Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!ucbvax!agate!saturn!ucscc.UCSC.EDU!haynes
From: haynes@ucscc.UCSC.EDU (Jim Haynes)
Newsgroups: comp.misc
Subject: Re: Looking for Computer Folklore
Summary: more of the same machine sold at two different prices
Message-ID: <6230@saturn.ucsc.edu>
Date: 4 Feb 89 06:38:58 GMT
References: <7143@pyr.gatech.EDU> <532@geovision.UUCP> <4575@tekgvs.GVS.TEK.COM> <319@itcatl.UUCP> <2986@ficc.uu.net>
Sender: usenet@saturn.ucsc.edu
Reply-To: haynes@ucscc.UCSC.EDU (Jim Haynes)
Organization: California State Home for the Weird
Lines: 55

The GE 625/635 originally had no logic to detect unassigned op codes.
This, needless to say, was not a happy situation for reliability
with users constantly running new, untested programs.  About the same
time we were designing logic to detect and trap these op codes, one
of the engineers decided to do a systematic study to see what they
all did.  On at least one system there was at least one op code that
would repeatably cause a circuit breaker to trip in one of the power
supplies.

The 625 was convertible to the twice-as-fast 635 by removing one wire.
It wasn't originally planned to be this way; originally the 625 was to
be sold with 2 microsecond memory, while the 635 was to use 1 microsec
memory, which at the time was much more expensive.  But by the time
the systems got into full production the price of 1 microsec memory had
come down to match the price of 2 microsec memory.  Only a few of the
slower machines had been sold, and the maintenance organization didn't
want to have to stock parts for so many different kinds of memory;
so they said use one microsec memory for everything and just slow it
down for the slower machine.  (Memory was bought from two or three
outside vendors, so there were already enough different kinds of memory
parts to stock!).

Then there was an even slower 615, which was maybe a dozen wires
different from the 625/635.  It was made to suit marketing's need
for a competitor to a particular IBM 360 model number.  Simply
slowing the machine didn't satisfy marketing, as it was too fast on
some operations and too slow on others.  So there had to be several
changes to the timing logic to make it just slow enough but not too
slow.  These were easy to do because the machine didn't use a clock
oscillator; timing was controlled by pulses going through a number of
tapped delay lines.  So it was a matter of choosing the right taps
on each delay line.

Quite a number of the 635s in the field used two independent memory
controllers, each covering half the address range.  We tried a two-wire
change to allow these machines to run with interleaved memories.  It
turned out the speedup was only a couple of percent, so we didn't release
it to the field.  The logic was simply so well optimized for 1 microsec
memory that increasing the memory bandwidth by itself didn't gain
anything.

As long as we're talking about selling the same machine for two
different prices, there was the slightly different stunt advertised
by Amdahl some years back.  They had a switch out there for the customer
to use that controlled the speed of the machine.  I guess it was on rental
machines that the rate you paid depended on the position of the switch.
So they advertised that you could run the machine at the lower rate
so long as performance was adequate; and if you got behind on the workload
you could flip the switch and pay more to get more power.
haynes@ucscc.ucsc.edu
haynes@ucscc.bitnet
..ucbvax!ucscc!haynes

"Any clod can have the facts, but having opinions is an Art."
        Charles McCabe, San Francisco Chronicle