IEEE Spectrum January, 2011 - 40

top 11 technologies

of the

decade

Multicore
CPUs
PROCESSOR
PROLIFERATION
From multicore to
many-core to hardto-describe-in-asingle-word core
back in 1994, programmers figured that whatever
code they wrote would run
at least 50 percent faster on a
1995 machine and 50 percent
faster still on a '96 system.
Coding would continue
as it always had, with
instructions designed to be
executed one after the other.
But Kunle Olukotun, then
a newly minted professor
of electrical engineering
at Stanford, saw that the
party couldn't go on forever.
The microprocessors of the
day couldn't scale up as
efficiently as you'd expect
through the mere addition
of ever more and ever faster
transistors, the two things
that Moore's Law provided.
To solve that problem,
Olukotun and his students
designed the first generalpurpose multicore CPU.
This idea, more than any
other in the past decade,
is what has kept the
semiconductor industry
climbing the Moore's
Law performance curve.
Without multicore chips,
the computing capability
40

of everything from servers
to netbooks would not be
much better than it was
a decade ago. Everyone's
happy-except perhaps
for the programmers, who
must now write code with
threads of instructions
that must be executed
together-in pairs, quartets,
or even larger groupings.
It's not that old, singlecore CPUs weren't already
doing some parallel processing. When Olukotun
began his work, most
microprocessors had a
"superscalar" architecture.
In the superscalar scheme,
the CPU contained many
replicated components,
such as arithmetic units.

NA * IEEE SPECTRUM * JANUARY 2011

Individual instructions
would be parceled out to the
waiting components. Scaling
up such "instructionlevel parallelism" meant
building in more and
more such components
as the years rolled by.
Olukotun argued
that within a few more
generations, it wasn't going
to be worth the effort. You
needed to provide a quadratic increase in resources
for a linear increase in performance, he said, because
of the complexity of the
logic involved in parceling
out and keeping track of
all the instructions. If you
combined that with the
delays inherent in the mess

of interconnects among
all those parts, it seemed
a losing proposition. Doug
Burger and Stephen Keckler,
both computer scientists
at the University of Texas,
Austin, put a finer point
on it later in the decade,
calculating that instead of
the 50 percent improvements
everyone had gotten used
to, the computing industry should start thinking
12.5 percent. And 12.5 percent
isn't much of a reason to
buy a new computer, is it?
Olukotun's answer was
Hydra, a processor whose
parallelism came not from
redundant circuits within
a single complex CPU but
from building four copies of
a simpler CPU core on one
chip. That way, you save on
interconnects and on the
time lost casting instructions
out and reeling answers back
in. In Hydra, you got parallel
processing without all the
delay-inducing complexity.
In 1998 "we wrapped up
the hardware portion of
the project and declared
victory," says Olukotun.
It was a quiet victory. In
the computing environment
of the 1990s, Hydra seemed
a little crazy, Olukotun
says. Superscalar designs
were still delivering
50 percent performance
improvements every year.
"It was by no means clear
at the time that our view
of the world was going
to win," he recalls. And
indeed it would be years
before processor giants
like Intel, Advanced Micro
Devices, and IBM got the
multicore religion Olukotun
preached. And when they
did, it would largely be
for a reason he had hardly
considered: power.
SPECTRUM.IEEE.ORG

FRANK CHIMERO

no. 5
http://SPECTRUM.IEEE.ORG
Table of Contents for the Digital Edition of IEEE Spectrum January, 2011
