MORE THAN

MOORE BY M. MITCHELL WALDROP

1 4 4 | N AT U R E | VO L 5 3 0 | 1 1 F E B R UA RY 2 0 1 6
© 2016 Macmillan Publishers Limited. All rights reserved
FEATURE NEWS

THE SEMICONDUCTOR INDUSTRY

City. The Semiconductor Industry Association (SIA) in Washington DC,
which represents all the major US firms, has already said that it will cease
its participation in the road-mapping effort once the report is out, and

WILL SOON ABANDON ITS PURSUIT

will instead pursue its own research and development agenda.
Everyone agrees that the twilight of Moore’s law will not mean the
end of progress. “Think about what happened to airplanes,” says Reed.

OF MOORE’S LAW.
“A Boeing 787 doesn’t go any faster than a 707 did in the 1950s — but
they are very different airplanes”, with innovations ranging from fully
electronic controls to a carbon-fibre fuselage. That’s what will happen

NOW THINGS COULD GET A LOT with computers, he says: “Innovation will absolutely continue — but it
will be more nuanced and complicated.”

MORE INTERESTING. LAYING DOWN THE LAW

The 1965 essay1 that would make Gordon Moore famous started with
a meditation on what could be done with the still-new technology of

N
integrated circuits. Moore, who was then research director of Fairchild
ext month, the worldwide semiconductor industry will formally Semiconductor in San Jose, California, predicted wonders such as
acknowledge what has become increasingly obvious to every­ home computers, digital wristwatches, automatic cars and “personal
one involved: Moore’s law, the principle that has powered the portable communications equipment” — mobile phones. But the heart
information-technology revolution since the 1960s, is nearing of the essay was Moore’s attempt to provide a timeline for this future.
its end. As a measure of a microprocessor’s computational power, he looked at
A rule of thumb that has come to dominate computing, Moore’s transistors, the on–off switches that make computing digital. On the
law states that the number of transistors on a microprocessor chip basis of achievements by his company and others in the previous few
will double every two years or so — which has generally meant that years, he estimated that the number of transistors and other electronic
the chip’s performance will, too. The exponential improvement that components per chip was doubling every year.
the law describes transformed the first crude home computers of the Moore, who would later co-found Intel in Santa Clara, California,
1970s into the sophisticated machines of the 1980s and 1990s, and from underestimated the doubling time; in 1975, he revised it to a more real­
there gave rise to high-speed Internet, smartphones and the wired-up istic two years2. But his vision was spot on. The future that he predicted
cars, refrigerators and thermostats that are becoming prevalent today. started to arrive in the 1970s and 1980s, with the advent of micropro­
None of this was inevitable: chipmakers deliberately chose to stay on cessor-equipped consumer products such as the Hewlett Packard hand
the Moore’s law track. At every stage, software developers came up with calculators, the Apple II computer and the IBM PC. Demand for such
applications that strained the capabilities of existing chips; consumers products was soon exploding, and manufacturers were engaging in a
asked more of their devices; and manufacturers rushed to meet that brisk competition to offer more and more capable chips in smaller and
demand with next-generation chips. Since the 1990s, in fact, the semi­ smaller packages (see ‘Moore’s lore’).
conductor industry has released a research road map every two years to This was expensive. Improving a microprocessor’s performance
coordinate what its hundreds of manufacturers and suppliers are doing meant scaling down the elements of its circuit so that more of them
to stay in step with the law — a strategy sometimes called More Moore. could be packed together on the chip, and electrons could move between
It has been largely thanks to this road map that computers have followed them more quickly. Scaling, in turn, required major refinements in
the law’s exponential demands. photo­lithography, the basic technology for etching those microscopic
Not for much longer. The doubling has already started to falter, elements onto a silicon surface. But the boom times were such that this
thanks to the heat that is unavoidably generated when more and more hardly mattered: a self-reinforcing cycle set in. Chips were so versa­
silicon circuitry is jammed into the same small area. And some even tile that manufacturers could make only a few types — processors and
more fundamental limits loom less than a decade away. Top-of-the- memory, mostly — and sell them in huge quantities. That gave them
line microprocessors currently have circuit features that are around enough cash to cover the cost of upgrading their fabrication facilities,
14 nanometres across, smaller than most viruses. But by the early 2020s, or ‘fabs’, and still drop the prices, thereby fuelling demand even further.
says Paolo Gargini, chair of the road-mapping organization, “even with Soon, however, it became clear that this market-driven cycle could not
super-aggressive efforts, we’ll get to the 2–3-nanometre limit, where sustain the relentless cadence of Moore’s law by itself. The chip-making
features are just 10 atoms across. Is that a device at all?” Probably not process was getting too complex, often involving hundreds of stages,
— if only because at that scale, electron behaviour will be governed by which meant that taking the next step down in scale required a net­
quantum uncertainties that will make transistors hopelessly unreliable. work of materials-suppliers and apparatus-makers to deliver the right
And despite vigorous research efforts, there is no obvious successor to upgrades at the right time. “If you need 40 kinds of equipment and only
today’s silicon technology. 39 are ready, then everything stops,” says Kenneth Flamm, an economist
The industry road map released next month will for the first who studies the computer industry at the University of Texas at Austin.
time lay out a research and development plan that is not centred on To provide that coordination, the industry devised its first road map.
Moore’s law. Instead, it will follow what might be called the More The idea, says Gargini, was “that everyone would have a rough estimate
than Moore strategy: rather than making the chips better and letting of where they were going, and they could raise an alarm if they saw
the applications follow, it will start with applications — from smart­ roadblocks ahead”. The US semiconductor industry launched the map­
phones and supercomputers to data centres in the cloud — and work ping effort in 1991, with hundreds of engineers from various companies
downwards to see what chips are needed to support them. Among working on the first report and its subsequent iterations, and Gargini,
ILLUSTRATION BY REBECCA MOCK

