Path: utzoo!attcan!uunet!aplcen!samsung!cs.utexas.edu!tut.cis.ohio-state.edu!ucsd!helios.ee.lbl.gov!nosc!marlin!aburto
From: aburto@marlin.NOSC.MIL (Alfred A. Aburto)
Newsgroups: comp.sys.amiga.hardware
Subject: Re: URGENT How fast is 2630
Message-ID: <1335@marlin.NOSC.MIL>
Date: 2 Mar 90 21:08:34 GMT
References: <4ZvUba200XNI88ZUU8@cs.cmu.edu>
Reply-To: aburto@marlin.nosc.mil.UUCP (Alfred A. Aburto)
Distribution: comp.sys.amiga.hardware
Organization: Naval Ocean Systems Center, San Diego
Lines: 45

In article <4ZvUba200XNI88ZUU8@cs.cmu.edu> Michael.Witbrock@cs.cmu.edu writes:
>I do neural network research. That means I to an awful lot of
>calculations of the form
>
>x= sum(over i) w(i)*o(i)
>
>i.e. multiplies and sums.  I don't care about savages or mandlebrot
>tests, and I'm willing to write
>assembly code for my inner loops. 
>
>I just found out that I can buy an unbundled 2630 at edu discount if I
>act within the next month.
>
>EXACTLY HOW MANY FLOATING POINT OPERATIONS PER SECOND CAN A 2630 BOARD
>DO, IF THEY ARE EQUAL NUMBERS OF MULTIPLYS AND ADDS (they can be short
>floats, not doubles, if this helps).  I need to know whether my typical
>run would be a week (acceptable) or a month (unacceptable) (and you
>thought ray tracing was bad!).
>
>regards. Michael
>

-----------

The 68882 @25 MHz can do typical Double Precision Scalar FADD/FSUB/FMUL
and FDIV's at a rate of approximately 0.46 MFLOPS Peak --- This is for
scalar operations with variables in FPU registers and fast 32-bit memory. 

When working with arrays, as you indicated, where the variables can not
be assigned to FPU registers then the MFLOPS rating goes down (perhaps
alot depending upon the code a compiler generates).  The Livermore Loops
program (LLOOPS) can give you a good estimate of MFLOPS rating in this
case.  I think the A2630 with 32-bit fast memory and a good 020 compiler
(Like Lattice C 5.04) will achieve 0.3 MFLOPS or so when working with 
FP arrays.

If you want high performance with FP arrays then you will need to use
an 'Array Processor' which can do A[I] * B[I] in one operation for all
elements for example --- this is the way to get the big MFLOPS ratings ...

There might be a way to interface an 'Array Processor' to the Amiga, but
I'm not aware of such a system ...

Al Aburto
aburto@marlin.nosc.mil