By Don Clark
Companies that make computer chips are struggling with some of
the toughest times in the industry's history, thanks to slowing
demand for certain devices and diminishing performance gains from
making smaller circuitry.
Yet the pressures are spurring a renaissance of semiconductor
innovation, along with a growing band of startups aiming to exploit
it, industry executives say.
Big and small companies in the $335 billion global semiconductor
industry are pushing to develop new chip designs, materials and
manufacturing processes. One reason is the widening use of the
artificial-intelligence technique known as deep learning, in tasks
like classifying images, translating speech and autonomous driving,
which benefit from new calculating techniques.
Some of the new development efforts are aimed squarely at
disrupting entrenched incumbents like Intel Corp. which has
responded by modifying some of its own time-honored strategies.
"It is the best of times and the worst of times together," said
Dharmendra Modha, a chief scientist at International Business
Machines Corp. who is leading an effort to develop an unusual
brain-like chip.
The surge in activity stems partly from the waning growth in
sales of smartphones and personal computers, which has also led to
unprecedented consolidation among chip makers. Dealogic tallies 707
mergers and acquisitions totaling $246 billion over the past two
years.
At the same time, the industry's standby strategy for increasing
performance -- steadily shrinking the tiny transistors on each
sliver of silicon -- is running into diminishing returns. For
decades, chip makers followed Moore's Law, the roughly two-year
cadence for packing more transistors on chips, named after Intel
co-founder Gordon Moore. Lately, though, the gains in calculating
speed and power consumption have been less dramatic, while
designing chips with more transistors costs more.
The Semiconductor Industry Association and its research
affiliate have enlisted 22 tech companies to launch a broad study
of technologies that might bring computing advances. Alternatives
range from stacking circuitry in space-saving layers to making
chips from biological materials such as proteins.
Development is particularly intense in deep learning. The
technique involves training systems by exposing them to immense
quantities of data rather than programming them with explicit
instructions, which can take a long time and tends to produce
less-reliable results. Web companies using deep learning have taken
a particular interest in spurring the creation of hardware that can
get faster results.
Deep-learning systems often use a combination of Intel
processors and Nvidia Corp. or Advanced Micro Devices Inc. chips
that were originally designed to render videogame images. Those
chips comprise hundreds or thousands of simple processors
performing calculations simultaneously, compared with dozens of
sophisticated computing cores on a high-end Intel chip.
Some companies say even more specialized hardware is needed.
Alphabet Inc.'s Google unit recently took the unusual step of
designing its own chip from scratch for some deep-learning tasks.
IBM is targeting deep learning with TrueNorth, a chip unveiled in
2014 and composed of one million structures patterned after the
brain's neurons. Mr. Modha said it has shown startling acceleration
of deep-learning applications and is on track to create a "business
at scale" by 2019.
Venture capitalists have taken notice. Technological challenges
and stiff competition in the chip industry have led most venture
capitalists to put their money elsewhere. But some entrepreneurs
and investors see new opportunities to develop chips for markets
like networking, where customers may want to diversify their
sources rather than relying on one or two leading suppliers.
Cerebras Systems, a 25-employee company that plans to design
processors targeting deep learning, found it surprisingly easy to
raise venture-capital funds, said founder Andrew Feldman. "We
raised money in eight days," he said, declining to say how
much.
Other startups designing chips for deep learning include KnuEdge
Inc., Graphcore Ltd., Cornami and Wave Computing. Wave says a
system powered by a specialized processor it is developing can
complete a typical text analysis job in 6.75 seconds, compared with
69 minutes on a system using a combination of Intel Xeon and Nvidia
processors.
Intel, which supplies roughly 99% of the central processing
units, or CPUs, used in server systems, is in the crosshairs of
many of the new chip efforts. Compatible chips from AMD remain an
option, and some web giants have been testing IBM's Power chip
technology or designs from ARM Holdings PLC used in mobile
phones.
Other chip buyers are augmenting Intel's processors with
specialized chips for particular jobs. Microsoft Corp., for
example, recently said it was outfitting each server in its Azure
cloud service with chips called field programmable gate arrays, or
FPGAs, which can be configured after leaving the factory. The
company says the technology accelerates computing chores like
Microsoft's Bing search service while making the servers
communicate faster.
"I need more than what Intel and AMD are doing in CPUs," said
Doug Burger, a distinguished engineer in a Microsoft research group
spearheading the effort. "That is why we have this Cambrian
explosion of new processor architectures."
Intel is mounting its own offensives. The company paid $16.7
billion in late 2015 to buy Altera Corp., allowing it to sell FPGA
chips to Microsoft and others. This year Intel purchased Nervana
Systems and plans to add its deep-learning technology to its own
processors.
Intel's approach poses fewer programming hassles for customers
than shifting over to specialty chips that tend to be supplanted
over time, said Jason Waxman, vice president and general manager of
Intel's data center group.
"I'm excited to see a lot of innovation and ideas," he said. "I
tend to believe very few of them are going to last as discrete
products."
(END) Dow Jones Newswires
January 11, 2017 14:16 ET (19:16 GMT)
Copyright (c) 2017 Dow Jones & Company, Inc.
Intel (NASDAQ:INTC)
Historical Stock Chart
From Aug 2024 to Sep 2024
Intel (NASDAQ:INTC)
Historical Stock Chart
From Sep 2023 to Sep 2024