Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!rutgers!usc!bbn!bbn.com!slackey
From: slackey@bbn.com (Stan Lackey)
Newsgroups: comp.arch
Subject: Re: John von Neumann, sqrt instr
Message-ID: <44558@bbn.COM>
Date: 21 Aug 89 14:30:55 GMT
References: <21353@cup.portal.com> <25643@obiwan.mips.COM> <1513@l.cc.purdue.edu> <2376@wyse.wyse.com>
Sender: news@bbn.COM
Reply-To: slackey@BBN.COM (Stan Lackey)
Distribution: usa
Organization: Bolt Beranek and Newman Inc., Cambridge MA
Lines: 24

In article <2376@wyse.wyse.com> stevew@wyse.UUCP (Steve Wilson xttemp dept303) writes:
>In article <1513@l.cc.purdue.edu> cik@l.cc.purdue.edu (Herman Rubin) writes:
>>There is no reason why a machine with hardware division should not have
>>hardware square root.  It costs almost nothing.
>Ah, but that presumes that hardware division is warranted!
>Does the occurance rate of divide/square root in scientific computing
>justify the cost?
>How does the scientific computing community feel about this functionality?

I was once involved in a trade-off between doing a divider gate array,
or just iterating through the multiplier.  Performance was that for a
multiply of 3 cycles, the division would take 18.  Also, vector
multiplication would be one chime, and vector division 18. (We still
considered it worthwhile to keep the vector division instructions, in
case a future implementation had a faster one and also to make things
easier for the compiler.)  We found that over a wide range of
applications the ratio of mul to div was like 3:1 to 5:1.  Worse
still, Ahmdahl's Law predicted that with those numbers we'd be
spending nearly 50% of our time doing division!!  In the applications
we wanted to do well at, that is.

We did the gate array.  And yes it has square root.  I personally
don't know how important sqrt is, but it was cheap enough.
-Stan