A Sea Change for Semiconductors

This is part of IEEE Spectrum's special report: What's Wrong—What's Next: 2003 Technology Forecast & Review.

Skies were blue and trade winds were steady as the SS Semiconductor set sail in early 2001. The previous year's haul had set records, and forecasters were heralding yet another banner year. But powerful forces were to shake even this vessel's sturdy timbers.

A low-pressure center caused by the bursting dot-com bubble and warm moist Y2K air from around the globe stirred up a storm so violent that it shredded the mainsail. Gales of corporate malfeasance from the southwest and typhoons of terrorism from the east swept wave after wave of stock-market volatility over the decks. While many of the crew were tossed overboard, cargo destined for distant ports sat idle in the holds.

It was a long time before the winds grew calm enough for the crew to repair the battered boat and limp to port. But customers there, who had also been buffeted by the foul weather, were scarce. And the ship's officers were forced to sell their cargo at a fraction of its original value.

Now, almost two years after the near-fatal voyage, forecasters are again predicting blue skies, and officers and crew are guardedly optimistic that good fortune awaits in the year ahead.

"It was a perfect storm of external events," says Ray Burgess, corporate vice president and director of strategy for Motorola Inc.'s Semiconductor Products Sector (Austin, Texas). "Y2K, the Internet bust, and the telecom depression combined with an overzealous industry believing itself to be looking at 25 percent growth." The result was tremendous overcapacity in the semiconductor industry.

"I have been in this industry since 1980 and I have seen a lot of ups and downs," Burgess told IEEE Spectrum, "And this is the worst downturn that the industry has ever had, both psychologically and from the standpoint of the numbers. If you take a rolling five-year average growth rate, this year [2002] we are at the lowest five-year compound growth rate in the history of the semiconductor industry."

According to Burgess, the scenario has been something like this: the last two years of the 20th century were extremely profitable for the industry. Corporations were spending a lot of money on Y2K, upgrading their computers and network infrastructures for better fault tolerance. As late as the fourth quarter of 2000, analysts were predicting that revenues would grow 25 percent in 2001 to US $250 billion from an all-time high of more than $200 billion in 2000. Anticipating huge increases in demand for chips, semiconductor companies were investing billions in new chip fabrication lines and equipment.

But then in early 2001 the implosion of the dot-com bubble—which was actually sparked in late 2000—began to take its toll. Sales of networking equipment, disk drives, and computers for Web hosting farms plunged, driving down demand for chips that go into these systems. And because the chip fab lines had to be kept full in order to be cost-effective, IC prices dropped dramatically.

Semiconductor revenues are cyclic, reflecting the boom-bust capital equipment cycle, notes Burgess. "And then if the demand drops in the low part of the cycle, it dramatically exacerbates the situation."

But Moore's Law may also be playing into the IC price decline. With each new generation of semiconductors, the same functions cost less to produce, observes Bijan Davari, vice president of technology and emerging products at IBM Microelectronics Division (Hopewell Junction, N.Y.) and an IEEE Fellow. The reason is that the same functionality can be fit on a smaller chip in a smaller package. Smaller chips mean that more of them can be diced from a wafer at little extra cost. Productivity enhancement is Davari's term for this increasing functionality for a given cost.

As long as chip consumption grows faster than productivity enhancement, revenues remain healthy. "From the late 1970s up to the mid-1990s, the productivity enhancement was about 100 times for every 10 years. And the consumption rate was about 400 for every 10 years. So revenue increased about four times for every 10 years," says Davari. But since the mid-1990s, productivity enhancement has risen by 200-300 times per 10 years. So in effect the capabilities of chips are increasing faster than customers are buying them. "We are selling more and more chips, but revenues are staying flat or going down."

