44

This post seems not to have made it the first time.  Apologies if this
is a duplicate.

Here is further information on the Pentium bug and SPSS Inc software.

The SPSS testbed is designed to control variations in random number
generators and other factors.  We always expect, however, that there
will be some differences from one hardware platform to another.  That
is why when we run our testbed, we normally compare results only to
five decimals, and even then sometimes results will differ slightly.
Computers are finite-precision devices.   Floating point models differ.
Compilers differ from platform to platform in precisely how they reorder
computations,  when register to memory precision losses occur, and
other factors.

We would not expect to see any output differences in our testbed when
a 486 is compared to a Pentium -- absent the bug -- and in fact the
testbed does agree exactly to printed precision except in the instances
mentioned in my previous post.  This is not like comparing the Pentium
testbed output with, say, the IBM mainframe testbed output.  However,
there may be some algorithmic differences between the Pentium and the
486 in the evaluation of built-in functions that could produce a visible
difference.

We have not proved that the differerences are necessarily due to the
division problem because we have not trapped and compared all of the
arithmetic operations in the places where results differ.  We have
found differences in the results between the Pentium and a 486 in the
statistical output after controlling the other known sources of variation.
We did not find differences big enough to matter in the testbed runs,
but we did reproduce a known incorrect division result via a transformation.

We can't turn off the floating point hardware with the NO87 switch --
it doesn't work under Windows.  If we built the system with software
floating point, we would introduce variation due to the compiler and
run-time libraries.  If we had a corrected Pentium chip, we could compare
our testbed output with output generated on that hardware.

Intel has today promised to Fed Ex us a corrected chip to try the cases
where differences were found minus the division bug.  We will rerun
these parts of the test and post the results as soon as possible.

For now, what we know is
1.  You can get a lot less precision than you would expect if you do the
right (wrong?) transformation (and hence analysis on the output of that
transformation).
2.  There are a few variations in the statistical testbed output, which
was not designed to include this division bug, from the results on other
Intel processors with the same floating point model.
3.  These variations are too small to matter in the testbed output, but
we cannot rule out more important variations.

We have not gotten any data from the field yet that adds any more
information to this.

Kim Peck
SPSS Inc.
peckaspss.com