Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!cs.utexas.edu!uwm.edu!psuvax1!rutgers!att!bellcore!texbell!attctc!chasm
From: chasm@attctc.Dallas.TX.US (Charles Marslett)
Newsgroups: comp.sys.ibm.pc
Subject: Re: FasMath Math Coprocessor ?
Summary: It's fast, but not that fast (in the real world)
Message-ID: <10405@attctc.Dallas.TX.US>
Date: 29 Nov 89 15:44:08 GMT
References: <8911290805.AA05793@euler.Berkeley.EDU>
Organization: The Unix(R) Connection, Dallas, Texas
Lines: 32

In article <8911290805.AA05793@euler.Berkeley.EDU>, ndeng@EULER.BERKELEY.EDU writes:
> The latest issue of PC-Magazine (vol.8, no.21) reported a new 80387 compatible
> math-coprocessor, FasMath 83D87 from Cyrix Co. According to the company, this
> chip can be up to 10 TIMES FASTER than 80387, and is fully pin and SW compatible.
> I am extremely interested in this, but thought it's almost incredible. If this
> is proven, then 80387 and Weitek are all dead, and many accelerator boards 
> become obsolete. Does any number-crunching guru has hand-on experience about
> this chip? Could any of you share you experience/excitement with netters?
> 
> ndeng@euler.berkeley.edu

I have played a bit with it:  it is definitly faster than the intel 387.  It
also runs a lot cooler (they claim some despicably low power consumption, and
I do believe that!).  As I understand it, it often runs ten times as fast
internally that does the intel part, but after you add in the overhead (bus
cycles from memory to 386 to FPU to 386 to memory, etc.) the ratio is a lot
more like 2:1 or 3:1.

Cyrix seems to have done a bit more to exactly match the numbers intel generates
in their transcendental functions (good, bad, random, or whatever) that IIT
has, but the tests I have run show differences between the three vendors to
really be down in the mud (about 1 in 10 evaluations report differences, and
usually the difference is because the actual value is almost exactly halfway
between the two reported values!).

Recap:  FasMath is faster and much lower power, IIT's 3c87 is faster (about
the same power requirements?), and Intel's 387 is more available (and maybe
cheaper).

Charles Marslett
STB Systems, Inc.  <-- apply all standard disclaimers
[I wrote the SSDC/IIT demo/benchmark program, also.]