"So how do we fix the problem?" Davari asks. One solution would be to develop applications that will compel users to upgrade their electronics, be they PCs, PDAs, cars, or microwave ovens. The IEEE Fellow can see that happening in some interesting ways. "Take speech," he says. "It is very processing intensive, and we have not crossed the threshold on the capability that would motivate the average person to use it. It's not good enough. When speech recognition and speech translation cross that threshold, they can fuel a lot of growth."

A break in the clouds

It's not obvious, given the roller-coaster ride of technology stocks and wholesale layoffs, but the semiconductor industry is actually entering its second year of recovery. The low point in worldwide sales was the third quarter of 2001. In the same quarter of 2002, sales were up 20 percent. Over the entire year, though, total revenues in 2002 were about the same as in 2001.

"The recovery of the industry has picked up breadth and strength as the year [2002] has progressed," George Scalise, president of the Semiconductor Industry Association (SIA, San Jose, Calif.) observed in November. Global sales in the third quarter of 2002 rose 21 percent over the same quarter of 2001, he said. As to the future, the 2002-05 forecast released by the SIA in November predicts that worldwide sales of semiconductors will grow by 19.8 percent in 2003 and by 21.7 percent in 2004, surpassing the peak revenues of 2000.

The SIA believes that the huge growth of revenues in 2000 was an anomaly that pushed compound annual growth rates above 16 percent. "Over the long run," said W.J. ("Jerry") Sanders III, founder and chairman of the board of Advanced Micro Devices Inc. (AMD, Sunnyvale, Calif.), who delivered the forecast, "we believe that compound annual growth rates in the 8-10-percent range will be the norm."

Consumption in the Asia-Pacific region has risen particularly rapidly because of the growth of contract manufacturing there, according to Sanders. Semiconductor sales in that region are projected to increase 37 percent, up from 29 percent in 2001. This represents "the largest one-year shift in consumption patterns in history," he said.

In terms of market segments, ICs for consumer products, like digital video, audio, games, and cameras, have fared pretty well throughout this entire downturn, according to semiconductor analyst Christie Van Gaal of In-Stat/MDR (San Jose, Calif.), speaking at the 2002 Microprocessor Forum, held October 14-17 in San Jose, Calif. ICs for automotive applications have also done reasonably well.

And cellphone sales, too, are helping the semiconductor industry out of its doldrums. The SIA also reported in November that "the wireless market continues to be the strongest single growth market for chips, as new consumers come online in record numbers in Asian markets and existing subscribers upgrade to new 2.5/3G cellphone technologies."

But these are much smaller markets than communications and computers, which together made up about 80 percent of the semiconductor market in 2000, according to the SIA. The importance of the communications market to the IC industry has been growing since the late 1990s when semiconductor companies, perceiving a maturing of PC technology, shifted their R&D emphasis from computer chips to communications ICs. So when the bottom fell out of the communications industry, it was even worse for chip makers than it might have been otherwise.

"We all believed our own hype and so we all invested heavily in areas that had to do with the Internet and the supporting infrastructure. We were anticipating that high growth of Internet services would create lots of traffic and so from there lots of systems," says Dado Banatao, a managing partner of Tallwood Venture Capital (Palo Alto, Calif.).

Still, the coming years should see an increase in demand for ICs for the equipment that supports manufacturers putting their product specifications and price lists up on the Web, and allows customers to order them online. In the same way, a growing number of consumers are also ordering products on the Web. Banatao points out, though, that new technologies will have to be developed if companies are to take full advantage of such Web services.

"Hopefully, in the next two years we will be able to get out of the mess that we are in and provide some solutions for the real-time requirements of the enterprise [to interact with its suppliers and customers online]," he says.

Hopeful signs

The first signs of recovery in the communications market will occur at the edges of the network—PCs, routers, and wireless local-area networks, says Motorola's Burgess. The need for companies to move far more data faster is prompting them to upgrade their local-area networks from 100-Mb/s to 1-Gb/s Ethernet. In doing so, they will also have to upgrade their computers, switches, and routers to handle the higher data rates.

