Google Isn't Playing Games With New Chip
May 18 2016 - 3:26PM
Dow Jones News
By Robert McMillan
When Google's AlphaGo computer program bested South Korean Go
champion Lee Se-dol in March, it took advantage of a secret weapon:
a microprocessor chip specially designed by Google. The chip sped
up the Go-playing software, allowing it to plot moves in the
time-limited match and look further ahead in the game.
But the processor, built in secret over the last three years and
announced on Wednesday, plays a more strategic role in the company.
Google, a division of Alphabet Inc., has been using it for more
than a year to accelerate artificial intelligence applications as
the software techniques known as machine learning become
increasingly important to its core businesses. Overall the chip,
known as the Tensor Processing Unit, is 10 times faster than
alternatives Google considered for this work, the company said.
The company has been rumored to have been working on its own
chip designs, but Wednesday's announcement marked the first time it
confirmed such an effort.
Google isn't the only firm speeding up artificial intelligence
with new chip designs. Microsoft Corp. is using programmable chips
called Field Programmable Gate Arrays to accelerate AI
computations, including those used by its Bing search engine.
International Business Machines Corp. designed its own
brain-inspired chip called TrueNorth that is currently being tested
by Lawrence Livermore National Laboratory.
Nvidia Corp. has been pushing its chips, known as graphical
processing units, into artificial intelligence as well. GPUs are
designed to render videogame images on personal computers but have
turned out to be well suited to performing calculations used by
machine learning applications.
Google, which relies mainly on standard Intel Corp. processors
for most computing jobs, also has used Nvidia GPUs for artificial
intelligence calculations including its early tests of the AlphaGo
software.
Large companies have begun using new processor designs to
augment general-purpose processors as the pace of improvement in
that field has slowed, said Mark Horowitz, a professor of
electrical engineering at Stanford University.
"They're not doing this to replace the Intel processors," he
said. "These are addendums to these processors."
Google and Apple Inc. lately have been aggressively hiring chip
designers and engineers, Mr. Horowitz said. Apple launched its own
chip-making effort around 2009 to improve the power and efficiency
of its devices and develop new features.
Google believes its new chip will give it a seven-year advantage
-- roughly three processor generations -- over currently available
processors when it comes to machine learning. That is important
because Google is betting its future on such software. It uses
machine learning in more than 100 programs for applications
including search, voice recognition, and self-driving cars. Such
programs require intensive calculation, and supplying the
processing power and electricity to do this math quickly is
expensive.
"In order to make them feasible to roll out, economically, with
the required latency for users and all that stuff, we looked around
at the existing alternative and we decided that we needed to do our
own custom accelerators," said Norman Jouppi, a distinguished
hardware engineer at Google.
It is unclear how much of Google's overall computation runs on
its new processors. Mr. Jouppi said Google uses more than 1,000 of
the chips, but he wouldn't say whether that meant the company was
buying fewer processors from vendors such as Intel or Nvidia.
"We're still buying literally tons of CPUs and GPUs," he said.
"Whether it's a ton less than we would have otherwise, I can't
say."
Google began using the Tensor Processing Unit in April 2015 to
speed up its StreetView service's reading of street signs. It
allows the company to process all the text stored in its massive
collection of StreetView images -- things such as street signs and
address numbers attached to the sides of houses -- in just five
days, much faster than previous methods, Mr. Jouppi said.
The chip also is used in Google search ranking, photo
processing, speech recognition, and language translation. The
company plans to make the chips available as part of its Google
Cloud Platform computing-on-demand service, he said.
TPU chips are soldered onto cards that slide into the disk-drive
slots in Google's standard servers, where they handle the
specialized calculations required by Google's machine learning
software.
--Don Clark contributed to this article
Write to Robert McMillan at Robert.Mcmillan@wsj.com
(END) Dow Jones Newswires
May 18, 2016 15:11 ET (19:11 GMT)
Copyright (c) 2016 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