Tuesday20 August 2019 ★ † FINANCIAL TIMES 13computing memory that sits alongside
the chip’s many processing cores, so that
they can handle tasks with the mini-
mum possible delay and use only a tiny
amount of energy to shuttle information
back and forth.
The next task is to link cores in a
matrix pattern so that they can commu-
nicate with each other, like the synapses
in the brain. By connecting 400,
cores, Cerebras claimed it was taking
this process to its ultimate level while
still keeping the efficiency of handling
all the processing on a single chip.
Intel, which bought Nervana, the AI
chip start-up, four years ago as the AI
race was getting started, has been trying
to achieve the same effect by network-
ing together many individual chips in a
vast array.
Thathad brought a breakthrough in
the past six months andwould bring
much greater efficiency to systems due
out later this year, saidAmir Khosrow-
shahi, the chief technology officer of
Intel’s AI efforts.
However, even if Intelsucceeded in
linking thousands of cores into a huge
matrix-like system, itwas unlikely to be
as efficient as the Cerebras chip thanks
to the inbuilt advantage that the com-
panyhad of placing everything on a sin-
gle chip, said Mr Moorhead.ple compared with the cores in general-
purpose CPUs, which have to handle
many different computing tasks. But
the chips’ makers must find ways to har-
ness vast numbers of cores in ways that
speed up the time it takes to train a large
deep learning model, while saving on
electricity, hence cost.
Taking the challenge to its logical
extreme, Cerebras has carved a single,
square chip out of a 300mm diameter
circular wafer, the largest silicon disc
that can be produced in today’s chip
“fabs”, or factories.
Wafers are normally sliced up into
dozens of individual chips, because the
technology does not exist to etch cir-
cuitry into anything bigger than a large
postage stamp-sized area. Cerebras has
sought to overcome this limitation by
connecting the many different sectors
on wafers, known as dies, enabling them
to communicate directly with each
other, in effect turning the entire silicon
plate into a massive processor.
Most of the companies building spe-
cialised deep learning chips have
adopted designs that push data intowhich are used in deep learning systems
to handle tasks such as image recogni-
tion and language understanding.
The networks operate as giant feed-
back loops, recycling information as
they learn to find patterns in the form-
less data.
The computing “cores”, or brains, in
the chips needed for this work are sim-RICHARD WATERS— SAN FRANCISCOThe race among semiconductor makers
to gain an edge in the booming market
for specialised artificial intelligence
processors hasproduced the biggest
computer chip.
While chip circuitry continues to get
smaller, the slab of silicon, developed by
Cerebras, a Californian start-up, has a
surface area slightly larger than a stand-
ard iPad and is more than 56 times
bigger than its closest competitor.
It eats up as much electricity as all the
servers in one-and-a-half racks — the
towers of computers in data centres that
stand more than six feet tall.
The chip, due to be unveiled yester-
day after nearly four years of develop-
ment, is the starkest sign yet of how
traditional thinking is being turned on
its head as the chip industry struggles
with the demands of AI.
It highlights leaps in the computing
power being thrown at complex AI
problems, something that prompted
OpenAI, the US research group, to raise
$1bn fromMicrosoftlast month, hoping
to ride the exponential hardware curve
toreach human-level AI.
Most chipmakers have been looking
to create smaller, modular elements
known as “chiplets”, out of which
today’s most advanced chips are assem-
bled, said Patrick Moorhead, a US chip
analyst. Cerebrashas jettisoned that
approach andcome up with what is in
effect an entire computing cluster on a
single chip.
The race to build a new generation of
specialised AI chips, under way for sev-
eral years, isreaching a critical point,
with several companies — including
Intel,Habana LabsandGraphcore,aUK
start-up — either just starting or promis-
ing to deliver their first chips to custom-
ers before the end of this year.
Cerebras did not identify what it said
was a number of customersreceiving its
chips, although they are likely to be best
suited to the large-scale tasks under-
taken by the biggest internet compa-
nies.
More than 50 companies have been
trying to develop specialised chips for
AI. Most of these are used for inference,
the task of applying a trained AI system
to real-world examples, rather than the
far more data-intensive job of training
the deep learning models in the first
place. That challenge has been taken on
by a handful of start-ups such as Cere-
bras, Graphcore andWave Computing,
as well asCambriconof China.
The length of time it has taken for
companies such as these to start ship-
ping products shows that the technical
challenges were much greater than
most had expected, said Linley Gwen-
nap, principal analyst at the Linley
Group, a US chip research firm.
That has not prevented some of the
productless start-ups from attracting
high valuations.
Cerebras had raised more than
$200m in venture capital, with its latest
round, late last year, valuing it at about
$1.6bn, saidAndrew Feldman, chief
executive.
The need for a new type of processor
for AI stems from the massive amounts
of data needed to train neural networks,COMPANIES
AI race spawns chip bigger than an iPad
California start-up’s silicon slab underlines how traditional thinking in the sector is being turned on its head
Projected growth of deep
learning chipsets
Global revenue ( bn)Source: Tractica ROBERT ARMSTRONG— NEW YORKThe inversion of theyield curve — the
market indicator that has rattled Wall
Street — has had a particularly pro-
nounced effect on US regional banks,
whose shares are among the worst per-
formers of the past few weeks.So-called asset-sensitive banks, those
with a business heavily weighted
towards floating rate loans such as busi-
ness loans, face a hit to their profits
because of the slump in yields, which
pulls down the interest they will receive.
