Re: Number Cruncher FPUs

[dinarich@di.epfl.ch (Mirko Dinarich)]

In article <dirk-0406961121130001@muffin.maschinenbau.uni-dortmund.de>, dirk@gaga.maschinenbau.uni-dortmund.de (Dirk Froehling) writes:
>
>I think the difference between the 881 and 882 is that the 882 has more
>instructions hard-wired and uses less clock cycles. There is a big performance
>difference between the two when used on a Mac.
>So with a 881 your point is valid, but I don't know about the 882...
>

I've checked from MC68881/MC68882 Floating-Point Coprocessor User's Manual
and the instructions time are quite similar; 68882 most important new features 
are concurrent instruction execution and high speed hardware float conversion 
format.
Also some instructions are executed faster (but not as fast as you think),
here is a little comparison table (taken from MC68881/MC68882 Floating-Point Coprocessor User's Manual) of instruction's execution time in clock cycles: 

        68881        68882
FABS      60 (*)       57     (*) sorry, this value was wrong in my 
FASIN    606          603                previous post
FCOS     416          413
FDIV     130          127     (PS: these # are without effective address 
FLOGN    550          547          calculation time and are suitable for 
FMUL      98           95          double calculation)


Even if we half the execution time for a two concurrent instructions execution
(and technically is not true) the average execution time is generally greater 
than 1 Mhz.

These are theorical considerations, in practice it should be verified how
NumberCruncher interface the fpu with the Apple system bus.
-- 
!\  /!  !  !  /  /--\  /--\  /--\  /--\     
! \/ !  !  ! /   !  /  !  !  !  /  !  !  dinarich@di.epfl.ch 
!    !  !  ! \   ! \   !  !  !  \  !  !  http://diwww.epfl.ch/~dinarich
!    !  !  !  \  !  \  \--/  \--/  \--/