The Junkyard Powerhouse
------------------------------------------------------------------ This article first appeared in Forbes Magazine in September, 1998. ------------------------------------------------------------------
BUNCH OF DUTCH engineering students
interning at Hewlett-Packard's Palo Alto
laboratories in 1995 wanted to translate magnetic
resonance imaging data into a three-dimensional
map of the brain's arteries. HP gave them a
refrigerator-size gizmo made of junk
parts-processor chips with high defect rates. The
experimental machine, called a Teramac, ran 100
times faster than a top-end workstation, and the
students went home thrilled.
What is this Teramac? Tera means "trillion,"
signifying the designers' intention that the machine
perform a trillion operations per second. It is the
first computer purposely made out of
junk-processor and memory chips that "any sane
engineer would refuse to ship," in the words of
Philip Kuekes, 51, who helped to design it. The
software detects any hardware flaws and simply
steps around them.
Meet a manufacturing idea that may supplant a
concept at the very heart of modern life: the
concept that parts should be so precisely formed
that they become interchangeable. "We're
proposing to do the exact opposite," says Kuekes.
"Get rid of mechanical precision and use computer
power to make up for the imprecision." In the
Teramac, quantity substitutes for quality.
There is precedent for the Teramac approach-in
the peripherals business. Since flawless disk drives
are costly, it is cheaper to supply highly reliable
mainframe storage via arrays of inexpensive
redundant disks. If each file is stored in duplicate on
multiple disks, and if the multiple copies are
carefully compared whenever the file is retrieved,
the data can be made as close to error-free as you
want. It helps that small disk drives of the sort that
go into consumer products are getting dirt cheap.
The same kind of economics surfaces in the chip
business. Quantity is easy to come by: As Moore's
Law dictates, the transistor count on chips is
already in the millions and will someday reach into
the billions. But quality is costly. Producing
defect-free chips with ever more microscopic
detail will demand factories costing tens of billions
of dollars. One day it will cost too much to continue
the process, and Moore's Law will slow to a crawl.
Yet you could extend the Law's life quite a bit if you
could teach computers to tolerate defects.
HP got the defect-ridden behemoth to work by
having software make up for the failings of
hardware. First, the software spent a week of
processing time to find, map and catalog the
220,000 defects that sullied the hardware. For most
computers, a single one of those defects would
have been fatal.
Then the software set about rewiring the machine
to get around the trouble spots. It had the means to
do so because Teramac's chips, made by HP, came
with extra logic elements (called "gates")-backup
systems, as it were. "You could even clip wires or
purposely break a part and then reconfigure the
computer, and it would work fine," says James
Heath, a UCLA chemistry professor who
collaborated on Teramac.
You might wonder how a machine built of defective
parts could correctly test itself for defects in the first
place. It turns out that 7% of its parts-the ones that
diagnose and correct errors-do have to be perfect.
So why not switch immediately to these self-fixing
computers? Because it's not economical yet.
Teramac relies on especially large chips, which use
more silicon and are therefore expensive. For now,
it's still cheaper to junk defective Pentiums when
they come out of the factory, and use only the
perfect ones.
As a result, the Teramac architecture is most likely
to find its first applications in niche markets that
require bigger chips. One likely prospect: digital
cameras, which need a large light-sensing chip to
get a sharp picture. Other applications include
special-purpose computing, as in a television
set-top box or in medical electronic instruments.
Because Teramac rewires itself readily, it can
optimize its hardware, first for one task, then
another. In other words, it can substitute hardware
for software.
Further down the road lies perhaps Teramac's
most promising application, as an enabling
technology for super-small circuitry. Such
nanoelectronic devices, as they are called, still
smack of science fiction, with parts that aren't
etched but are self-assembled, almost like living
things. The hoped-for result would be a
workstation-standard computer smaller than a
grain of sand.
You could never guarantee the perfection of a
trillion transistors that had been cooked up in this
fashion. You'd need a computer architecture that
could live with the defects.
Says Heath, who first did work on nanocomputer
components in 1992: "Teramac is very significant;
we would have been wandering in the woods
without it. We'd have continued making wiring and
gates in beakers without knowing how to assemble
them into a computer."
First on the agenda is getting Teramac to pay for
itself. If the cost of perfect parts rises fast enough, it
will make sense to use a larger quantity of junkyard
parts. It could happen in as little as five years, says
Stanley Williams, a physical chemist who has used
Teramac in his own research.