快猫短视频

A DARKNESS is descending on Silicon Valley. The lights that carve microchip
components on Frisbee-sized wafers of silicon are about to be extinguished. The
reason is simple: to make smaller and smaller transistors takes light of ever
diminishing wavelength. And after 35 years of continually packing more and
smaller components onto each chip, the limit has been reached. No known material
can focus light at the wavelengths needed to make tomorrow鈥檚 chips.

The crisis is bringing together experts in fields as diverse as laser fusion,
astronomical interferometry and spy-satellite imaging. By drawing on their
skills the chip makers hope to ensure that the next generation of super-dense
chips can go ahead as planned.

Each layer of a microchip鈥檚 components is fashioned with a stencil-like mask.
Light shining through the mask projects a pattern onto a wafer that has been
coated with a photoresist鈥攁 chemical that hardens in the light. Unexposed
resist can then be washed away, leaving areas that are ready for a layer of
material to be deposited to make anything from part of a transistor to a
wire.

Today鈥檚 chips, such as Intel鈥檚 Pentium processor, are etched using deep
ultraviolet light, which has a wavelength of 193 nanometres. This can create
transistors just 250 nanometres across. At this size, it鈥檚 possible to squeeze
11 million transistors onto a chip. According to Sematech, the American chip
industry鈥檚 research consortium, deep ultraviolet should be carving features as
small as 130 nanometres across by 2003. Beyond this, new techniques will be
needed鈥攁nd quickly. To keep up the momentum of technical advance, Sematech
estimates that chips with 200 million transistors, each just 100 nanometres
across, will have to be ready by 2006.

With no materials available that are transparent at wavelengths below 193
nanometres, lens-based chip-making tools are headed for the scrap heap. The
鈥減ost-optical era鈥� is on the way, though nobody yet knows which technologies
will usher it in. IBM and Canon want to replace light with X-rays. Bell Labs
proposes to use multiple beams of electrons to inscribe chip patterns. Siemens
of Germany has put its money鈥攑lus a contribution from the European
Commission鈥攐n a projection system based on beams of ions. Meanwhile, a
pair of Silicon Valley giants, Intel and Advanced Micro Devices (AMD), have
teamed up with Motorola from Austin, Texas, to work out how to etch chips with
鈥渆xtreme ultraviolet鈥� (EUV)鈥攖he chip makers鈥� name for soft X-rays with a
wavelength of 10 to 15 nanometres.

They are confident that EUV will have the edge for the first post-optical
chips. 鈥淢ost of the smart money says EUV will appear for the 100-nanometre
generation,鈥� says Bill Siegle, chief scientist at AMD in Sunnyvale. Further down
the line, for dimensions going below 100 nanometres, rival technologies may take
over.

But nobody is suggesting that even EUV will be ready tomorrow. Siegle openly
admits that there is a huge amount of engineering work still to be done on EUV.
That鈥檚 why the three companies have paid $250 million to a trio of US
national laboratories with sites across the bay from Silicon
Valley鈥擫awrence Livermore, Sandia and Lawrence Berkeley鈥攖o help
tackle the scientific challenges. 鈥淭his is the first time that these three
laboratories have teamed up to apply resources to a basic problem,鈥� says Donald
Sweeney, the EUV programme manager at Livermore. The project is employing 150
scientists at the three sites, he says.

The chief challenge is find a way to project a sharp EUV image of a mask onto
a silicon wafer. 鈥淏ecause there鈥檚 nothing transparent like a piece of glass at
EUV wavelengths, we have to build a system totally out of mirrors,鈥� says
Sweeney. That means using a 鈥渞eflective mask鈥�, from which an image of the EUV
radiation that bounces off the pattern on the mask is focused onto a wafer by a
series of concave mirrors. The mask is four times the size of the image, which
makes it easier to manufacture.

To generate the EUV radiation, Livermore researchers have used their
experience of inertial confinement fusion. They take a supercold pellet of solid
hydrogen isotopes and blast it from all sides with beams of radiation from
powerful lasers (see 鈥淪uperlaser for baby bombs鈥�, 快猫短视频, 29
August, p 32). The resulting plasma, in which nuclear fusion can be triggered,
emits EUV radiation. 鈥淚n our proposed EUV system we use a big laser that hits a
small supersonic gas jet in a vacuum,鈥� says Sweeney. 鈥淭he result is a plasma
that radiates similarly to a laser fusion target.鈥� The EUV system is smaller
than a fusion machine, adds James Brase, deputy leader of Livermore鈥檚 advanced
microtechnology programme. 鈥淭his one would fit in a small room.鈥�

The four mirrors in the EUV machine also borrow from past research. The
details are secret, but Sweeney will say that they are made with multilayered
coatings that are highly reflective to X-rays. They must be machined to
incredible accuracy鈥攚ithin 0.2 nanometres of the original design. 鈥淭hat鈥檚
roughly the size of a silicon atom,鈥� says Brase. To achieve this, the Livermore
researchers are using an interferometry method they dreamt up to check the
accuracy of the mirrors at the 10-metre Keck telescope in Hawaii.

