EMBEDDED COMPUTING

How Many System

power operation. InfoPad, one of the
first portable networked multimedia
devices, included specialized hardware
for graphics, video, and wireless oper-

Architectures?
ations, leaving the CPU free to perform
housekeeping functions. A special-pur-
pose unit can perform an operation
using less energy than a general-pur-
pose CPU—just eliminating the
Wayne Wolf, Princeton University
instruction fetch and decode logic
saves significant energy. Further, a spe-
cial-purpose unit can be shut down

P
reviously, I talked about CPU
architectures and their 15 Constant pressures on the cost and
minutes of fame (“Whither
power consumption of embedded
Warhol’s Law,” Computer,
Sept. 2002, pp. 96-97). Quite systems will likely continue to
a few instruction sets vie for attention
in the embedded systems marketplace
spawn a diversity of uniprocessor
today. However, instruction sets pro- and multiprocessor platforms.
vide only one piece of the design puz-
zle. An embedded system also requires
input and output, memory, and, fre- system fails. Real-time operation when it is not needed. This is especially
quently, parallel processing. How best requires more than just having enough important today, when leakage cur-
to structure this overall system is the computing power on average—it rents constitute a large fraction of a
fundamental architectural question in requires enough power for the worst- unit’s total energy consumption.
embedded computing. case scenario.
Developers often call an embedded We must also consider power con- GROWING THE CHIP ECOLOGY
system’s hardware architecture a plat- sumption when choosing a platform. The push toward higher VLSI levels
form. The idea is that you select a When confronted with real-time per- has added new urgency to the question
hardware platform, then write soft- formance problems, developers tend to “How many platforms are necessary?”
ware to fit onto it. If only things were respond reflexively by overdesigning The number of platforms required to
that easy. The platform provides the the platform and providing anywhere support embedded systems translates
computing resources required to run from a little to a lot of extra computing directly into the number of different
the embedded application, and the power. However, that computing chip types that semiconductor manu-
decisions about what hardware should power consumes electrical power. facturers need to supply.
go into the platform often depend on Battery-operated devices are always
what sort of software you want to run. highly sensitive about energy con- Why more is better
sumption, but even systems plugged Having more platforms offers
PLATFORM CHOICES into the wall must watch their con- advantages for semiconductor manu-
Choosing a platform is a critical sumption because power dissipates as facturers, particularly as the industry
decision in embedded system design: heat. Fans to cool the electronics not consolidates. More platform options
Too much hardware and you’re wast- only cost money, but they can also be mean more market niches for chip
ing money and probably power, too lit- noisy and unacceptably annoying in companies, just as a richer ecology
tle hardware and you can’t run the many situations. For example, would spawns more species to occupy its
application. Fitting the application you buy an MP3 player with a fan? many environmental niches.
onto the platform means more than Building a platform with some spe- But it takes a lot of customers to
just making sure you have all the right cialized hardware to support common make a new chip type economically
I/O devices. Embedded systems have operations in the application can save viable. Designing a chip takes great
strict performance goals, such as real- significant amounts of energy. The engineering effort, a cost that must be
time deadlines. If you don’t have InfoPad project at UC Berkeley in the spread over its manufacturing life.
enough computing power of the right 1990s clearly demonstrated the impor- Even the cost of the masks used to
type available to meet the deadline, the tance of specialized hardware for low- manufacture chips has become a major