There is also much work to do on improving corporate network security. Today, there are the secure socket layers standard and the firewall. But the system software also has to be impenetrable to hackers, and servers must have the interfaces for building in security code. Another factor is encryption. "Those algorithms are so compute-intensive that the only way to do it without creating bottlenecks in the server is to do it in hardware," says Banatao. So the demand for high-powered ICs that can rapidly encrypt and decrypt data will intensify.

Recovery for the ICs to be found in the core of the network, where large switches and routers that handle long-haul transmissions reside, will be slower. Burgess thinks that it will be at least six months before that market segment begins to rebound. When the demand for bandwidth from the edge devices catches up with the core infrastructure, the network will be stretched again, driving demand for core telecom equipment.

"We need demand [for bits] to increase before network infrastructure gets back to a growth phase," says Burgess. "Demand increases when more people buy set-top boxes, cellphones, higher-speed PCs, and when more companies start to audio- and videocast." Developing countries, such as China and India, will also do their part, fattening demand for core components as their citizens come online over the next several years [see "Some Chip Makers Sink, Some Swim"].

Converging currents

Pushed largely by advances in wireless technology, communications and computing will continue to converge. Note the ability to surf the Web from your cellphone and the growing pervasiveness of IEEE 802.11 hot spots. Not just the local coffee shop, but also airports, hotels, conference centers, and even whole city centers are putting in wireless access points. Imagine: sit down on a park bench, turn on your laptop, and voilà! You are connected.

Intel is banking on convergence. At last September's Intel Developer Forum, Paul Otellini, the company's president and chief operating officer, described his vision for the third era of computing, which will be built around devices that meld the capabilities of computing and the wireless infrastructure. He admits convergence is not a new topic, but believes it's finally about to become reality. As chip technology advances, it will become easier and less expensive to deliver the local intelligence needed to make convergence happen.

The star in Intel's convergence plans is its Banias platform, a microprocessor and chipset designed for laptops, PDAs, and other portables. Banias is the first microprocessor that Intel has designed from the ground up for portable applications. Previous Intel chips for portables have been recycled desktop processors, says Tom Halfhill, senior analyst for In-Stat/MDR. And then there is just the hint of a possibility that the microprocessor giant is feeling a wee bit of pressure from the Crusoe microprocessor from Transmeta Corp. (Santa Clara, Calif.), introduced in 2000 as the first designed specifically for low power.

Mooly Eden, director of Intel's Israel Development Center (Haifa) where Banias was developed, told the audience at last year's Microprocessor Forum that his engineers paid strict attention to tradeoffs between power consumption and performance in selecting features for the technology. For one thing, the stages that an instruction must step through during execution are fewer than in Intel processors for desktop PCs. For another, the processor has become better at predicting which of two possible streams of instructions will be executed. These improvements mean fewer wasted cycles, hence less wasted power.

Banias is not only low power, said Mike Trainor, Intel mobility enabling manager, at last September's developer forum. The Banias chip set gives developers many ways to let electronics communicate over wired and wireless networks. It contains circuitry for both IEEE 802.11a and 802.11b and is also compatible with Bluetooth and wireless wide-area-network technologies.

In the future, PC users will be able to wirelessly send digital photos and videos to their TVs for better viewing or to send downloaded music to their compact-disk players for better listening, encouraging demand for new generations of electronics with these capabilities.

But if the availability of network communications doesn't sell computers, maybe rock-bottom prices will. At press time, Wal-Mart Stores Inc.'s online store was promising a Microtel PC for US $199, minus display and floppy drive. The operating system (OS) and the microprocessor are what make this PC unbelievably cheap. The OS, called Lindows, is based on the Linux operating system, which costs almost nothing. The VIA C3 microprocessor comes from Centaur Technology (Austin, Texas), a subsidiary of Taiwan's VIA Technologies Inc., and sells for under $25.