A beam of light is sent down an optical fibre that is aimed at a spot on a
mirror. This beam interferes with its own reflection coming back off the mirror,
to create a pattern of black and white fringes on a semi-transparent film on the
face of the fibre. The nature of that pattern, which is captured by a camera,
gives a precise measure of any unevenness in the mirror. Any anomalies found
this way can then be corrected.

The masks needed to make chips with features of just 100 nanometres will need
to be examined for defects at a resolution of around 25 nanometres. But how to
do it? 鈥淭he sheer number of pixels that you have to look at makes this a
tremendous imaging problem,鈥� says Brase.

This is where the satellite technology comes to the rescue鈥攖he
Livermore group that used to work on spy satellites has now been transferred to
EUV research. Radar imagers from spy satellites routinely scan the sea surface.
鈥淲e have for many years had programmes looking for ship wakes on the ocean,鈥�
says Brase. 鈥淚t鈥檚 a disturbance on a surface . . . Our processing techniques can
find that disturbance,鈥� says Brase. The defects on a chip may be different to
ships鈥� wakes, but the principle of finding them is the same.

鈥淒efects, at some point are going to be one of the biggest challenges,鈥� says
Brase, 鈥渘o matter which lithography ends up being the winner in this post-optical era.鈥�

Looking for Leonardo

They said he knew the secret of creativity, but how to find him鈥�

鈥淲HAT people don鈥檛 realise is the extraordinary creativity required to design
chips,鈥� says Tom McWilliams. We鈥檝e been talking about a topic that has the
potential to send an insomniac to sleep鈥攁utomated chip design. But
suddenly I鈥檓 all ears. How on earth can chip design be creative?

鈥淎 custom chip is like a piece of fine artwork. They are masterpieces,鈥�
continues McWilliams, an expert in automated chip design at Sun Microsystems
Laboratories near Palo Alto. We鈥檙e not talking about off-the-shelf Pentium or
Alpha processors here鈥攁utomation plays a big role in their design. We鈥檙e
talking about chips for specific tasks鈥攖horoughbreds designed for the
highest performance. 鈥淗uman designers are still the best.鈥�

My mind begins to race. What does this tell us about the way the brain works?
Or the nature of creativity? And what is it about chip design that is so tough
for a computer?

鈥淪o who is the most creative chip designer in Silicon Valley?鈥� I ask.
Silence.

鈥淲hat do you mean?鈥�

鈥淚 mean that if I wanted to interview the Leonardo da Vinci of the
chip-design world, who should I call?鈥� McWilliams thinks for a while. 鈥淭here鈥檚 a
designer at DEC. What鈥檚 his name? Dan Doverpull or Doberpool, I think.鈥�

鈥淗ow do you spell it?鈥�

鈥淒-O-B . . . err . . . I鈥檓 not sure.鈥�

Maybe I should track down Dan Whatshisname. His views on the nature of
creativity would be fascinating. Perhaps there鈥檚 a human-suffering
angle鈥攖he creative genius whose great works of art go unrecognised by an
ignorant and unappreciative world. If only I can find him.

I start constructing the story in my head. The week before I鈥檇 been talking
to Burt Sutherland, the director of Sun Microsystems Laboratories. He explained
that there is a crisis in chip design. Chips now run so fast that their speed is
limited by the rate at which data can move between the parts inside them. The
solution is to minimise the distance data has to travel between these different
parts. 鈥淲hy is that difficult to automate?鈥� I ask him.

Sutherland suddenly seems more animated. 鈥淭hink about how a computer would
have to do it. The computer could arrange a set of components on a chip at
random, then link them together with wires and then measure the length of each
wire. If any wire is longer than x then the design isn鈥檛 suitable. It
would then try another random configuration, then another, then another. That is
a long process.鈥�

The penny drops. 鈥淪o it鈥檚 like the travelling salesman problem,鈥� I
venture.

鈥淓虫补肠迟濒测.鈥�

The travelling salesman problem is a famous conundrum in mathematics. It
sounds simple: if a salesman has x places to visit, what route should
he take to minimise the distance he has to travel? The solution is
straightforward: measure all possible routes and see which is the shortest.

The travelling salesman problem is one of a class called nonpolynomial (NP),
which takes its name from the fact that the time a computer needs to solve them
grows as some nonpolynomial function of the size of the problem. Finding a
faster way of solving these problems is one of the holy grails of
mathematics鈥攁nd of chip design. Today鈥檚 chips have millions of
transistors, each one of which can be thought of as a destination. Solving such
a huge problem would take for ever, or close to it.

