Path: utzoo!attcan!uunet!husc6!rutgers!rochester!pt.cs.cmu.edu!sei!sei.cmu.edu!firth
From: firth@sei.cmu.edu (Robert Firth)
Newsgroups: comp.arch
Subject: Re: A simple question on RISC
Message-ID: <7722@aw.sei.cmu.edu>
Date: 16 Nov 88 13:04:14 GMT
References: <76083@sun.uucp> <559@dms.UUCP> <1021@raspail.UUCP> <7819@winchester.mips.COM> <1688@scolex> <13844@lll-winken.llnl.gov>
Sender: netnews@sei.cmu.edu
Reply-To: firth@bd.sei.cmu.edu (Robert Firth)
Organization: Carnegie-Mellon University, SEI, Pgh, Pa
Lines: 28

In article <13844@lll-winken.llnl.gov> brooks@maddog.UUCP (Eugene Brooks) writes:

>Seymour made the observation that floating point divides are much less
>frequent than multipy or add...

True

>... so one could implement them with a less
>complex reciprocal approximation unit followed by a Newton Raphson hit or
>two...

False.  This leads to results less accurate than those given by a true
divide.  For instance, 21.0/7.0 gives an exact 3.0 with a true divide,
but usually will not if computed as 21.0*(1/7.0).

>...This disproves the notion that Seymour just hacked in what he could do
>fast...

Really?  I'd say it confirms the notion.

>... he obviously made careful decisions about what to put in hardware
>based on how heavily instructions were used and how fast they could be
>implemented.

Too bad he didn't base his decision in part on whether users performing
floating-point calculations prefer better answers to worse ones.  Too
bad also that the 'megaflops per millisecond' benchmarking culture tends
to be heavily biased in favour of "wrong answers fast".