91av

IT MAY not look like it, but that computer on your desk is a racing car. It
responds instantly to every command, has an engine that’s built for pure speed,
and prefers to run flat out. It also wastes power by the tank load. In fact,
it’s a reckless gas guzzler, and its extravagance threatens to slam the brakes
on the IT revolution.

For decades, computer engineers have been fixated on speed. Transistors have
grown steadily smaller and quicker, and microchip makers have crammed more and
more of them into their microprocessors to make them run faster. That’s why
today’s machines can run rings around those of yesteryear. But they’ve
inadvertently created an energy crunch. Transistors have an enormous appetite
for electricity, and computer engineers are struggling to keep them
satisfied.

“It’s been creeping up on us,” says David Albonesi, a computer scientist at
the University of Rochester in New York. “Power consumption has gotten to the
point where it could be the limiting factor in computer performance.”

The problem spans the entire computing spectrum, from portable devices to the
Internet. Power-hungry microprocessors will drain batteries in a few hours flat,
keeping the owners of laptops, mobile phones and PDAs in constant search of wall
sockets. If microprocessors keep demanding so much juice, there’s no way
computers can become more mobile and ubiquitous, says Trevor Mudge, a computer
scientist at the University of Michigan in Ann Arbor. “If you think like I do,
that the future of computing will be in a phone-like device, then power
consumption becomes an obvious constraint.”

Meanwhile, the tens of thousands of microchips used by large Internet
companies suck up electricity nearly as quickly as power plants can produce it.
A single “server farm” can slurp a staggering 150 megawatts, about a quarter of
the entire output of an average power plant.

There’s another problem, too. Almost all the power that flows into a
microprocessor flows back out in the form of heat, and chips are getting too hot
to handle. Today’s high-performance microprocessors give off roughly three times
more heat per square centimetre than a hotplate. If they get much hotter they’ll
cook themselves, says Krste Asanovic, a computer scientist at the Massachusetts
Institute of Technology.

It’s a serious problem, and computer engineers have had to alter their
priorities to tackle it. “Power hasn’t been a concern in the design of a
processor,” says Asanovic. “But what people are now finding is that power is the
primary constraint in how fast you can run your processor.”

So where can engineers make savings? Ironically, the best approach to making
chips more efficient is to instil in them the habits of lazy people. Thus the
microprocessors of the near future are likely to put things off to the last
nanosecond, work only as hard as they have to, juggle several tasks at once and
doze off whenever they can get away with it. The first of the slothful silicon
chips have just hit the market and use only a fraction of the power consumed by
standard microprocessors. And researchers are even working on chips that need no
electrical power at all.

A microprocessor consumes power primarily in the act of flipping microscopic
switches known as transistors. The processor reduces all of its tasks—from
moving an image across the screen to dispatching an e-mail—to binary
mathematics, strings of ones and zeros. To make a 0 it sets the voltage across
one of its transistors to 0; to make a 1 it sets the voltage to a higher value,
typically around 2.5 volts. Each flick of each switch consumes power by an
amount roughly proportional to the square of the voltage difference.

To avoid confusion, the microprocessor must switch its transistors in the
proper order. To keep everything coordinated, the chip employs an internal
“clock”, a pacemaker that sends electrical pulses sweeping through its circuitry
several hundred million times a second. Transistors can switch only with the
arrival of a clock pulse, so the calculation proceeds step by step. The clock
frequency determines how fast a processor can run through its tasks. An 866
megahertz processor, for example, can switch each transistor 866 million times a
second. The faster a transistor switches, the more power it consumes.

So when it comes to saving power, computer engineers have a limited number of
choices. They must either reduce the voltage, slow down the clock, or switch
fewer transistors with each clock pulse. They’re doing all three.

The obvious approach is to reduce the number of transistors that are active
at any one time. To do this, engineers try to make their microprocessors think a
bit like your parents, who no doubt nagged you to turn off lights and the
television when you weren’t using them.

The microprocessor may be a single chip, but it contains several different
units that can act more or less independently. There are units for crunching
integers, units for storing data, units for tracking decimal points and so on.
At any moment only a few of them have any real work to do, says Anantha
Chandrakasan, an electrical engineer at the Massachusetts Institute of
Technology. For example, in calculations using integers only, the decimal point
unit isn’t needed.

But even an idle unit soaks up power, because its transistors continue to
switch as they await their next instructions. To prevent such waste, designers
simply stop sending clock pulses to idle units and put them to sleep, a
technique called clock gating.