What fires my imagination is the idea of humans solving this problem more
effectively than a computer because we鈥檙e creative and they just plain are not.
I decide that I must ask DD.

But there is a problem. How do I find a man who鈥檚 name I cannot spell? I put
various spellings into a Web search engine. I try Doverpull. Nothing. Doberpul,
Doberpuhl, Doverpoll, Doverpuhl . . . nothing. Next I call DEC and eventually
get through to their public affairs department but they can鈥檛 help either.

I decide to look elsewhere for my Leonardo. Perhaps I鈥檒l have more luck at
the University of California at Berkeley.

Richard Newton is an impressive figure. In his office at the department of
electrical engineering and computer science, he introduces me to chip design in
an hour and a half. And he is full of sound bites: 鈥淚f the chip wasn鈥檛 cooled,
it would glow like an electric stove . . . Without design automation you鈥檇 need
every person in America working on the layout of a billion transistor chip . . .
If you were to draw the layout with each wire the width of a street, you鈥檇 have
a freeway system that covered the entire surface area of the US . . .鈥� It鈥檚
brilliant stuff. We even touch on creativity鈥攁 fascinating discussion of
the future of computing and how it could begin to mimic human thought. But it is
ultimately unsatisfying. Newton does not know the secret of creativity.

There is one last chance. I call the DEC research labs in Palo Alto. The
switchboard answers. 鈥淐an you put me through to Dan Dover . . .鈥� I splutter to
hide that I don鈥檛 know his name.

鈥淲hat was the last name?鈥� the woman asks, and I repeat it.

鈥淥h Dobberpuhl, I鈥檒l just put you through.鈥�

Bingo! But it turns out that our man left DEC two months ago to start his own
company, called SiByte. I track it down and dial.

鈥淚s that Dan Dobberpuhl?鈥�

鈥渊别蝉.鈥�

It鈥檚 like the voice of God. I explain my story and ask if he鈥檇 be willing to
be interviewed. He pauses, then drops the bombshell:

鈥淎ctually no, I think I鈥檒l pass on the interview.鈥�

No interview, no story. This is a disaster.

It gets worse. Hearing the sound of Dan鈥檚 voice is a sharp dose of reality.
If I asked an artist or musician about their process of creative thought, what
would they say? Probably nothing of real interest or importance. How likely is
it that Dan has some extraordinary insight? He is a human being after all, not a
god.

With or without the interview, the secret of creativity is going to remain
hidden. But I have an end for my story. 鈥淥K then,鈥� I reply with resignation.
鈥淭hanks anyway.鈥�

Copper in the valley

For really high speeds, we need to rethink what chips are made of

PICTURE a skyscraper with foundations reaching down 20 metres and a towering
superstructure rising 3 kilometres into the sky. At ground level and below, its
rooms are jam-packed with machinery. Above them rise storey after storey of
glass shot through with a fine mesh of aluminium. Reduced by a factor of 100
million, this shining edifice is, of course, a silicon chip.

The number of storeys apart, it is a picture that would have been
recognisable even two decades ago. But it won鈥檛 be for much longer. Aluminium
wiring is on the way out and is about to be superseded by copper. Time is almost
up for glass鈥攕ilicon dioxide鈥攁s well. Contenders to take its place
include such outlandish materials as 鈥渟pin-on鈥� plastics and foam-like gels.

These is no escaping these changes. They are dictated by the laws of physics.
As components shrink, to squeeze more of them onto each chip, the old materials
are hitting their limits. 鈥淭he industry is at a critical crossroad,鈥� says Paul
Ho, professor of materials science and engineering at the University of Texas,
Austin. 鈥淭hey鈥檝e got to uproot from technologies they鈥檝e used for 25 years and
move into areas they know little about.鈥�

Eventually the transistors at the base of our skyscraper will have to be
modernised. But for now, attention is focused on the
interconnect鈥攖he above-ground levels of the skyscraper housing the five or
six layers of wiring that link the transistors. The problem is that this complex
structure is succumbing to a creeping inertia.

鈥淲e work in a world of voltages,鈥� says Hans Stork of Hewlett-Packard鈥檚
research labs in Palo Alto. 鈥淚f you want to establish a voltage you need to
charge up a capacitance and to do this as fast as possible you want the smallest
possible capacitance and the largest possible current.鈥� But as wires get
thinner, they can carry less current, and as they get closer together, the
capacitance between them grows. The simple fact is that charging up the
aluminium web embedded in glass is starting to take too long. The solution is
twin-pronged: find a substitute for aluminium that can handle more current, and
a replacement for glass that doesn鈥檛 store so much charge, that is, has a lower
dielectric constant.

