May 2005

Published by the IEEE Computer Society

C

omputer performance has

been driven largely by

decreasing the size of chips

while increasing the number

of transistors they contain. In

accordance with Moore’s law, this has

caused chip speeds to rise and prices

to drop. This ongoing trend has dri-

ven much of the computing industry

for years.

However, transistors can’t shrink

forever. Even now, as transistor com-

ponents grow thinner, chip manufac-

turers have struggled to cap power

usage and heat generation, two critical

problems. Even performance-enhanc-

ing approaches like running multiple

instructions per thread have bottomed

out.

For these reasons, processor perfor-

mance increases have begun slowing.

Chip performance increased 60 per-

cent per year in the 1990s but slowed

to 40 percent per year from 2000 to

2004, when performance increased by

only 20 percent, according to Linley

Group president Linley Gwennap.

“We could build a slightly faster

chip, but it would cost twice the die

area while gaining only a 20 percent

speed increase,” noted Marc Tremblay,

chief architect for Sun Microsystems’

Scalable Systems Group.

In response, manufacturers are

building chips with multiple cooler-

running, more energy-efficient pro-

cessing cores instead of one in-

creasingly powerful core. The multi-

core chips don’t necessarily run as fast

as the highest performing single-core

models, but they improve overall per-

formance by handling more work in

parallel, as Figure 1 shows.

“Multicore chips are the biggest

change in the PC programming model

since Intel introduced the 32-bit 386

architecture,” stated Gwennap.

“Multicores are a way to extend

Moore’s law so that the user gets more

performance out of a piece of silicon,”

said John Williams, Advanced Micro

Devices’ technical director for server

microprocessor planning.

Chip makers AMD, IBM, Intel, and

Sun are now introducing multicore

chips for servers, desktops, and laptops.

Current transistor technology limits

the ability to continue making single

processor cores more powerful.

For example, as a transistor gets

smaller, the gate, which switches the

electricity on and off, gets thinner and

less able to block the flow of electrons.

Thus, small transistors tend to use elec-

tricity all the time, even when they

aren’t switching. This wastes power.

Also, increasing clock speeds causes

transistors to switch faster and thus

generate more heat and consume more

power. Gwennap said thermal-design

advances have mitigated some prob-

lems. However, he noted, this approach

can’t keep pace with processors’

increasing power and heat buildup.

These and other challenges have hurt

manufacturers’ plans for new, faster

single-core processors. For example,

Intel canceled two next-generation

Pentium 4 processors last year, noted

Jeff Austin, the company’s desktop

product manager. Intel also postponed

and then cancelled a 4-GHz, current-

generation Pentium. And IBM could

build so few of its G5 chips that Apple

Computer had to delay last year’s intro-

duction of its new iMac G5 desktop,

which uses the processor.

A dual-core chip running multiple

applications is about 1.5 times faster

than a chip with just one comparable

core, according to University of Texas

assistant professor Steven Keckler. He

said each core in a typical multicore

chip includes everything a micro-

processor has except level-2 cache

memory and the memory hierarchy,

which is located elsewhere on the sili-

con for all the cores to use.

“The compiler handles the schedul-

ing of instructions for a program,” said

Bill Roth, vice-president of product

marketing for software vendor BEA

Systems.

The operating system controls the

overall assignment of tasks in a multi-

core processor. Based on this, either the

OS or a multithreaded application

parcels out work to the multiple cores.

Generally, when a multicore proces-

sor has completed a task, one core takes

the completed data from the other cores

and assembles the final result.

Chip Makers

Turn to Multicore

Processors

Computer

When a single-core chip runs multi-

ple programs, it assigns a time slice to

work on one program and then assigns

different time slices for others, noted

assistant professor Keckler. This can

cause conflicts, errors, or slowdowns

when the processor must perform mul-

tiple tasks simultaneously.

“If you have multiple tasks that all

have to run at the same time, you will

see a boost with multicore processors,”

said Keckler. For example, the chips

could use a separate core for each task.

Because the chips’ cores are on the

same die, they can share architectural

components, such as memory elements

and memory management. They thus

have fewer components and lower

costs than systems running multiple

chips. Also, the signaling between

cores can be faster and use less elec-

tricity than on multichip systems.

Several companies are making or

planning to make multicore chips.

AMD will release Opteron enter-

prise-server multicore first, then the

Athlon 64 and Sempron desktop chips,

and finally Turion mobile chips. “We

will ship them all this year,” said

Williams.

AMD says it didn’t have to change its

chip architecture to accommodate mul-

Totake advantage of multicore chips,

vendors must redesign applications so

that the processor can run them as mul-

tiple threads. “It is more challenging to

create software that is multithreaded,”

noted AMD’s Williams.

Programmers must find good places

to break up the applications, divide the

work into roughly equal pieces that

can run at the same time, and deter-

mine the best times for the threads to

communicate with one another.

Vendors also must redesign applica-

tions so that they can recognize each

core’s speed and memory-access capa-

bilities as well as how fast cores can

communicate with one another.

Intel provides a Threading Toolkit

to help game and other software devel-

opers design multithreaded applica-

tions to be used on its new chips.

Each of the two cores in AMD’s

Opteron and Intel’s Itanium chips for

servers and workstations will have its

own cache. IBM, on the other hand,

doesn’t use separate caches in its mul-

ticore server chips.

Separate caches eliminate the extra

work needed to design chips so that

multiple cores can work with a single,

centralized cache. In some chip

designs, though, single caches can

function more rapidly than multiple

caches.

ticore capabilities because it anticipated

that the technology would become

viable and developed its architecture

several years ago with that in mind.

Intel is working on 16 multicore-

chip projects.

