IEEE Spectrum April, 2015 - 35

DATA sourcEs: AMD, KooMEy ET AL. (2011)

Moore's Law

There are many ways to gauge a
computer's efficiency, but one of
the most easily calculated metrics is
peak-output efficiency, which measures the efficiency of a processor
when it's running at its fastest.
Peak-output efficiency is typically
quoted as the number of computations
that can be performed per kilowatthour of electricity consumed. And
according to a peer-reviewed paper
published in 2011 in the IEEE Annals
of the History of Computing, it doubled
like clockwork every year and a half or
so for more than five decades.
This trend started well before
the first microprocessor, way back
in the mid-1940s. But it began
to come to an end around 2000.
Growth in both peak-output efficiency and performance started to
slow, weighed down by the physical
limitations of shrinking transistors.
Chipmakers turned to architectural
changes-such as putting multiple
computing cores in a single microprocessor-but they weren't able to
maintain historical growth rates.
These days, we've found, it takes
about 2.7 years for peak-output
efficiency to double. That's a
substantial slowdown. Historically,
a decade of doubling boosted
efficiency by a factor of a hundred;
at current rates, it would take
18 years to see a hundredfold gain.
Fortunately, the news isn't all bad.
Our computing needs have changed.
For years after Moore's landmark
1965 paper, computers were expensive, relatively rare, and regularly
pushed to their computing peak. Now
that they're ubiquitous and cheap,
the emphasis in chip design has
shifted from fast CPUs in stationary
machines to ultralow-power processing in mobile appliances, such as
laptops, cellphones, and tablets.

Today, most computers run at
peak output only a small fraction
of the time (a couple of exceptions
being high-performance supercomputers and Bitcoin miners). Mobile
devices such as smartphones
and notebook computers generally operate at their computational
peak less than 1 percent of the time
based on common industry measurements. Enterprise data servers
spend less than 10 percent of the

We've recently defined a measure
of efficiency that's more in sync
with how chips are used nowadays,
which we call "typical-use efficiency."
Like peak-output efficiency, it's
measured in computations per
kilowatt-hour. This time, however, it's
calculated by dividing the number
of computations performed over
the course of a year by the total
electricity consumed-a weighted
sum of the energy a processor

enerGY efficiencY relative to 1985 (1985 = 1.0)
10 m

100 k

2008

Typical-use
efficiency
(historical
and projected)

2000-2009 trend
in peak-output
efficiency (historical
and projected)

Peak-output
efficiency
(historical and
projected)

2012

year operating at their peak. Even
computers used to provide cloudbased Internet services operate at
full blast less than half the time.
In this new regime, a good powermanagement design is one that
minimizes how much energy a device
consumes when it's idle or off. And
the better indicator of energy efficiency is how much electricity a computer consumes on average-not
when it's operating at full blast.

in the way the semiconductor industry describes itself. In
the 1980s and early 1990s, the technology generations, or
"nodes," that define progress in the industry were named after
dynamic RAM generations: In 1989, for example, we had the
4-megabyte DRAM node; in 1992, the 16-MB node. Each generation meant greater capability within a single chip as more
and more transistors were added without raising the cost.
By the early 1990s, we'd begun to name our nodes after the
shrinking features used to make the transistors. This was only

2016

2020

and its supporting circuitry use in
different modes over that same
period. For example, a laptop might
operate at peak power when its
user is playing a game, but this only
happens a tiny fraction of the time.
Other common activities, such as
word processing or video playback,
might consume a tenth as much
electricity, since only a fraction of the
chip is needed for these functions,
and smart power management can

50 Years

actively shut off circuitry between
keystrokes and video frames.
Encouragingly, typical-use efficiency seems to be going strong,
based on tests performed since
2008 on Advanced Micro Devices'
chip line. Through 2020, by our
calculations for an AMD initiative,
typical-use efficiency will double
every 1.5 years or so, putting it back
to the same rate seen during the
heyday of Moore's Law.
These gains come from aggressive
improvements to circuit design,
component integration, and software,
as well as power-management
schemes that put unused circuits
into low-power states whenever
possible. The integration of
specialized accelerators, such as
graphics processing units and signal
processors that can perform certain
computations more efficiently, has
also helped keep average power
consumption down.
Of course, as with any exponential trend, this one will eventually
end, and circuit designers will have
become victims of their own success. As idle power approaches zero,
it will constitute a smaller and smaller
fraction of the energy consumed by a
computer. In a decade or so, energy
use will once again be dominated by
the power consumed when a computer is active. And that active power
will still be hostage to the physics
behind the slowdown in Moore's Law.
Over the next few decades, we'll
have to rethink the fundamental
design of computers if we want to
keep computing moving forward
at historical rates. In the meantime,
steady improvements in everyday
energy efficiency will give us a bit
more time to find our way.
-JonaThan kooMey
& saMueL naffziger

natural. Most chips didn't need to carry as many transistors as
possible. Integrated circuits were proliferating, finding their
way into cars and appliances and toys, and as they did so, the
size of the transistor-with the implications for performance
and cost savings-became the more meaningful measure.
Eventually even microprocessors stopped scaling up as fast
as manufacturing technology would permit. Manufacturing
now allows us to economically place more than 10 billion transistors on a logic chip. But only a few of today's chips come
SPECTRUM.IEEE.ORG

nORTh aMERICan

aPR 2015

http://SPECTRUM.IEEE.ORG

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