Earlier than most people expected, the first substitute has arrived. IBM
began research on copper interconnects for its mainframe computers as long ago
as 1980. When the PC took off, the work was sidelined, but it was never
abandoned. Boosted by this head start, IBM launched the first copper chips in
September.

Copper鈥檚 conductivity is about 40 per cent higher than that of aluminium,
says John Heidenreich, manager of interconnect technology at IBM in East
Fishkill, New York. So it can cope with higher currents than aluminium even when
made into thinner wires.

But improved conductivity doesn鈥檛 solve all the problems. As wires get
thinner and the density of the current increases you come up against the
phenomenon known as electromigration. 鈥淲ith very high current densities,
electrons impart some of their momentum to the atoms and push them away,鈥�
Heidenreich says. 鈥淭his can create voids in the metal, so it wears out
谤补辫颈诲濒测.鈥�

Fortunately, copper is not as susceptible to this degradation as aluminium.
But it does bring other problems. It diffuses into silicon and contaminates it,
making it more conductive, according to Paolo Gargini, director of technology
strategy at Intel in Santa Clara. So something is needed to isolate the copper
from the silicon. Gargini says the most likely barrier materials are thin films
of tantalum or titanium compounds, but IBM is keeping the nature of its barrier
material a secret.

To build a conventional aluminium interconnect, chip makers first lay down a
blanket of the metal across the surface of the chip. They then etch parts of the
metal away to leave the desired pattern of wires. Finally they deposit silicon
dioxide around the wires and polish it down to the level of the metal. For
copper, IBM has found it gets better results by turning this process on its
head: growing a layer of silicon dioxide first, etching out channels for the
wires, then depositing copper and polishing it back to the level of the glass.
IBM has also pioneered electroplating on the chip assembly line which,
Heidenreich says, creates better copper wires than other techniques.

The move to copper, says Ho, will buy time to find a replacement for silicon
dioxide. Which is just as well, as a substitute is proving hard to find. The
crucial property here is the material鈥檚 dielectric constant (K) a
measure of its ability to store charge. The lower it is, the more quickly a
given current can change the voltages in the wires of the
interconnect鈥攚hich is what everyone is after. The ideal value is 1, but
only air has a K that鈥檚 this low. For silicon dioxide it鈥檚 about 4.

If you are prepared to settle for a K above 3, you can simply add a
little fluorine to the silicon dioxide. Otherwise, you have to switch to an
organic polymer from, say, the parylene or polynaphthalene families. Below a
K of 3, things start to get wilder, with thin films of fluorinated
amorphous carbon and inorganic polymers that form cages, such as
hydrogensilsesquioxanes. Lower still are the truly exotic xerogels and
aerogels鈥攑orous forms of silicon dioxide with K values as low as
1.3, thanks chiefly to the air they contain.

Some of the alternatives have already been tested. Texas Instruments, for
example, has demonstrated a chip with copper separated by a xerogel. These
alternatives are applied differently and, like copper, are likely to create
production problems. Some are dumped on the silicon wafer, which is then spun to
create an even film, while others are deposited as vapour. They may also be heat
sensitive, so parts of the production process may need to be kept cooler than is
ideal. And some dielectrics are so mechanically weak that engineers describe
them as 鈥渃heese reinforced with copper鈥�.

According to Sematech, low-K devices are still five years away. But
its advice to the chip makers is to begin preparing for the new technologies.
鈥淚t鈥檚 not a trivial change,鈥� Monnig says. 鈥淓ven if production demands don鈥檛 need
us to start now, we need to start the learning process now.鈥�

TROUBLE is in store for the ever-shrinking transistor. The problem is with
the gate dielectric, a layer of silicon dioxide that sits on the silicon wafer and is topped with
an electrode (see Diagram). This layer, just a few atoms
thick, serves as the transistor鈥檚 switch. When a positive voltage is applied to
the gate, it attracts electrons to the boundary between the wafer and the
dielectric. Current is then able to flow between the source and drain.

As transistors get smaller, gate dielectrics will dwindle to just 2
nanometres thick to achieve the required capacitance. With a layer this thin,
electrons start to start to leak through by the quantum process known as
tunnelling. The tiny 鈥減arasitic current鈥� this produces is enough to spoil the
operation of the transistor.

Chip designers, then, need a thicker dielectric to stop the tunnelling, but
want to keep the capacitance across the gate at least as high as it is now, so
that it still acts as a switch. This means finding a material with a higher
dielectric constant (K) than silicon dioxide, which has a K of
4. As a first response, the industry is likely to add nitrogen to the existing
dielectric, to make silicon oxynitride. This will roughly double the K
value. Other promising candidates include tantalum pentoxide, which has a
K close to 20 and ferroelectric materials such as bismuth strontium
titanate, which go much higher.