By Michael Kanellos
Staff Writer, CNET News.com
April 19, 2005 4:00 AM PDT
Princeton professor Stephen Chou believes he has come up with a way to keep the semiconductor industry rolling forward, and it resembles something Henry VIII might have worn on his royal vestments.
Chou, who also founded Nanonex, is one of the prime advocates of "imprint lithography," a process that involves pressing an ornate template into a liquefied
"Ten years ago, people said, 'This is crazy. You will never use technology to make things this small,'" Chou said.
Imprint lithography--along with silicon nanowires, phase change memory, spintronics
Intel co-founder Gordon Moore's early observation about the rate of progress in the electronics industry--specifically, that the number of transistors on a microchip double every one to two years--turns 40 on Tuesday. Under this principle, chipmakers have managed to steadily boost the performance of their products while simultaneously dropping the price, a rare confluence that has allowed digital technology to seep into virtually every segment of the world economy.
Success, however, has created its own problems. Decades of doubling transistors has led to chips containing several million transistors and multibillion-dollar factories to produce them. Shrinking the size of transistors and the copper wires that connect them to fit more onto a chip has led to problems with electric leakage, power consumption and
And there's only so much room left for shrinking today's transistors. Transistors consist of four basic parts: a source (which stores electrons), a drain (where they go to create a "1" signal), a gate and a gate oxide (which control the flow from the source to the drain). After several shrinks, the gate oxide is only about 10 atoms thick in some cases, meaning an end to shrinking or an arduous chemical or architectural makeover.
Around 2023, the source and drain will be so close that electrons will travel freely between the two and corrupt data unless major changes are made in chip manufacturing techniques, design and materials. Self-assembling circuits, a potential benefit of nanotechnology, could keep the industry chugging for 50 more years, some researchers believe. Unfortunately, any answers to these problems are vague at best.
Tom Theis, the director of physical sciences at IBM, said he knows of no fewer than 20 ideas from credible sources for replacing silicon memory chips. "It is quite difficult to figure out a long-term winner," he said. "There is nothing that is an obvious replacement for a silicon transistor."
Chipmaker Intel believes that Moore's Law, with substantial technological breakthroughs, will continue for at least a decade. A gradual conversion from silicon to other materials will then take place from 2015 to the 2020s. Another strong possibility is that silicon survives, but new materials get added to it.
"You can scale it down another 10 to 15 years," Intel CEO Craig Barrett said. "Nothing touches the economics of it."
As daunting as these challenges are, devising a solution could be worth billions of dollars in licensing fees and royalties because competitors will likely agree to adopt similar standards.
HP, for its part, is encouraging chipmakers to license its "crossbar latch," a matrix of interconnections that could replace transistors or interconnect wires. The concept, which relies partly on mathematical theories devised by mathematician Claude Shannon in the 1950s, took about 10 years to develop.
Not only are the crossbar components small, the sheer number of them means that many can fail but still allow the chip to function, a necessity for chips that will contain billions of components made of just a few atoms each.
"We're moving to an era where we can't be sure that the individual components are reliable," HP's Williams said. "What we decided to do is get way ahead of the train and build a side track that everyone would want to use."
Why Moore works
The magic of Moore's Law is that it allowed chipmakers--and, by extension, hardware manufacturers, software developers and even network carriers--to offer customers rapid technological improvements. Adding transistors to chips also adds performance or additional capabilities without necessarily raising costs.
"If you could get twice as much gas mileage three years down the road, you'd probably be prompted to buy a new car," said Dan Hutcheson, CEO of VLSI Research, which studies the semiconductor equipment industry.
The progress has been staggering. In 1955, the annual production of transistors could be measured in the millions. In 2003, production came to around a quintillion, or a trillion million, Hutcheson said.
In 1954, the average price of a transistor cost $5.52. In 2004, the average cost was 191 nanodollars, or 191 billionths of a dollar.
"The early reactions were, 'Wow, is this really true?" recalled Hans Stork, the CTO at Texas Instruments who was a graduate student when Moore's Law first began to gain traction. "It is unprecedented."
Yet boosting transistor production has required complex tools. In 1957, Hutcheson points out, transistors could be painted in wax with a 10-cent camel hair brush. Now, lithography machines cost up to $20 million while factories cost several billion.
Cramming millions of transistors into a tiny space has also created an energy crisis. Transistors need electricity, and pumping power into a confined space creates heat, which in turn can harm the entire system. A few decades ago, a lab at the University of Michigan used the heat pumped out by its Cray supercomputer to heat the garage.
Heat also raises costs. When IBM made chips with
In the past few years, the heat problem has crept into PCs. The water cooling in the latest Apple G5 computer adds about $50 to the price tag, Hutcheson said.
Watch for rocks
The next major change for chipmakers will come later this year when some companies begin to experiment with "immersion" lithography in 65-nanometer manufacturing. This involves drawing the circuits on chips with a laser while the wafer is immersed in purified water. The water refracts light, permitting it to draw finer circuits.
That's going to be a relatively easy step compared with 45-nanometer manufacturing, which should debut in late 2007 or early 2008. Then, many manufacturers will likely replace their silicon gates with gates made of metal.
Two to three years later, the move to 32-nanometer manufacturing will bring an overhaul in lithography systems, the machines used to draw circuits.
"By the time you reach 32-nanometer manufacturing, you are going to have to go from photo lithography to something else," said Michael Falcon, marketing director for Molecular Imprints. Molecular's Imprinter 250 will cost about $10 million, fully installed. An equivalent EUV, or Extreme Ultraviolet, machine will cost $45 million to $50 million, he said, a big difference considering that manufacturers will have to install about 10 in each factory.
Sometime around 2012 or so, the 22-nanometer manufacturing will debut. By then, the bag of tricks for extending silicon may start wearing out.
"I think we will get to 22-nanometer and maybe even a generation beyond that, but it's going to take real breakthroughs to make that happen," IBM's Theis said. By this time, chips may feature multiple gate transistors where electrons flow across more than one channel, as in today's transistors.
Subsequently, the industry may see the emergence of transistors or other structures made of silicon nanowires, strands of pure silicon. Besides being small and able to transport electrons quickly, nanowire and nanotube transistors could be cheap to make because they can assemble themselves.
Paolo Gargini, director of technology strategy at Intel, says 2015 will mark the beginning of the "integrated solutions" era, when nanotubes or other materials will be grafted onto silicon for hybrid chips. What these chips might consist of should be known by 2010, he said.
Other alternatives include quantum well transistors made of III-V materials. Phase change materials may also hit the market.
Then again, 40 years of investment and know-how in silicon transistors will make them tough to get rid of. Three-dimensional chips would allow engineers to continue to increase density without necessarily changing materials, although the feasibility is uncertain.
Predictions about the death of Moore's Law have been wrong several times, too. In 1978, IBM scientists predicted Moore's Law had only 10 years left. When they got to 1988, they said it would end in 10 years again. Moore himself thought it would end at 250-nanometer manufacturing, a landmark the industry passed in 1997.
"I am an optimist in that the creativity of people continues to surprise me," Texas Instruments' Stork said. "I am cautious in that there could be a time when the economics and the physics are at odds with each other."
21 commentsJoin the conversation! Add your comment