home
***
CD-ROM
|
disk
|
FTP
|
other
***
search
/
Light ROM 4
/
Light ROM 4 - Disc 1.iso
/
text
/
net_news
/
1994
/
2.doc
/
000179_ernie@gaspra.pd.com_Thu Dec 8 14:20:25 PST 1994.msg
< prev
next >
Wrap
Text File
|
1995-03-23
|
3KB
|
56 lines
Article: 432 of comp.graphics.packages.lightwave
Xref: netcom.com comp.graphics.packages.lightwave:432
Path: netcom.com!ix.netcom.com!howland.reston.ans.net!gatech!newsxfer.itd.umich.edu!zip.eecs.umich.edu!newshost.marcam.com!news.kei.com!ddsw1!news.rtd.com!gaspra.pd.com!ernie
From: Ernie Wright <ernie@gaspra.pd.com>
Newsgroups: comp.graphics.packages.lightwave
Subject: Re: Pentium Bug WAS Re: LightWave 4.0 Delays
Date: Wed, 7 Dec 1994 23:43:42 -0700
Organization: RTD Internet Access, a division of RTD Systems & Networking, Inc.
Lines: 38
Message-ID: <Pine.SUN.3.91.941207233425.4133A-100000@gaspra.pd.com>
References: <1fc65d78.1d8eecab@amuc.mtroyal.ab.ca> <shfD0938M.4vu@netcom.com> <3bt0b7$5ae@news.cs.tulane.edu> <3c3vlg$8ri@nnrp.ucs.ubc.ca> <3c51l2$e4u@news.cs.tulane.edu>
NNTP-Posting-Host: pd.com
Mime-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
In-Reply-To: <3c51l2$e4u@news.cs.tulane.edu>
The discussion here about the Pentium FPU bug's affect on LightWave
has, I think, greatly exaggerated the size of the problem.
The worst possible relative error is about 0.00006 (6e-5). Relative
errors larger than 0.0000001 (1e-7), approximately epsilon for single
precision, occur on average once every 65 billion divides.
If the LW renderer is spending more than 1 or 2 percent of its time in
FDIV instructions, I'd be very surprised. At that rate, it would take,
on average, more than a month of continuous rendering before a single
relative error greater than 1e-7 occurred, and on that basis, the MTBF
(mean time before failure) for Modeler is several years.
Most of the calculations LW does aren't that sensitive to small errors.
Even if an error greater than 1e-7 survives the mayhem of roundoff that
is always happening anyway, the chance that this error will actually be
visible is vanishingly small. It takes an error 100,000 times larger
to produce a visibly incorrect pixel, for example.
Will the Pentium bug affect LightWave? The right answer is a simple
"No." Just keep in mind that *somebody* wins the lottery. And it's
always possible that some perverse coincidence will cause LightWave to
preferentially generate the problem numerator-denominator pairs, or
that the bug is worse than we thought. But stuff like that is ALWAYS
possible. As computer users, you've got failures to worry about that
have MUCH higher probabilities.
My thanks to Dave Gilinsky of Pixel Dust, Inc., for forwarding his
collection of information on the bug.
And a final clarification. In my "Pentium test" post, I said that Tim
Coe was using MATLAB when he discovered the worst case error. Coe is
a chip designer for Vitesse Semiconductor and discovered the worst case
by studying the bit patterns that had previously produced errors. I
read about it in a post about MATLAB, which is where that reference
entered my head.
- Ernie