those chips will be new generations of sensors, power-management then the director of technology strategy at Intel, as its chair. In 1998,
circuits and other silicon devices required by a world in which com­ the effort became the International Technology
puting is increasingly mobile. [Link] Roadmap for Semiconductors, with participa­
The changing landscape, in turn, could splinter the industry’s long To hear more about tion from industry associations in Europe, Japan,
tradition of unity in pursuit of Moore’s law. “Everybody is struggling what will come after Taiwan and South Korea. (This year’s report, in
with what the road map actually means,” says Daniel Reed, a computer Moore’s law, visit: keeping with its new approach, will be called the
scientist and vice-president for research at the University of Iowa in Iowa [Link]/nppjyx International Roadmap for Devices and Systems.)
1 1 F E B R UA RY 2 0 1 6 | VO L 5 3 0 | N AT U R E | 1 4 5
© 2016 Macmillan Publishers Limited. All rights reserved
NEWS FEATURE

But in the early 2000s, when the features began to shrink below about

SOURCE: TOP, INTEL; BOTTOM, SIA/SRC

MOORE’S LORE 90 nanometres, that automatic benefit began to fail. As electrons had
to move faster and faster through silicon circuits that were smaller and
smaller, the chips began to get too hot.
For the past five decades, the number of transistors per microprocessor That was a fundamental problem. Heat is hard to get rid of, and no
chip — a rough measure of processing power — has doubled about every
two years, in step with Moore’s law (top). Chips also increased their ‘clock
one wants to buy a mobile phone that burns their hand. So manufac­
speed’, or rate of executing instructions, until 2004, when speeds were turers seized on the only solutions they had, says Gargini. First, they
capped to limit heat. As computers increase in power and shrink in size, a stopped trying to increase ‘clock rates’ — how fast microprocessors
new class of machines has emerged roughly every ten years (bottom).
execute instructions. This effectively put a speed limit on the chip’s
1010 electrons and limited their ability to generate heat. The maximum clock
rate hasn’t budged since 2004.
Second, to keep the chips moving along the Moore’s law performance
10 8 curve despite the speed limit, they redesigned the internal circuitry so
that each chip contained not one processor, or ‘core’, but two, four or
10 6 more. (Four and eight are common in today’s desktop computers and
smartphones.) In principle, says Gargini, “you can have the same output
with four cores going at 250 megahertz as one going at 1 gigahertz”. In
10 4
practice, exploiting eight processors means that a problem has to be
Transistors per chip broken down into eight pieces — which for many algorithms is dif­
10 2 ficult to impossible. “The piece that can’t be parallelized will limit your
improvement,” says Gargini.
1
Even so, when combined with creative redesigns to compensate for
electron leakage and other effects, these two solutions have enabled
Clock speeds (MHz) chip manufacturers to continue shrinking their circuits and keeping
10 –2 their transistor counts on track with Moore’s law. The question now is
1960 1974 1988 2002 2016
what will happen in the early 2020s, when continued scaling is no longer
possible with silicon because quantum effects have come into play. What
comes next? “We’re still struggling,” says An Chen, an electrical engineer
1013
who works for the international chipmaker GlobalFoundries in Santa
1012 me Clara, California, and who chairs a committee of the new road map that
in fr a
10 11
Ma is looking into the question.
That is not for a lack of ideas. One possibility is to embrace a
1010 te r
m pu completely new paradigm — something like quantum computing,
ic o
10 9
Mi
n which promises exponential speed-up for certain calculations, or
10 8 neuro­morphic computing, which aims to model processing elements
l
na on neurons in the brain. But none of these alternative paradigms has
10 7 r so er
Size (mm3)

