By Don Clark
SAN JOSE, Calif.-- International Business Machines Corp. is
claiming a major advance in emulating the brain in silicon.
The technology company has developed a microchip that simulates
functions of neurons, synapses and other features of the brain to
perform calculations. IBM says the chip, a sharp break from the
fundamental design used in most computers, excels at chores like
recognizing patterns and classifying objects while using much less
electrical power than conventional hardware.
IBM's new chip is the latest in a series of efforts by the
company and others to design brain-like chips as traditional chip
manufacturing yields fewer dramatic breakthroughs. But its latest
offering, described in a paper in the journal Science, has novel
features that include its large size and the use of standard
digital technology rather than esoteric materials or production
processes.
Though it is giving few details on timing, IBM says it is
already talking to potential partners about ways to bring the chip
to market. The company has connected multiple chips together to
test potential system designs, and sees applications of the
technology ranging from room-sized supercomputers to floating
jellyfish-shaped devices that could sense tsunamis or other aquatic
conditions.
"We have huge commercial ambitions," said Dharmendra Modha, a
researcher at IBM's Almaden Research Center here whose titles
include chief scientist for brain-inspired computing.
The chip, dubbed TrueNorth, was built for IBM by Samsung
Electronics Co. using the same manufacturing technology the South
Korean company uses to make microprocessors for smartphones and
other mobile devices. IBM collaborated on the underlying design
with researchers at the New York campus of Cornell University in a
project that has received $53 million in funding since 2008 from
the Pentagon's Defense Advanced Research Projects Agency.
IBM's announcement comes as scientists and engineers are
pondering the prospect of slowing improvements of conventional
microprocessors. Historically, manufacturers steadily have shrunk
chip circuitry, reducing costs while improving calculating speed
and reducing power consumption.
But chip manufacturers can no longer rely on traditional
processes to produce dramatic improvements, particularly for
scientists grappling with supercomputers whose power needs already
approach those of small cities. Radically different chip
architectures like TrueNorth may help.
"Power is the fundamental constraint as we move forward," says
Horst Simon, deputy director of Lawrence Berkeley National
Laboratory, a major supercomputer user. "This chip is an indication
that we are really at the threshold of a fundamental change in
architecture."
The underlying design used in most computers and microprocessors
since the 1940s--named after mathematician John von
Neumann--separates components that carry out calculations from
memory circuity that stores data. Bits are shuttled between those
components through a conduit called a bus, with activity
synchronized by an internal clock. The scheme works well for tasks
like adding repeated sets of numbers, Mr. Modha said. And chips
have become much faster for such jobs as manufacturers have
increased the frequency of the clock's timing pulses.
But that trend also tends to boost a chip's power consumption.
Moving data back and forth over a bus, meanwhile, tends to slow
calculations, he said.
Brains, by contrast, are compact and particularly efficient at
chores like recognizing a person's face or distinguishing one sound
from another, Mr. Modha said. Cells called neurons process and
transmit information that is stored nearby, connected by structures
called synapses.
TrueNorth, IBM says, uses 5.4 billion transistors--four times
more than a typical PC processor--to yield the equivalent of one
million neurons and 256 million synapses. They are organized into
4,096 structures called "neurosynaptic cores," each able to store,
process and transmit data to any other using a communications
scheme called a crossbar.
The design is "event-driven," Mr. Modha said. That means that
individual cores fire up only when they are needed, rather than
running all the time.
This scheme makes the chips more power-efficient. Where a
comparable standard microprocessor draws 50 to 100 watts per square
centimeter, TrueNorth draws just 20 thousandths of a watt, IBM
says.
In a demonstration, Mr. Modha showed how the technology used
with a video camera atop a building can pick out and track people
walking below. Besides daisy-chaining TrueNorth chips to make large
systems, IBM expects to distribute a smaller, simpler chip for
applications where space is paramount or humans can't easily go.
Potential applications described by the company include ball-sized
rolling robots with cameras to inspect disaster sites and
leaf-sized sensing devices that could be scattered during a forest
fire.
"It's not going to replace conventional computers," Mr. Modha
says. "It is a complementary relationship."
Other companies, including chip giants Intel Corp. and Qualcomm
Inc., have their own designs for what engineers call "neuromorphic"
chips. Once working devices have been created, engineers face the
equally daunting task of persuding programmers to learn new methods
for creating useful software.
IBM has tried to address the obstacles, developing a special
programming language and tools to simulate such chips. The
technology "is much closer to being useable than a lot of other
neuromorphic systems others have developed," said Rajit Manohar, a
longtime collaborator who is professor of electrical and computer
engineering at Cornell Tech.
But other experts say it is too early to identify front-runners
in the field. One is Jeff Hawkins, a mobile-device pioneer and
co-founder of Numenta, a startup has been building hardware and
software based on a lengthy study of the brain.
Mr. Hawkins believes that largely two-dimensional chips like
TrueNorth will give way to stacks of chips or other approaches that
more closely emulate the brain's many connections. It's "a
many-year process to find out what the right neural architecture
is," he said.
Write to Don Clark at don.clark@wsj.com
Subscribe to WSJ: http://online.wsj.com?mod=djnwires