By Don Clark
Silicon Valley pioneer Gordon Moore laid out a bold theorem 50
years ago. Engineers would cram twice as many transistors on tiny
squares of silicon every year or so, producing more and more power
in ever-smaller machines.
His extrapolation, known as Moore's Law, has been one of the
most enduring precepts of the technology industry, foretelling the
revolutionary emergence of personal computers, mobile phones, Web
servers and network routers. Each generation of chips usually
brought more performance at a lower cost.
But Moore's Law is hitting some painful limits.
The design and testing of a chip with the latest technology now
costs $132 million, up 9% from the previous top-of-the-line chip,
estimates International Business Strategies Inc., a consulting firm
in Los Gatos, Calif. A decade ago, designing such an advanced chip
cost just $16 million. Meanwhile, some companies for the first time
are unable to reduce the cost of each tiny transistor.
The changes are triggered partly by the many new processing
steps needed to turn silicon wafers into the latest computer chips.
Circuitry for the latest chips has a width of 14 nanometers, or
billionths of a meter, which enables manufacturers to squeeze
hundreds of millions more transistors on a chip than they could in
the past. But designing products that use so many more components
takes lots of time and money.
While companies say they likely can keep shrinking the size of
silicon chips for another decade or so, that work is bringing
diminishing financial returns. Some chip designers already are
limiting their use of the newest technology to high-end products
where performance is more important than cost.
"We are being very careful," said Henry Samueli, co-founder,
chairman and chief technology officer at Broadcom Corp., which is
based in Irvine, Calif., and makes chips for about half the world's
tablets and smartphones. "The price of these chips is going up
dramatically."
Micron Technology Inc. Chief Executive Mark Durcan adds: "There
will be smaller and smaller pieces of the market that will pay for
the improvement." Micron, of Boise, Idaho, makes flash memory
chips, which are used in smartphones, digital cameras and tablets
to store photos.
Mr. Moore was director of the research and development
laboratories at Fairchild Semiconductor, a unit of Fairchild Camera
and Instrument Corp. and seminal Silicon Valley startup, when
Electronics magazine published his predictions on April 19, 1965
under the headline: "Cramming more components onto integrated
circuits."
He extrapolated that the number of components on a single
silicon chip would double every year from about 60 to as many as
65,000 by 1975. In 1975, he adjusted the formula to predict a
doubling every two years.
Fairchild sold its first transistors for $150 each, but prices
for the company and its rivals dropped year after year in the wake
of Mr. Moore's projection.
Semiconductor giant Intel Corp.'s Core i5 microprocessors
include 1.3 billion transistors, each costing $0.00000014, or a
penny per 70,000 transistors, according to the Santa Clara, Calif.,
company.
Mr. Moore, now 86 years old, didn't use the words "Moore's Law"
in his article, but they took hold as an axiom in Silicon Valley
and as general shorthand for just about any kind of progress,
technological or otherwise.
"I googled 'Moore's Law' and I googled 'Murphy's Law' and
'Moore' beats 'Murphy' by at least two to one," he said in a
January interview by Intel.
Mr. Moore co-founded Intel three years after his 1965 prediction
and retired as the company's chairman in 1997. He now lives in
Hawaii. Mr. Moore couldn't be reached, and Intel said he wasn't
available to comment.
At first, Moore's Law was largely a yardstick for chip
engineers. It gradually became a competitive imperative, spurring
companies to relentlessly innovate.
Until the mid-2000s, Moore's Law helped chip makers boost a key
aspect of computing performance known as operating frequency, or
clock speed. But higher clock speeds generated too much heat and
consumed too much power as the market shifted to portable computing
devices.
That led to a strategic shift by Intel and other chip makers,
which are using tricks like changing the shape of transistors to
make them switch faster and use less energy.
But the industry's costs keep rising, with new chip-fabrication
plants costing as much as $10 billion. Cost pressures led
International Business Machines Corp. last year to pay $1.5 billion
to another company to take over its semiconductor operations.
Companies that can afford to keep pushing Moore's Law are
finding it increasingly hard to keep up the pace. Intel's
introduction of 14-nanometer technology was two quarters late
because of delays in reducing manufacturing defects.
Intel said it is confident that the process will result in
greater savings per transistor than past advances. "It was a little
bit harder but has got us to a better place in the end," said Mark
Bohr, a senior Intel fellow who helps lead development of
production processes.
More transistors provide the biggest benefits when tasks can be
broken up and tackled by many processor cores at once. Nvidia Corp.
chips render ultrarealistic images on computer screens by
simultaneously painting colors on thousands of pixels.
Some big computer users are going beyond such chips to try other
new designs, since smaller transistors alone aren't boosting
computing speeds enough.
"Moore's Law is not having the same effect on the rate of gains
we are seeing," said Gordon MacKean, a senior director of Google
Inc.'s hardware platforms team.
Some makers of data-storage chips are taking more dramatic
steps. Producers of chips called NAND flash memory used in
smartphones and an increasing number of computers have decided to
stop shrinking transistors, worried that smaller circuitry won't
store data reliably.
Instead, they plan to stack circuits in three dimensions--32 or
48 layers per chip--rather than on a flat square of silicon to keep
boosting the capacity of their devices.
Micron and Intel expect to produce so-called 3-D NAND chips that
initially store as much as 384 gigabits of data, or three times
more than conventional memory chips.
Later this year, Intel expects to deliver a chip for specialized
applications with eight billion transistors--or 133 million times
more than chips than when Mr. Moore made his projection.
Access Investor Kit for Broadcom Corp.
Visit
http://www.companyspotlight.com/partner?cp_code=P479&isin=US1113201073
Access Investor Kit for Intel Corp.
Visit
http://www.companyspotlight.com/partner?cp_code=P479&isin=US4581401001
Access Investor Kit for Micron Technology, Inc.
Visit
http://www.companyspotlight.com/partner?cp_code=P479&isin=US5951121038
Access Investor Kit for NVIDIA Corp.
Visit
http://www.companyspotlight.com/partner?cp_code=P479&isin=US67066G1040
Subscribe to WSJ: http://online.wsj.com?mod=djnwires