It is a stripped-down X86 microprocessor that runs at 800 MHz, with performance on a par with Intel Celeron and AMD Duron chips running at the same speed. Keep in mind though, that Celerons and Durons usually run much faster than the C3: the Microtel PC also comes in slightly pricier versions, which include either a 1.3-GHz Duron or a 1.8-MHz Celeron. The Celeron version is being offered at $498.

But Centaur president Glenn Henry founded the company because he believes in inexpensive, low-power microprocessors just fast enough for most people's needs. "The difference between $200 and $500 is a lot for applications that are not performance driven," he told 2002 Microprocessor Forum attendees.

Moving up to 64 bits

Higher up on the microprocessor value chain, two companies hope to spur sales by offering 64-bit microprocessors for PCs, low-end workstations, and servers. Sixty-four-bit microprocessors, like the UltraSparc from Sun Microsystems Inc. (Palo Alto, Calif.) and IBM's Power4, are typically used in high-end servers and engineering workstations, whereas desktops and low-end servers use 32-bit chips. The main advantage of a 64-bit design is that it can directly address an address space in a computer's main memory that is almost unfathomably large—more than 18 EB. An EB, or exabyte, is 1018 bytes, which is more than 10 times all the print, film, optical, and magnetic content produced in the world in 1999. (Computer systems will not have that much memory for some time.) A 64-bit design also has 64-bit-wide registers and data paths for handling larger chunks of data at a time.

AMD and IBM have each decided that it is time for 64 bits to move closer to the low end. As Peter Sandon, senior PowerPC architect at IBM Microelectronics, explains, 32-bit processing is fine for many market segments, but "we feel that efficient implementation of a 64-bit architecture will enable more applications to take advantage of that 64-bit functionality." Such applications include games and large storage systems and databases. High-performance support for 32-bit processing in 64-bit systems is also needed so that legacy 32-bit applications can continue to run efficiently.

AMD will offer two versions of its 64-bit architecture. For desktop and portable systems, the company was to have begun volume shipments of its Athlon 64 by the end of 2002. Shipments of its enterprise server processor, Opteron, will begin in the first half of this year. It will go head to head with Intel's Xeon and Itanium 2 processors. Xeon is a 32-bit X86 design and Itanium 2 comes out of the 64-bit architecture, IA-64, co-developed with Hewlett-Packard Co. (Palo Alto, Calif.).

Itanium 2 is the second-generation IA-64 design from Intel. The first Itanium began volume shipments in May 2001, to a lukewarm reception. Its mediocre integer performance and huge power dissipation gave server developers little to cheer about. Nevertheless, many climbed onto the Itanium bandwagon because Intel had promised to do much better in subsequent generations.

Now Itanium 2 is indubitably outperforming leading server-class processors—except for Intel's own 2.5-GHz Xeon. Power remains an issue, however. Early in 2002, Hewlett-Packard and Intel microprocessor architects reported a power dissipation of 130 W for Itanium 2. That number will probably decline as Intel moves Itanium 2 to a more advanced manufacturing process, but it may not decline enough to satisfy managers of large server farms, who must deal with the heat generated by thousands of processors contained in a small volume.

All in all, the jury is still out on Itanium. The biggest challenge is "convincing IT managers to adopt a new architecture during these troubled times," wrote Peter Glaskowsky, editor-in-chief of Microprocessor Report, in the August issue of that newsletter. "Even after HP and Intel have pumped more than $1 billion into Itanium, its future success is not yet secure."

Another factor is the need for software and system developers to recompile their software code for a new architecture. Of course, the need for developers of Itanium system to adopt a new architecture may work to AMD's advantage. The company is banking on the fact that its X86-64 is still an X86 architecture, which should give developers a relatively painless migration path from 32- to 64-bit applications.

Intel, meanwhile, has announced no plans to offer a 64-bit X86 chip. It believes that hyper-threading, now available on Xeon and Pentium 4 processors, will do much more for performance than a move to 64 bits. In hyper-threading, code is broken up during compilation into multiple streams that can be run simultaneously on the same processor, so that applications may exploit chip resources that would otherwise be idle.