Two Pentium chips will use the dual-

core technology code-named Smithfield.

Intel’s high-end PC chip, the Pentium

Processor Extreme Edition 840, has

begun shipping and runs at 3.2 GHz but

outperforms today’s high-end, single-

core, 3.8-GHz Pentium 4. The Pentium

D, slated for release this year, will be a

mainstream desktop chip.

Intel plans to release two dual-core

Xeon server and workstation proces-

sors early next year: Xeon MP chips

for servers running at least four proces-

sors and Xeon DP chips for servers and

workstations.

A dual-core Itanium server chip

scheduled for release later this year will

contain 1.7 billion transistors. It will

be Intel’s first processor with more

than 1 billion transistors. Intel has not

released information about the chip’s

clock speed.

Intel has developed energy-saving

dynamic power-coordination technol-

ogy that, when workloads permit, lets

the OS tell one processing core to sleep

or slow down while the other works.

Intel plans to integrate DPC, which

would extend battery life, into Yonah,

the company’s first dual-core laptop

chip. Yonah is slated for release later

this year.

In 2002, Intel introduced its hyper-

threading technology, now supported

by Windows XP and most Linux

releases. Hyperthreading lets multi-

threaded software’s threads execute in

parallel on a single core, thereby

improving performance.

Hyperthreading accomplishes this by

enabling more efficient use of all exe-

cution units—including arithmetic

logic units and floating-point units—in

a core. The technology also informs the

OS that it supports multiple threads

and coordinates their execution.

2000

Performance

(based of benchmarks)

2004

2008

Single core

Multicore

May 2005

have eight cores, each handling four

threads. It will also feature compilers

that will generate parallel threads auto-

matically from an application. The OS

could then map these threads to the

hardware automatically, Tremblay

explained.

I

n the future, manufacturers will

make their multicore chips faster by

increasing the speed of each core, as

Sun is already doing. During the next

few years, said AMD’s Williams, the

decrease in chips’ feature sizes from

today’s 90 nanometers to 65 nanome-

ters will leave room for more cores.

Multicore processors will find a nat-

ural home in servers, said Keckler, but

won’t be very useful on the desktop

until vendors develop considerably

more multithreaded software.

Until this occurs, Williams said, sin-

gle-core chips will continue to com-

pete. Also, he added, single-core chips

are inexpensive to manufacture, so

they will continue to be popular for

lower-priced PCs for a while.

According to the Linley Group’s

Gwennap, the widespread transition

IBM released the industry’s first

dual-core server chip, the Power 4, in

2001. Last year, it introduced the dual-

core Power 5, which runs four times

faster than its predecessor.

IBM, Sony, and Toshiba have com-

pleted design of the Cell processor

optimized for compute-intensive work-

loads, broadband data transmission,

and multimedia processing. The com-

panies plan to begin production dur-

ing the second half of this year, said

Ted Maeurer, IBM’s lead Cell software

engineer.

They designed Cell for use in con-

sumer-entertainment devices such as the

Sony PlayStation III game console. The

companies plan to implement the chip

this year in an IBM-Sony workstation

primarily for handling computer ani-

mation and other demanding graphics

tasks, and next year in a Sony-Toshiba

high-definition TV and a Sony server.

Cell will use one 64-bit Power core

to run the operating system, divide up

tasks, and assign them to eight 128-bit

processing cores optimized for the

floating-point matrix algebra associ-

ated with computer entertainment and

rich media. The processor will have

considerable bus bandwidth between

cores and to memory.

The first version will run at about

4.6 GHz and perform 256 Gflops. It

will use IBM’s silicon-on-insulator

technology, in which pure crystal sili-

con sits on pure silicon-oxide insula-

tion. The purity lets the chips operate

faster, more efficiently, and cooler.

In 1999, Sun released the Micro-

processor Architecture for Java Com-

puting dual-core, multimedia, desktop

chip. MAJC was never widely adopted

as a desktop processor but has been

used as an embedded-systems chip.

Sun has since built UltraSparc IV dual-

core server chips.

Now, Sun’s Tremblay said, the com-

pany is working on the Niagara multi-

core chip for high-end servers. Planned

for release early next year, Niagara will

to multicore chips will occur during the

next two years. However, this could

give the semiconductor industry time

to find new ways to improve single-

core chip technology via, for example,

exotic materials and advanced manu-

facturing techniques.

The per-processor fees that enter-

prise-software vendors charge their

customers could be a challenge to mul-

ticore chips’ success, as the “Are

Multicore Processors One or Many

Chips?” sidebar explains.

Nonetheless, Williams expressed

optimism and said, “Multiple cores are

the new megahertz. Multicore will be

the transition from brute-force perfor-

mance to architectural elegance.” I

Software vendors charge customers in various ways for using their prod-

ucts. One prominent practice is to charge customers for each processor that will

run the software. A customer running an application on 20 machines with sin-

gle-core processors would thus pay a set amount per chip.

A key issue for multicore-chip makers such as Intel is whether software ven-

dors will consider a processor to be a single core or an entire chip.

Intel defines a processor as a unit that plugs into a single socket on the

motherboard, regardless of whether it has one or more cores, and advocates

that software vendors charge accordingly, explained Jeff Austin, the company’s

desktop product manager.

Microsoft and Novell agree and don’t charge extra for using their software

on multicore processors.

BEA Systems and Oracle, on the other hand, charge more to use their soft-

ware on multicore chips for per-processor licensing. “Customers get added

performance benefit by running our software on a chip with two cores, so we

charge a fraction of the single CPU price for additional cores,” said Bill Roth,

the company’s vice president of product marketing.

Multicore-chip makers are concerned that this type of policy will hurt their

products’ sales.