Path: utzoo!utgpu!jarvis.csri.toronto.edu!rutgers!ucsd!sdcsvax!celece!fellman
From: fellman@celece.ucsd.edu (Ronald Fellman)
Newsgroups: comp.sys.next
Subject: Re: Floating point
Message-ID: <6636@sdcsvax.UCSD.Edu>
Date: 14 Jun 89 05:49:06 GMT
References: <89157.122728UH2@PSUVM> <680008@hpcuhc.HP.COM>
Sender: nobody@sdcsvax.UCSD.Edu
Reply-To: rfellman@ucsd.edu (Ronald Fellman)
Organization: UCSD Department of ECE
Lines: 34

In article <680008@hpcuhc.HP.COM> edwardm@hpcuhc.HP.COM (Edward McClanahan) writes:
>> I have compared a MacII, a NeXT, a Sun 4/280, and a DecStation3100
>> running the CD3040 opamp in Berkeley spice3b1. (I don't have the
>> graphics part running on the NeXT but that didn't matter for this
>> benchmark.)
>
>> Here's the results:
>> MacII: 59 sec., NeXT: 21 sec., Sun 4/280: 8.7 sec., DecStation 3100:4.0sec.
>
>Did this application take advantage of the Motorola 56001 DSP chip
>in the NeXT?  Is it reasonable for users to expect applications developed
>for multiple machines (MacII,NeXT,Sun4,DecStation3100,Sun3,etc...) to
>take advantage of such "product differentiators" as this?  And, if we
>are only comparing the CPU(and memory systems, e.g. cache, main, secondary),
>why was the NeXT (at 25MHz) so much faster than the MacII (at 15.67MHz)?
>
>Ed "enquiring minds want to know" McClanahan

That didn't use the DSP chip in the NeXT. In fact, a 56001 uses 24 bit
fixed-point multiplication, NOT double-precision floating point.

As for the speed difference between the NeXT and the MacII:
the NeXT uses a 68030/68882 combination. The Mac II I had used a
68020/68882 combination.  The 68030 has a data cache on-chip which must
have made most of the difference. Spice probably had a high hit rate in
some of its inner loops.  As for the 68882 on my MacII, I just stuck it
there without telling Lightspeed C 3.0 about it (in fact it doesn't know
how to optimize code for a 68882.) Simply plugging a 68882 into a MacII
just helps by about 10%.  Apparently, if your compilers know how to
reorder instructions properly, one could get more leverage out of the
'882s pipelining. I don't know if the NeXT C compiler does this or not.

-ron fellman
(rfellman@ucsd.edu)