The technique also allows multitasking, in which separate applications are run at the same time on the same processor. (On nonthreaded X86 machines, applications may appear to be doing that—for example, playing a music CD while writing a document in Microsoft Word. But in actuality, the processor is switching between the two applications many times a second.) The precursor to hyper-threading is multithreading, which has been around for many years. Multithreading is a technique for hiding the clock cycles wasted when the processor must wait for data with which to execute an instruction. But instead of sitting idle until data arrives, a multithreaded processor switches to a different thread and continues executing instructions.

Shrinking size calls for new materials

Economic downturn or no, semiconductor companies are pushing ahead with new technologies. At the top of just about everyone's list is 90-nm technology, a manufacturing process capable of producing circuit structures as small as 90 nm. Most major semiconductor companies and foundries have announced their plans to bring 90 nm on line this year, but it won't be easy.

"Everybody is talking as if 90-nm technology is already here and it's no big deal," says Kazu Yamada, general manager of the technology foundation division, Broadband LSI Technology, NEC Electronics Corp. (Kawasaki, Japan). "But some people are still seeing processing instabilities in the current 130-nm technology as they bring it into mass production." One reason is that some of the materials used to fabricate the wafers have changed. The aluminum wires, or interconnects, that join the chip's circuits to one another were replaced by copper. And the insulating layers, which until recently were silicon dioxide, were replaced with new materials that reduce interference between adjacent wires because they have a lower dielectric constant.

These changes have taken effort to perfect, and 90-nm technology will add even more challenges for process engineering, production yield, and process stability, says Yamada. Design complexity is also of concern. "As we move from 180 nm to 130 nm to 90 nm, the difficulty of both developing the manufacturing technology and doing design in the technology is going up very fast," says Dennis Buss, vice president of R&D for Texas Instruments Inc. (Dallas) and an IEEE Fellow.

In nurturing a new manufacturing process, materials changes are the hardest to deal with, says IBM's Davari. "Each takes about 10 years to go from the feasibility stage to the manufacturing stage. We spent more than 10 years getting copper [interconnect technology] ready....And a company that wants to make that switch without that experience—it's a difficult thing to do."

Certainly a cost and benefit tradeoff confronts companies moving to more advanced manufacturing processes. State-of-the-art fabrication lines are now well over a billion dollars. A mask set for 90-nm designs can run more than a million. But what companies get in return for the move is a reduction in the cost of making a chip. "Every generation, the chip size goes down by a factor of two," notes Buss. "And even if the wafer cost goes up a little bit, the overall cost comes down dramatically. But that means increasingly large design teams and increasingly large design challenges."

Keeping a lid on power density

Another high-priority item on chip makers' worry list is power density. "As you shrink the minimum feature dimensions with every successive generation of the technology, the power density pretty much doubles if the same set of materials is used," says IBM's Davari. Success here presumes that the increase in power entailed by higher operating frequencies can be counterbalanced by a reduction in operating voltage.

"Keeping a lid on power is one reason that process engineers keep reducing the operating voltage," says Joe Mogab, director of Motorola's advanced products research and development laboratory (Austin, Texas). "But it's not working well. We continue to see—for the high-performance chips—an increase in power consumption."

No more. As semiconductor technology moves from 90 nm to 65 nm (due in 2005), the power density can no longer be permitted to double. New materials will be a must not just for high performance but for all applications. Almost every part of the silicon circuit from the wires to the transistor materials to the silicon wafer itself will be replaced by materials engineered to use less power and improve performance.

Nor are circuit designers oblivious to the problem. They are enabling chips to turn off circuits when they are not in use and to turn down the operating voltage and frequency for applications that demand less than peak performance. In short, it will take all the tricks that process and design engineers can devise to keep power consumption within bounds.