March 2003 93
 Embedded Computing

The PC has been a huge cash cow for block diagram closely reflects the
the chip and computer industries. Vast design of the early PCs. Modern PCs
CPU Memory I/O numbers of PCs have been sold, all are considerably more complex, with
based on a relatively small number of graphics operations, audio, and net-
Bus
different parts. Customers and small working commonly offloaded from the
companies have been left the task of main CPU into separate specialized
making those PCs useful for individual processing units.
DMA Timers Displays applications. Quite a few people in the
electronics industry would like to see Uniprocessor simplicity
another killer category like the PC to The uniprocessor architecture’s
sustain the industry for another decade biggest advantage is simplicity. Its
Figure 1. A simple bus-based architecture or two. That means finding a common design means that the hardware is
for embedded systems. hardware architecture that can serve fairly straightforward—the bus,
all sorts of markets efficiently enough caches, and other components are sim-
factor in overall cost. Masks for a chip for economies of scale to overwhelm ple to design and manufacture. More
manufactured in a 0.25-micron line- any inefficiencies the platform engen- importantly, the software is simple. To
width process a few years ago cost ders for a particular operation. develop an application for this plat-
about $120,000. Today, the masks in For embedded system designers, form, programmers need only a
the latest 90-nanometer process cost knowing that a small number of plat- straightforward programming model.
$1 million. Given that the typical con- forms can adequately cover the design The programmer doesn’t need to
sumer-grade chip sells in the $10 range, space leads to comfort. Rather than worry about synchronizing data, hand-
you have to sell a whole lot of chips to searching high and low for the right shaking, or any of those other com-
avoid having the design’s fixed costs combination of hardware and perform- plexities that make parallel programs
drive the chip’s unit cost to an unac- ing extensive experiments to measure harder to debug.
ceptably high level. performance and power consumption, Uniprocessing is attractive if it works.
However, today’s technology has cre- knowing that the choice lies in two or In the embedded world, that means the
ated several large chip markets. For three platforms makes choosing one of CPU must be fast enough to perform
example, consumers buy about 30 mil- them straightforward. Because a plat- the real-time operations with enough
lion CD players every year—that’s just form can be chosen quickly, work on the capacity to provide a reasonable oper-
audio CD players, not computer CD system can swiftly proceed to software ating margin. As VLSI technology
drives. While this may seem surprising, design and implementation. Further, the improves and CPUs run faster, more
remember that CD players are standard characteristics of a platform—its speed applications will fall into this category.
equipment on boom boxes and other for various operations, power con- The hardware block diagram of a
audio systems. Increasingly, automo- sumption, and so on—can be more basic PDA is not much more complex
biles also ship with CD players installed. thoroughly characterized. Better char- than that shown in Figure 1—we need
Consumer audio and video components acterization leads to easier and faster only a little extra hardware to read the
alone provide a long list of high-volume software development because the touch screen. A single CPU performs
devices: DVD players and recorders, implications of a software design deci- handwriting recognition and the data-
digital still cameras, digital video cam- sion can be predicted by looking at the base functions associated with the
eras, digital set-top boxes, and more. code, rather than performing a series of PDA applications. A flash-based MP3
PDAs constitute another major segment custom-designed experiments. player is also simple—basically the
of consumer electronics. And there are architecture shown in Figure 1 with a
plenty more major markets both within DEFINING THE NICHES headphone amplifier for an I/O device.
consumer electronics and in the indus- What sorts of platforms might we A single CPU is more than fast enough
trial and military markets. envision to support embedded appli- to decode MP3 audio in real time.
cations? The simplest such platform is
Fewer is simpler what I call the PC architecture because Moving to multiprocessors
The other side of the debate argues it resembles an early personal com- However, consumer appetites con-
that having fewer platforms makes life puter. This architecture, shown in tinue to grow. As people see more, they
simpler for both semiconductor man- Figure 1, has a simple bus-based orga- want more. Not only does this mean a
ufacturers and their customers. The nization. A single CPU is supported by demand for more functionality, such as
winners in the battle to provide the some memory, timers, a direct memory a move from audio to video, it also
standard embedded systems platform access (DMA) controller, some I/O, requires a device that does more things
would open up enormous markets. and perhaps a display driver. This at one time. PDA applications are de-