Instructing units to shut down is fairly easy, but it doesn’t generate big
power savings. What computer engineers would really like to do is lower the
voltage across the transistors. The problem is, they can’t turn the voltage down
just as low as they please because this reduces the speed at which the
transistors can flip, leaving the microprocessor running in slow motion.

Fortunately, there’s a way of getting around this. Not every task requires
the chip to run flat out. Sometimes it can coast, and the trick is to let it do
so whenever possible.

Suppose, for example, you’re using your computer to listen to a CD. The
microprocessor needs to read the disc at a minimum rate of 40,000 bits per
second. But being an eager go-getter, it fetches each bit as soon as it
can—within a few billionths of a second—and then spends thousands of
clock cycles waiting around for its next instruction.

There’s no real need for such haste. “Instead of reacting instantly, it
should react just under your level of perception,” says Mudge. That means the
voltage can be altered on the fly to keep the computer running only as fast as
necessary, a technique called dynamic frequency scaling. If you’re playing the
latest video game, the voltage will climb as high as it can. But if you’re
typing the opening lines of your first novel, it’ll plummet while you sit there
stroking your chin.

Another way to slow the clock without losing processing capacity is to make
the processor execute more than one instruction at each clock pulse. This
strategy is known as parallelism, and it allows the processor to do the same
amount of work at a lower clock frequency.

For example, one instruction might say “add two numbers”, and the second
“store the sum in memory”. These obviously can’t be executed at the same time.
But if the second instruction is “retrieve an unrelated number from memory”, the
instructions don’t interfere and can be run simultaneously.

Many of today’s microprocessors already employ some degree of parallelism.
Intel’s Pentium III, for example, can execute up to three instructions per clock
pulse. But as usual the purpose is speed, not efficiency. The chip must
continually figure out which incoming instructions can be run in parallel, and
thinking this through burns up as much power as it saves. “The hardware cost of
finding parallelism probably swamps any savings from running it,” Asanovic
says.

To make parallelism pay, researchers are trying a couple of tacks. Some are
tagging each instruction so that the processor can tell immediately which can be
run together. Others are tackling the problem on the next level up, by rewriting
the programs that translate computer languages such as C into machine code.
Called compilers, these programs are traditionally built for speed. Now
researchers are trying to develop compilers that maximise parallelism.

The first microprocessors incorporating power-saving technology have already
hit the market. Last month Intel unveiled the latest low-power version of the
Pentium III for laptop computers. The chip has a low-voltage mode that kicks in
whenever it’s running on battery power. It’s also vigilant about shutting down
idle units, so much so that the chip basically falls asleep between keystrokes.
Also on the market is the Crusoe microprocessor, made by Transmeta of Santa
Clara, California. Crusoe can ratchet its voltage between 1.0 and 1.6 volts and
its frequency between 300 and 600 megahertz. Both Crusoe and the low-power
Pentium III run on roughly half a watt, a pittance compared with the 20 to 40
watts a typical desktop chip gets through.

Such power savings are enough to keep a laptop running all day. But some
researchers want to go even further. The biggest power savings may come from
specialised microprocessors that do fewer, more specific tasks. That may sound
like a bad deal to the technophile who would like to browse Web pages on a
cellphone, but it can pay off in spades in certain situations. A processor built
to do just one thing consumes about a thousandth of the power a general-purpose
processor needs to do the same deed, says Jan Rabaey, an electrical engineer at
the University of California at Berkeley. “It’s a trade-off of flexibility for
efficiency,” he says.

Rabaey is developing a system of smart sensors to control climate within
large buildings. Each of the sensors will have equipment to monitor conditions
such as temperature and light levels, a tiny antenna to send signals to its
neighbours and a microprocessor with just enough capability to process the data.
The system should be smart enough to know, for example, when to turn the air
conditioning on and when to turn the lights off.

Most remarkably, it should require no electrical power at all. Each processor
uses so little energy that it can satisfy its needs by soaking up dribs and
drabs from its environment, a strategy called energy scavenging. “We basically
take the energy out of light, vibration and pressure,” Rabaey says.

Within five years or so, researchers hope to reduce by a factor of 100 the
power consumed by a typical microprocessor. But what they are really aiming for
is a revolution in the speed-first culture of computing. “We don’t buy cars by
horsepower any more,” Mudge says. “Maybe we’ll get past the frequency number for
computers someday.” If so, the next time you buy a new computer you can happily
trade in your gas guzzler for a compact, because you’ll never notice the
difference.