Pe p u t
co
m made it very far out of the laboratory. And many researchers think that
10 6
quantum computing will offer advantages only for niche applications,
10 5
p rather than for the everyday tasks at which digital computing excels.
pto “What does it mean to quantum-balance a chequebook?” wonders
10 4 La
e John Shalf, head of computer-science research at the Lawrence Berkeley
10 3 on
r tph National Laboratory in Berkeley, California.
100 S ma ed
dd
b e or s
10 Em ce s s
o
MATERIAL DIFFERENCES
pr A different approach, which does stay in the digital realm, is the quest
1
to find a ‘millivolt switch’: a material that could be used for devices at
0.1
1950 1960 1970 1980 1990 2000 2010 2020
least as fast as their silicon counterparts, but that would generate much
less heat. There are many candidates, ranging from 2D graphene-like
compounds to spintronic materials that would compute by flipping
electron spins rather than by moving electrons. “There is an enormous
“The road map was an incredibly interesting experiment,” says research space to be explored once you step outside the confines of the
Flamm. “So far as I know, there is no example of anything like this in established technology,” says Thomas Theis, a physicist who directs the
any other industry, where every manufacturer and supplier gets together nanoelectronics initiative at the Semiconductor Research Corporation
and figures out what they are going to do.” In effect, it converted Moore’s (SRC), a research-funding consortium in Durham, North Carolina.
law from an empirical observation into a self-fulfilling prophecy: new Unfortunately, no millivolt switch has made it out of the laboratory
chips followed the law because the industry made sure that they did. either. That leaves the architectural approach: stick with silicon, but
And it all worked beautifully, says Flamm — right up until it didn’t. configure it in entirely new ways. One popular option is to go 3D.
Instead of etching flat circuits onto the surface of a silicon wafer, build
HEAT DEATH skyscrapers: stack many thin layers of silicon with microcircuitry
The first stumbling block was not unexpected. Gargini and others had etched into each. In principle, this should make it possible to pack
warned about it as far back as 1989. But it hit hard nonetheless: things more computational power into the same space. In practice, however,
got too small. this currently works only with memory chips, which do not have a
“It used to be that whenever we would scale to smaller feature size, heat problem: they use circuits that consume power only when a
good things happened automatically,” says Bill Bottoms, president of memory cell is accessed, which is not that often. One example is the
Third Millennium Test Solutions, an equipment manufacturer in Santa Hybrid Memory Cube design, a stack of as many as eight memory
Clara. “The chips would go faster and consume less power.” layers that is being pursued by an industry consortium originally

1 4 6 | N AT U R E | VO L 5 3 0 | 1 1 F E B R UA RY 2 0 1 6
© 2016 Macmillan Publishers Limited. All rights reserved
FEATURE NEWS

launched by Samsung and memory-maker Micron Technology in issue,” says Bottoms. Some companies, notably Intel, are still trying to
Boise, Idaho. shrink components before they hit the wall imposed by quantum effects,
Microprocessors are more challenging: stacking layer after layer of hot he says. But “the more we shrink, the more it costs”.
things simply makes them hotter. But one way to get around that prob­ Every time the scale is halved, manufacturers need a whole new
lem is to do away with separate memory and microprocessing chips, generation of ever more precise photolithography machines. Building a
as well as the prodigious amount of heat — at least 50% of the total — new fab line today requires an investment typically measured in many
that is now generated in shuttling data back and forth between the two. billions of dollars — something only a handful of companies can afford.
Instead, integrate them in the same nanoscale high-rise. And the fragmentation of the market triggered by mobile devices is mak­
This is tricky, not least because current-generation microprocessors ing it harder to recoup that money. “As soon as the cost per transistor at
and memory chips are so different that they cannot be made on the same the next node exceeds the existing cost,” says Bottoms, “the scaling stops.”
fab line; stacking them requires a complete redesign of the chip’s structure. Many observers think that the industry is perilously close to that
But several research groups are hoping to pull it off. Electrical engineer point already. “My bet is that we run out of money before we run out
Subhasish Mitra and his colleagues at Stanford University in California of physics,” says Reed.
have developed a hybrid architecture that stacks memory units together Certainly it is true that rising costs over the past decade have forced a
with transistors made from carbon nanotubes, which also carry current massive consolidation in the chip-making industry. Most of the world’s
from layer to layer3. The group thinks that its architecture could reduce production lines now belong to a compara­
energy use to less than one-thousandth that of standard chips. tive handful of multi­nationals such as Intel,

GOING MOBILE “MY BET IS Samsung and the Taiwan Semi­conductor

Manufacturing Company in Hsinchu. These

THAT WE
The second stumbling block for Moore’s law was more of a surprise, but manufacturing giants have tight relationships
unfolded at roughly the same time as the first: computing went mobile. with the companies that supply them with
Twenty-five years ago, computing was defined by the needs of materials and fabrication equipment; they are
desktop and laptop machines; supercomputers and data centres
used essentially the same microprocessors, just packed together in RUN OUT already coordinating, and no longer find the
road-map process all that useful. “The chip