Another result is that in some
cases, their shares appear to have
started trading in lockstep with changes
in the yield curve.
“The moves up and down on a daily
basis [in response to] the curve have
been so violent and so quick that I’m not
sure it is traditional asset managers that
are causing this,” saidAnton Schutz,
who manages a portfolio of small-cap
bank stocks atMendon Capital.
He suspects that macro hedge funds
and algorithmic traders have the banks
in their sights. “Machines are trading
this,” he said.
The yield curvedescribes the yields
on different maturities of US govern-
ment debt, and its inversion — when
rates on long-term bonds fall below
those on short-term debt — has been
a reliable indicator of recessions in
the past.It also makes it much harder for asset-
sensitive banks to make money, since
they cannot offset falling income from
floating rate loans by cutting the rates
they pay depositors, since these are
already at rock bottom. Larger banks
can have other fee-based businesses
such as wealth management or pay-
ments that make their profits less sensi-
tive to the yield curve.
Shares in institutions including
Silicon Valley Bank,ComericaandCitizens, Wall Street darlings just a year
or two ago, have traded in lockstep with
the change in the relationship between
three-month Treasury bills and the 10-
year note.
That relationship first inverted in
March, meaning the three-month
offered a higher rate of interest, and has
been consistently and ever more deeply
negative since May. At the close on Fri-
day, the 10-year was yielding 32 basis
points less than three-month bills.
Silicon Valley Bank is one of the hot-
test growth stories in the industry. It
specialises in serving young tech com-
panies, and has a deep base of non-inter-est bearing deposits, as start-ups hold
their venture financing or IPO proceeds
with the bank. The asset side of the bal-
ance sheet is dominated by floating-rate
business loans.
Comerica is similarly fast growing,
and concentrated in corporate deposits
and lending. Shares in both banks tri-
pled between 2016 and 2018, while the
yield curve was broadly stable, but have
sold off since late last year, as prospects
for Federal Reserve rate hikes have
dimmed. Other banks that have fallen
into line with the fluctuations in the
curve includeZion Bancorp,Regions
Bank,Bank OZK, andPeople’s United.
Bank investors and executives argue
that opportunistic traders are overlook-
ing the asset-sensitive banks’ capacity to
adjust their balance sheets and increase
profits despite an inverted yield curve.
Bruce van Saun, chief executive at Cit-
izens, whose shares have tumbled 14 per
cent since late July, said that while it has
a high concentration of floating-rate
assets on its balance sheet, lending mar-
gin was not the only contributor to earn-
ings growth. “Fee growth is positive,
expense control is positive, and credit
[quality] is positive... three out of four
profit [drivers] are working — so we can
hang in there and have a good year.”
Citizens, like many other banks, has
been adding interest-rate hedges to
limit the impact of further declines in
short-term rates.Financials
Yield curve pain for regional US banks
Cerebras has connected the sectors on wafers, enabling them to communicate directly with each other, turning the entire silicon plate into a massive processorThe chip is more than 80 times
larger than its closest rival‘The moves up and down
have been so violent that
I’m not sure it is traditionalasset managers causing this’
RELEASED
those on short-term debt — has beenRELEASED
those on short-term debt — has been
a reliable indicator of recessions inRELEASED
a reliable indicator of recessions inBYrates on long-term bonds fall below
BYrates on long-term bonds fall below
those on short-term debt — has beenthose on short-term debt — has beenBY"What'son different maturities of US govern-"What'son different maturities of US govern-
ment debt, and its inversion — when
"What'sment debt, and its inversion — when
rates on long-term bonds fall belowrates on long-term bonds fall below"What'sNews"yield curve
News"yield curve
on different maturities of US govern-on different maturities of US govern-News"VK.COM/WSNWSon different maturities of US govern-VK.COM/WSNWSon different maturities of US govern-
ment debt, and its inversion — whenVK.COM/WSNWSment debt, and its inversion — when
rates on long-term bonds fall belowVK.COM/WSNWSrates on long-term bonds fall below
those on short-term debt — has beenVK.COM/WSNWSthose on short-term debt — has been
a reliable indicator of recessions in
VK.COM/WSNWSa reliable indicator of recessions in
the past.the past. VK.COM/WSNWSTELEGRAM:a reliable indicator of recessions inTELEGRAM:a reliable indicator of recessions in
the past.TELEGRAM:the past.t.me/whatsnwson different maturities of US govern-t.me/whatsnwson different maturities of US govern-
ment debt, and its inversion — whent.me/whatsnwsment debt, and its inversion — when
rates on long-term bonds fall belowt.me/whatsnwsrates on long-term bonds fall below
those on short-term debt — has been
t.me/whatsnwsthose on short-term debt — has been
a reliable indicator of recessions ina reliable indicator of recessions int.me/whatsnws