Google Groups

11! Cyberplus [2]

Eugene N. Miya Dec 24, 1991 7:17 AM
Posted in group: comp.sys.super
Conversions with several 'notable' computer scientists has affirmed
the value of this analysis.

Usual disclaimer:

        "This information is based on the recollections of the
        submitters, which may be a little foggy.  Corrections on
        technical details are welcome."

Summary of anonmyous and posted thus far:

Person 1:
I remember talking to someone in the mid 80's who worked in product
development for the Cyberplus.  My recollection is It was a working product
in the early 1980's, had 20 nsec clocks, with 15 functional units (including
load, store, communicate, various logical, etc) in each processor, and
I thought about 8 or 16 processors in a double ring.  It used VLIW
instructions to allow independent control of each function unit on each
clock cycle.  For some specialized applications, it's performance was
awesome for it's day, and still not too shabby today.  There was an effort to
develop a CMOS version for commercial sales, but I never heard what happened
to that effort.

Person 2:

foggy's the word, at least for *my* recollection.

The best thing to call it is an attached array processor, hooked up to a
CDC 6000/Cyber mainframe.

It was independent of the Star/Cyber 200 line. Thorndyke had nothing to do
with it.  (That may be a recommendation, though).
CDC had several other array processors (the MAP series) ; don't know if
they would qualify as supercomputers.

|> OS:

 Assembly language routines, Fortran-callable. Cross-assembler under NOS.
 Binaries downloaded.

|> but CDC marketing were poor at follow up.  
   As usual.

They had one at a spook shop in Bonn (presumably the German equivalent of
NSA) and were going to install something really big near the Czech border
- NATO rather than FRG, I think, and SigInt (don't know if cryptography
or just computer enhancement of signals.) I nearly got assigned to run the
site for CDC but luckily I got the 205 in Karlsruhe instead and I never
heard any more about it. This was in 1982.
There also was an outfit in Duesseldorf that ran financial futures.
A big operation finance-wise ; not a big customer for CDC but they (the
customer) had a billion-dollar business. This was 1983-85.

Person 3:

If you are really interested in the Cyberplus,
I have some old VIM stuff at home.  It hung
on a Cyber channel.  Compiler system resided on
the mainframe for cross compilation/linkage.

It was basically an attached array processor for
Cybers.  Circa 1984.  Used a VLIW type instruction
set but you had to code each pipeline stage/delay
explicitly...  One of the guys in our office here
(xxxxxxxxxx) actually used one in a former life.

Person 4:  (!)

Actually there are still Cyberplus's running today in non-"black"
environments. One that I know of is a company using a number
of the machines to predict trends in the world currency
markets and to advise traders (for a goodly fee I suspect)
on which currencies to buy and sell (we're talking mega-buck
transaction here).  The algorithms and methods used are
extremely proprietary but apparently they work well enough
that people make money rather than lose it.

Person 5:

Wayne Ray is the guy you really need for a decent history of the Cyberplus.

I actually saw one (one time) at ARH, but the Cyberplus is a mystery
shrouded in an enigma contained in a riddle.  Wrapped up, not surprisingly,
in yards of CDC red tape.

From: (Robert Hyatt)
Subject: Re: 11! CDC/Cyperplus [1]

I seem to recall that Ga. Tech has (had?) one of these monsters.  I recall
reading the manual and saying "what a pile of ----" that made the programmer
"do it all himself..."  might have been fast, but wasn't easy or intuitive
to program and/or get good performance from.  Any Ga Tech'ers out there with
more info?

From: (Gunter)
Subject: Re: 11! CDC/Cyperplus [1]

Theoretically still the fastest system around, if you built it up to the full
64 CYBERPLUS Configuration using all the Channel Paths (who needs disks :), but
seriously i believe you can hang 8 C+'s onto each Channel, and each C+ (size of
a C180-825), ran at somewhere from 250-700 MFLOPS, so in the full 64C+ config
you'd have something like 44.8 GFLOPS! A Special C+ FORTRAN was supplied for C+
systems. Considering it came out over 5 years ago....those are incredible
speeds...i know of about 4 sites that use C+'s.

From: (Bob Stearns)
Subject: Re: 11! CDC/Cyperplus [1]

Actually, the University of Georgia had a CYBERPLUS for a while on
"spec". We were writing a PROLOG interpreter for it as well as beta
testing the software. We terminated the PROLOG project about 2/3s of the
way through when analysis of the memory management showed that, for the
PROLOG (or any LISP-type language), the speed would not be better than a
fast PC-AT.

The original machine, designed for black sites, was an integer only
machine designated the AFP (Advanced Flexible Processor). Apocryphally,
the orders were supposed to say "Drive to the warehouse at 14th and K, find
the key under the mat, drive the truck inside, leave the key on the seat
of the truck; leave, close the door, and do not look back." The truck,
with a check would show up two weeks later back at Minneapolis.

The processor was far too specialized to compete against either the
CYBER 20x line or the normal CYBER general purpose machines, so it died
an early death.

There were 15 functional units in the processor, all connected with
a crossbar. Each functional unit's input and output could be changed
on each 20 ns clock, but only 4 opcodes could be issued. All the
units without opcodes continued the previous instruction until instructed
otherwise. The instruction word was 200 bits, with 4K words of instruction
store; this was mostly due to the expense of the 7 ns cycle time memory used.

The only effective language was MICrocode Assembler; the FORTRAN was
in development, looking towards using the box as an attached processor
for the CYBER line rather than as an independent processor.

The interprocessor topology was a pair of synchronous, 16-bit, opposed
direction, rings. The clock cycle of the ring was the same as the clock
cycle of the processors, so you could read two 16 bit words and write
two 16 bit words per 20 ns per processor. The maximum number of
processors in a ring was hardware address limited at either 16 or 32,
my memory is suffering bit rot.

We had the advanced floating point processor which added 40 bits to the
instruction word, 4 more functional units, another crossbar and sundry
extra memory; still only 4K words of program memory.

The original application for which the processor was developed was picture
comparison. It could compare two 1K*1K pixel images per second while
correcting for translational, rotational and weather differences. It did
require a 9 member ring to achieve this performance as I remember it.

If you want further information, I still have a complete set of manuals
for both the product CYBERPLUS and black machine AFP.


--eugene miya, NASA Ames Research Center,
  Resident Cynic, Rock of Ages Home for Retired Hackers
  {uunet,mailrus,other gateways}!ames!eugene