94 Computer
 NEW!
IEEE
signed along a traditional model in
which one thing at a time happens—the
to make designers’ lives a little easier.
I’ve covered rate-monotonic analy-
INFORMATION
CPU doesn’t have to worry about
spreading itself among multiple simul-
sis before (“Household Hints for
Embedded Systems Designers,”
TECHNOLOGY
taneous deadlines. However, imagine an
audio player that receives its music input
Computer, May 2002, pp. 106-108).
This little piece of theory provides the
LIBRARY (ITe L)
from a network. That device must now foundation for real-time systems and
juggle the audio and networking loads helps us predict how efficiently a CPU IEEE Journals, Magazines &
simultaneously. Because both these tasks can meet real-time deadlines. A key Conference Proceedings on:
run at relatively high rates, the CPU concept in rate-monotonic analysis is
operates under a much greater burden. the critical instant: the combination of ■ Computing
As another example, consider the events that leads to the highest proces- ■ Communications
lowly CD player. Compact disc tech- sor load. The critical instant occurs ■ Signal Processing
nology is an amazing triumph of sophis- when the deadlines line up so that all ■ Circuits and Systems
ticated electronics and signal processing the real-time processes are ready to
over cheap, low-quality components. run. This means that all the higher-pri- An ideal collection for businesses
Even a traditional CD player that plays ority processes must finish before any and universities focusing on
audio CDs without decompression lower-priority processes can run. these technologies!
requires a DSP to perform real-time Further, a high-rate process can have
The IEEE Information
servo algorithms, along with a separate several deadlines in the same period as
Technology Library (ITeL) brings
unit to perform the modified Reed- a single, slower-rate process.
you online access to 39 top-cited
Solomon decoding required to correct All this means that juggling several
errors. All this is in addition to an ana- tasks at once imposes a significant load
IEEE periodicals and more than
log front end that performs the truly on a processor. Rate-monotonic analy- 900 IEEE conferences, presenting
high-rate signal processing. If you want sis also teaches us that we can’t always the very latest technical research.
to play MP3 files, you need to add utilize 100 percent of the processor if Developed by the IEEE
another CPU for the audio decode. You we want to meet all our deadlines. This
Computer, Communications,
now have a fairly sophisticated multi- means that splitting up a complex set
Signal Processing and Circuits
processor running in real time. of tasks across several smaller CPUs
and Systems Societies.
Video applications require even more often makes it easier to meet perfor-
powerful multiprocessors. Philips’ mance goals. We’ve also seen that
Viper chip provides a digital television using several specialized processors can Request a
platform that contains a MIPS CPU improve power consumption.
along with a Trimedia processor. The
free trial today:
Trimedia is a very-long-instruction- +1 800 701 IEEE (4333)
word CPU that can execute five opera- he division into multiple proces-

T
(USA/CANADA)
tions per clock cycle. Each of these sors is application-specific. It
CPUs has its own bus, and a third bus might be possible to come up with +1 732 981 0060
feeds video data from bulk memory to a fairly generic architecture that would (WORLDWIDE)
the two processors. In addition, dozens let a designer turn on and shut down
onlineproducts@ieee.org
of devices and accelerators support var- parallel processing elements as neces- (EMAIL)
ious aspects of video operation. sary. But we’d also need good virtual
I/O to turn this sort of architecture into
CRITICAL INSTANTS a general platform. It is likely that we
Some have argued that only two or will see a diversity of embedded sys-
three platforms will survive to serve the tems platforms for quite some time to
embedded systems market. Others come. ■
claim that the diverse needs of embed- IEEE Information
ded applications require many more Wayne Wolf is a professor of electrical Driving Invention…
platforms. Who is right? The answer engineering at Princeton University and
depends in part on marketing, politics, author of Computers as Components:
in Information Technology
and other unforeseen forces. But a little Principles of Embedded Computing Sys-
www.ieee.org /onlinepubs
bit of science can give us some insight tem Design (Morgan Kaufman, 2000).
into how many platforms we may need Contact him at wolf@princeton.edu.