Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!uunet!samsung!gem.mps.ohio-state.edu!uwm.edu!lll-winken!maddog!brooks
From: brooks@maddog.llnl.gov (Eugene Brooks)
Newsgroups: comp.arch
Subject: Re: Evans and Sutherland quits the superbusiness
Keywords: Killer Micros
Message-ID: <38980@lll-winken.LLNL.GOV>
Date: 20 Nov 89 00:32:01 GMT
References: <27611@dhw68k.cts.com> <38966@lll-winken.LLNL.GOV> <1989Nov19.155445.27287@hellgate.utah.edu>
Sender: usenet@lll-winken.LLNL.GOV
Reply-To: brooks@maddog.llnl.gov (Eugene Brooks)
Distribution: usa
Organization: Lawrence Livermore National Laboratory
Lines: 49

If your basic processor is not as fast as a current killer micro,
someone with ONE Killer Micro will blow your pants off.  This was
the case in an extreme for the ES-1 and is why Evans and Sutherland
saw no hope at all for the future and closed shop.  One can quote
the offical words printed in the press, but the bottom line is that
you can't make money selling an expensive slow computer.  Had there
been any money at the end of the tunnel, they would have hung in
there and solved their hardware and software problems.  The Killer
Micros were moving in the on the territory and will have completely
dominated it before they could get their problems fixed.

It took the use of on the order of 10 processors (processing
elements) or more on the ES-1 to match traditional supercomputer
performance.  As noted in an earlier post, the ES-1 was a nice
micro architecture, but without any of the Killer part in either
performance or cost.  Judging from the Livermore Loops figures for
the latest Killer Micro from hell, the MIPS R6000, it matches
traditional supercomputer performance with ONE processor, and not
just for "scalar only" codes.  The other micro vendors will soon follow
with even more terrible critters, its a competitive world after all.

Killer Micros are certainly custom architectures which cost a lot
of money to develop.  The difference for them is that their development
costs are amortized over large markets and the parts end up sold at
"cookie cutter" costs.  Compare this to the costs of developing the
Cray-3, mentioned by Rollwagen at SC'89 to be more than a hundred
million dollars.  This development cost has to be amortized over
the sale base, and for a market which might only be few tens
of machines (after SSI, Cray Computers, Cray Research, Tera,
Convex, and the three Japanese companies divide up the total market
which the Killer Micros leave for them) puts quite a lower bound
on the machine price before you even get around to charging the cost
of the hardware.  This is in sharp contrast to the situation for
the Cray-1, the sale of the first copy of which more than paid the
entire development cost for the machine.  Gone are the days of
high profit margins for supercomputers.

Perhaps I am wrong and the R6000 powered box should be sold for
more than a million and MIPS is just dumping it on the market to
destroy the supercomputer vendors.  I don't see any "anti dumping"
legislation in the works, however.  One thing is clear, if
traditional supercomputers don't find another order of magnitude
in single CPU performance real soon, at fixed cost, they will
not survive The Attack of the Killer Micros.    For scalar codes
supercomputers need even more leverage, two orders of magnitude.
I don't think it is going to happen.


brooks@maddog.llnl.gov, brooks@maddog.uucp