Path: utzoo!utgpu!watserv1!watdragon!rose!sbnicol
From: sbnicol@rose.waterloo.edu (Scott Nicol)
Newsgroups: comp.sys.ibm.pc
Subject: Re: Math chips and CAD was: recommendations wanted
Message-ID: <20638@watdragon.waterloo.edu>
Date: 7 Feb 90 18:36:10 GMT
References: <25ce13f4.5fd6@polyslo.CalPoly.EDU> <8370003@hpavla.AVO.HP.COM>
Sender: daemon@watdragon.waterloo.edu
Reply-To: sbnicol@rose.waterloo.edu (Scott Nicol)
Organization: U. of Waterloo, Ontario
Lines: 33

In article <8370003@hpavla.AVO.HP.COM> boyd@hpavla.AVO.HP.COM (Bill Boyd) writes:
>About doing math-intensive operations with and without a coprocessor:
>
>I adapted a program for displaying the Mandelbrot set from an old issue
>of Turbo Technix.  It would take 20 minutes to run on my XT with 8087.
>I bought a 16-MHz 80386SX machine without a math coprocessor.  It took
>53 (!) minutes to run the same program!  I bought an 80387SX, and now
>it takes 3 minutes to complete.

  About five years ago, as a school project, a friend and I wrote a program
that accepted a function in three variables (x, y, and z), and plotted a
(accurate) 3D plot of the function.  The machine we were using was a
(brand new at the time) Compaq Deskpro (8 Mhz 8086), with an 8 Mhz 8087.
Using an inefficient C compiler (C86), the program could plot a reasonably
complex function in a minute.  without the coprocessor, we estimated it
would take 2 hours (we didn't run it for 2 hours, we ran it for 5 minutes
and estimated how far along it was).
   Last year, I discovered the source code lying around on one of my old
disks.  I ported it to Turbo C (2.0), and compiled it (with 286 code
generation) on an IBM PS/2 model 70 (25Mhz 8.  The
same function that took a minute on my old Deskpro took about 10 minutes
on the 386.  After that, I took the program to a 33Mhz Compaq Deskpro
(with a coprocessor and a real fancy EGA), it took less than 5 seconds.

   So, if software is numerically intensive, and if it is smart enough
to recognize the coprocessor, then it can make a huge difference.
Of course, if the program is I/O bound, the '387 will hardly make
a dent.

- Scott
Email: sbnicol@rose.waterloo.edu  -or- ...!uunet!watmath!rose!sbnicol
SlowMail: 546 FallingBrook Dr., Waterloo, Ont., Canada, N2L 4N4.
Phone: (519) 725-1980