[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

IP: Neural Computer to Rival Brain???

From: [email protected]
Subject: IP: Neural Computer to Rival Brain???
Date: Thu, 08 Oct 1998 09:15:57 -0500
To: [email protected]

Source:  The Economist


Semiconductors - Silicon smarts 

BRAINS and computers are very different things.
Brains consist of a trillion or so tiny elements, called
neurons, which are individually dumb but collectively,
thanks to the thousand trillion connections between
them, very powerful. Most computers, on the other
hand, depend on a single, complex component-a
microprocessor-to get things done. Even the most
advanced supercomputers, with hundreds or even
thousands of connected microprocessors, cannot
match the compactness or connection density of the
human brain. 

Two new chip-making techniques being developed at
Irvine Sensors Corporation (ISC) in Costa Mesa,
California, could be significant steps in the
long-running effort to make more brain-like
computers. Researchers at ISC have found a way to
pack silicon chips extremely tightly together and,
better still, to make large numbers of connections
between them. 

Their technique layers silicon chips on top of each
other, cramming 50 chips into the space normally
occupied by just one. This is done by grinding away
the underside of the silicon wafer on which the chip
circuitry is built-a thick, non-functional platform that
can be removed without affecting the chip's
operation. The result is a paper-thin but fully
functional chip, which can be stacked and bonded
with other chips to form a single unit. The chips are
wired together via connectors along their edges, and
the whole sandwich is embedded in epoxy resin. 

This space-saving technique is already being used
commercially in a four-layer memory chip that packs
128 megabits of data into an amazingly small
one-centimetre-square package. Earlier this year, ISC
won a $1.3m military contract from Boeing to build a
wearable, voice-activated computer the size of a
pack of cards. 

But while such stacking wizardry means computers
can be smaller, it does not make them more
brain-like. At present, the nearest approximation to a
silicon brain involves making electronic circuits that
behave like neurons, and connecting them up in small
networks. Such "artificial neural networks" can be
used for everything from image recognition to credit
scoring, but their size and complexity-and so their
deductive power-is limited. 

This is because it is only possible to fit a certain
number of silicon neurons on to a single chip; and
there is a limit to the number of connections that can
be made between adjacent neural chips. Researchers
would like to be able to build networks that are
larger and more densely connected-in short, more
brain-like. ISC's second technology should let them
do this, by allowing direct vertical connections to be
made anywhere on the adjoining surfaces of adjacent
chips in a stack. 

To achieve this, half of a special component called a
three-dimensional field-effect transistor, or 3DFET, is
constructed at the site of each connection, as part of
the usual chip-making process. When the chips are
stacked, the two halves of each 3DFET fit together,
allowing signals to pass up and down from one chip
to the other. A prototype 3DFET, developed with
financing from the US army's ballistic missile defence
organisation, has already been made and tested. With
new funding, ISC hopes to make a chip-stack
connected using 3DFETs within 18 months. 

After that, says ISC's chief technical officer, John
Carson, the long-term goal is to stack 1,000 neural
chips in a single cube. This would involve thinning
each chip down to less than the thickness of a human
hair. Already, ISC has produced a prototype that is
almost this thin. 

If ISC can squeeze a million silicon neurons on to each
chip, and pack a thousand chips into a one-inch
neural cube, the arithmetic starts to get interesting. A
thousand such cubes, which could fit in a shoebox,
would contain a trillion neurons, and a hundred trillion
connections. That would still not match the
connectivity of human grey matter. But it would be
the most brain-like computer ever made. 
NOTE: In accordance with Title 17 U.S.C. section 107, this material is
distributed without profit or payment to those who have expressed a prior
interest in receiving this information for non-profit research and
educational purposes only. For more information go to:

To subscribe or unsubscribe, email:
     [email protected]
with the message:
     (un)subscribe ignition-point [email protected]