IEEE Spectrum April, 2015 - 36

Moore's Law

50 Years

anywhere close to that total, in large part because our chip
designs generally haven't been able to keep up.
Moore's Law 1.0 is still alive today in the highest-end graphics processing units, field-programmable gate arrays, and
perhaps a handful of the microprocessors aimed at supercomputers. But for everything else, Moore's Law 2.0 dominates.
And now it's in the process of changing again.

This change is happening because the benefits of min-

iaturization are progressively falling away. It began in the
early 2000s, when an unpleasant reality started to emerge.
At that time, transistor sizes began to creep down below
100 nanometers, and Dennard's simple scaling rule hit its
limit. Transistors became so small that it was quite easy for
electrons to sneak through them even when the devices were
supposed to be off, leaking energy and lowering device reliability. Although new materials and manufacturing techniques helped combat this problem, engineers had to stop
the practice of dramatically lowering the voltage supplied to
each transistor in order to maintain a strong electrical clamp.
Because of the breakdown of Dennard scaling, miniaturization is now full of trade-offs. Making a transistor smaller
no longer makes it both faster and more efficient. In fact,
it's very difficult to shrink today's transistors and maintain
even the same speed and power consumption of the previous generation.
As a result, for the last decade or so, Moore's Law has been
more about cost than performance; we make transistors
smaller in order to make them cheaper. That isn't to say
that today's microprocessors are no better than those of
5 or 10 years ago. There have been design improvements. But
much of the performance gains have come from the integration of multiple cores enabled by cheaper transistors.
The economics has remained compelling because of an
important and unheralded feature of Moore's Law: As transistors have gotten smaller, we've been able to keep the cost
of manufacturing each square centimeter of finished silicon about the same, year after year after year (at least until
recently). Moore has put it at about a billion dollars an acre-
although chipmakers seldom think in terms of acreage.
Keeping the cost of finished silicon constant for decades hasn't
been easy. There was steady work to improve yield, which
started in the 1970s at around 20 percent and now sits at
80 to 90 percent. Silicon wafers-the round platters of silicon
that are eventually cut into chips-also got bigger and bigger. The progressive boost in size lowered the cost of a number of manufacturing steps, such as deposition and etching,
that are performed on a whole wafer at once. And crucially,
equipment productivity has soared. The tools employed in
lithography-the printing technology that's used to pattern
transistors and the interconnections between them-cost 100
36

aPR 2015

nORTh aMERICan

SPECTRUM.IEEE.ORG

times as much today as they did 35 years ago. But these tools
pattern wafers 100 times as fast, making up the cost increase
while delivering far better resolution.
These three factors-improved yields, larger wafers, and
rising equipment productivity-have allowed chipmakers
to make chips denser and denser for decades while keeping
the cost per area nearly the same and reducing the cost per
transistor. But now, this trend may be ending. And it's largely
because lithography has gotten more expensive.
Over the last decade, the difficulties of printing tiny features
have raised the manufacturing cost per unit area of finished
silicon about 10 percent per year. Since the area per transistor shrank by about 25 percent each year over the same
period, the cost of each transistor kept going down. But at
some point, manufacturing costs will rise faster than transistor area will fall, and the next generation of transistors will
be more expensive than the last.
If lithography costs rise fast, Moore's Law as we know it will
come to a quick halt. And there are signs that the end could
come quite soon. Today's advanced chips are made with
immersion lithography, which makes patterns by exposing
water-immersed wafers to 193-nm, deep ultraviolet light. The
planned successor is lithography using shorter-wavelength,
extreme ultraviolet light. That technology was supposed to
come on line as early as 2004. But it's suffered delay after
delay, so chipmakers have had to turn to stopgaps such as
double patterning, which doubles up some steps to fashion
the finest features. Double patterning takes twice as long as
single patterning. Nonetheless, chipmakers are contemplating triple and even quadruple patterning, which will further
drive up costs. A few years from now, we may look back on
2015 as the year the tide turned and the cost of transistors
stopped falling and started to rise.

i've been known for making grand pronouncements at

lithography conferences about the coming end of Moore's Law.
But the truth is, I don't think Moore's Law is over. Instead, I'd
argue it's on the verge of morphing again.
Going forward, innovations in semiconductors will continue, but they won't systematically lower transistor costs.
Instead, progress will be defined by new forms of integration:
gathering together disparate capabilities on a single chip to
lower the system cost. This might sound a lot like the Moore's
Law 1.0 era, but in this case, we're not looking at combining
different pieces of logic into one, bigger chip. Rather, we're
talking about uniting the non-logic functions that have historically stayed separate from our silicon chips.
An early example of this is the modern cellphone camera, which incorporates an image sensor directly onto a
digital signal processor using large vertical lines of copper
wiring called through-silicon vias. But other examples will

http://SPECTRUM.IEEE.ORG

Table of Contents for the Digital Edition of IEEE Spectrum April, 2015