🚀 go-pugleaf

On Wed, 14 Dec 1994, G. RAY MEESTER wrote:

> IBM said it talks the normal(?) comsumer 24 days to have an error on the
> pen...chip.  Intel said it took 27000 years for the normal(?) comsumer to
> have an error.  Can anybody explain if this is the same or different??
> How did they calculate these numbers????
> meesterr@mscd.edu

This post got a little long.  I hope it is highly informative, although I
get a little opinionated near the end.  For further discussion, go to
comp.sys.intel, where the discussion runs 10-1 against Intel, mostly by
unhappy Pentium owners.  I won't post further discussion of this to
Graph-TI, I promise.

Opinions expressed below are my own and are not intended to represent others.

If you have world-wide-web  access, try
	http://www.ibm.com/Features/pentium.html  ,

...where IBM explains in considerable detail their reasoning.  It focusses
considerable attention on their claim (which I find credible) that while
only one in a 9 billion divisors are affected, the bad ones could show up
highly-disproportionately-often in common spreadsheet applications.  As
of a day or two ago, IBM said they would no longer ship Pentium machines
until the problem was fixed.  Since Pentiums are a relatively small
portion of IBM's business, some say this is a move to discredit Intel as
IBM prepares to introduce its PowerPC machines. (i.e. they say IBM is
playing hardball).

Intel, on the other hand, assumes all numbers are used equally often for
divisors.  You can get Intel's side of the story at
	http://www.intel.com/product/pentium/white11/index.html ,
...which is an index for their white paper on the Pentium flaw.

For those who came late, all Pentium chips until recently, including
probably all now shipping, have a flaw in the floating-point division
section which can cause errors as large as in the fifth significant digit.
Intel says this is not a major problem, IBM disagrees.  If you want your
Pentium replaced, you must persuade Intel that you are using applications
that use floating-point division extensively, and for which high accuracy
is necessary.  Many computer sellers are referring owners of their
computers directly to Intel. Intel refuses to tell buyers what external
markings will distinguish the flawed chips from soon-expected-in-pipeline
fixed chips, apparently since they are determined to sell all their old
stock.

To test your pentium, find an application that you _know_ uses the
floating-point coprocessor and divide 4195835 by 3145727 (add ".0" if
your software requires it).  Compare with answer given by TI
calculators.  If the fifth sig fig is wrong on the computer, your Pentium
has the bug.  The windows calculator in scientific mode, although itself
buggy, suffices for a test.  Note that there are many more "bad"
divisor/numerator pairs, and that numbers close to the above are also bad
(say +/- .05 or so).  This error is consistent; you get the same bad
result every time you divide the same "bad" pair of numbers.

I've already had my Pentium replaced; it wasn't especially easy to do.  I
remain somewhat disgruntled that Intel now admits it knew about the
problem in "the summer" or even in June, yet still mailed me a flyer about
how the Pentium was superior for floating-point applications in late
August, with absolutely NO hint that they knew the chip was flawed.  (I
bought the machine at the end of August).  I _am_ biased in this
discussion, although all my connections to IBM (none) or Intel are fully
disclosed above.

Good luck and happy holidays to you all,

Mark Janeba			(503)370-6123
Dept. of Mathematics		e-mail:  mjaneba@willamette.edu
Willamette University		"Speaking for myself, not the university"
Salem, Oregon 97301-3922 USA