OF MONEY
much greater numbers. Not any more. Today, computing is increas­ manufacturer’s buy-in is definitely less than
ingly defined by what high-end smartphones and tablets do — not before,” says Chen.
to mention by smart watches and other wearables, as well as by the Take the SRC, which functions as the US
exploding number of smart devices in everything from bridges to the
human body. And these mobile devices have priorities very different BEFORE industry’s research agency: it was a long-
time supporter of the road map, says SRC

WE RUN
from those of their more sedentary cousins. vice-president Steven Hillenius. “But about
Keeping abreast of Moore’s law is fairly far down on the list — if three years ago, the SRC contributions went
only because mobile applications and data have largely migrated to the away because the member companies didn’t
worldwide network of server farms known as the cloud. Those server
farms now dominate the market for powerful, cutting-edge micropro­ OUT OF see the value in it.” The SRC, along with the
SIA, wants to push a more long-term, basic

PHYSICS.”
cessors that do follow Moore’s law. “What Google and Amazon decide research agenda and secure federal funding
to buy has a huge influence on what Intel decides to do,” says Reed. for it — possibly through the White House’s
Much more crucial for mobiles is the ability to survive for long periods National Strategic Computing Initiative,
on battery power while interacting with their surroundings and users. launched in July last year.
The chips in a typical smartphone must send and receive signals for voice That agenda, laid out in a report5 last September, sketches out the
calls, Wi-Fi, Bluetooth and the Global Positioning System, while also sens­ research challenges ahead. Energy efficiency is an urgent priority —
ing touch, proximity, acceleration, magnetic fields — even fingerprints. especially for the embedded smart sensors that comprise the ‘Internet
On top of that, the device must host special-purpose circuits for power of things’, which will need new technology to survive without batteries,
management, to keep all those functions from draining the battery. using energy scavenged from ambient heat and vibration. Connectivity
The problem for chipmakers is that this specialization is undermining is equally key: billions of free-roaming devices trying to communicate
the self-reinforcing economic cycle that once kept Moore’s law hum­ with one another and the cloud will need huge amounts of bandwidth,
ming. “The old market was that you would make a few different things, which they can get if researchers can tap the once-unreachable terahertz
but sell a whole lot of them,” says Reed. “The new market is that you have band lying deep in the infrared spectrum. And security is crucial — the
to make a lot of things, but sell a few hundred thousand apiece — so it report calls for research into new ways to build in safeguards against
had better be really cheap to design and fab them.” cyberattack and data theft.
Both are ongoing challenges. Getting separately manufactured These priorities and others will give researchers plenty to work on
technologies to work together harmoniously in a single device is often a in coming years. At least some industry insiders, including Shekhar
nightmare, says Bottoms, who heads the new road map’s committee on the Borkar, head of Intel’s advanced microprocessor research, are optimists.
subject. “Different components, different materials, electronics, photonics Yes, he says, Moore’s law is coming to an end in a literal sense, because
and so on, all in the same package — these are issues that will have to be the exponential growth in transistor count cannot continue. But from
solved by new architectures, new simulations, new switches and more.” the consumer perspective, “Moore’s law simply states that user value
For many of the special-purpose circuits, design is still something of doubles every two years”. And in that form, the law will continue as
a cottage industry — which means slow and costly. At the University of long as the industry can keep stuffing its devices with new functionality.
California, Berkeley, electrical engineer Alberto Sangiovanni-Vincentelli The ideas are out there, says Borkar. “Our job is to engineer them.” ■
and his colleagues are trying to change that: instead of starting from
scratch each time, they think that people should create new devices by M. Mitchell Waldrop is a features editor for Nature.
combining large chunks of existing circuitry that have known function­
1. Moore, G. E. Electronics 38, 114–117 (1965).
ality4. “It’s like using Lego blocks,” says Sangiovanni-Vincentelli. It’s a 2. Moore, G. E. IEDM Tech. Digest 11–13 (1975).
challenge to make sure that the blocks work together, but “if you were to 3. Sabry Aly, M. M. et al. Computer 48(12), 24–33 (2015).
use older methods of design, costs would be prohibitive”. 4. Nikolic, B. 41th Eur. Solid-State Circuits Conf. (2015); available at [Link]
com/wwljk7
Costs, not surprisingly, are very much on the chipmakers’ minds these 5. Rebooting the IT Revolution: A Call to Action (SIA/SRC, 2015); available at http://
days. “The end of Moore’s law is not a technical issue, it is an economic [Link]/urvkhw

1 1 F E B R UA RY 2 0 1 6 | VO L 5 3 0 | N AT U R E | 1 4 7
© 2016 Macmillan Publishers Limited. All rights reserved