Even the most ingenious chip designs can be undone by manufacturing problems: Consider IBM Corp.s struggle to raise the megahertz ante for its PowerPC processors.
The company has managed to keep delivering chips without major interruptions. Nevertheless, issues with two of its key clients point at problems manufacturing its chips at their rated speeds. The PowerPC chips that IBM supplies to Apple Computer Inc. are slower than Apple expected, and Nvidia Corp.s GeForce 6800 card has yet to ship in appreciable quantities.
For its part, Somers, N.Y.-based IBM claims that it has weathered the transition from 130-nm to 90-nm as well as its competitors. Thats at least partially true, even if its faint praise: Advanced Micro Devices Inc., Intel Corp. and foundries such as Taiwan Semiconductor Manufacturing Co. have all stumbled during transitions to new process technologies.
However, on June 23, 2003, Apple CEO Steve Jobs told a crowd at its Worldwide Developers Conference in San Francisco that the company would be shipping 3GHz PowerPC desktops within 12 months. “Its a fantastic architecture, and its got legs,” Jobs said at the time. Its now a year later, and Apples fastest G5 desktops run at only 2.5GHz.
So far, most of the discussions with IBMs fab have dealt with the “defect density,” the number of devices on the wafer that actually work. But as Apples case points out, the issue may also be with IBMs ability to “speed-bin” chips to meet its customers demands.
A semiconductor fabs output is rated according to “yield,” which usually means the number of functional processors contained on the silicon wafer. If a chip fails to operate, it cant be sold. But a chip that fails to operate at a rated speed may also be a wasted effort, if a customer is positioning a microprocessor or other chip against a competitor.
“The issue today in most modern fabs is not whether it works or not but whether it yields effectively at a marketable speed,” said Nathan Brookwood, an analyst at Insight64 in Saratoga, Calif.
IBMs yields are, of course, a secret, and IBM executives declined to divulge specifics. “I dont believe IBM made that prediction publicly,” said Jesse Stein, IBMs PowerPC marketing programs manager, of the 3GHz prediction actually made by Jobs.
“Were ahead of our competition, although I dont want to disparage anybody,” Stein added. “The frequency jumps weve been able to achieve with the [PowerPC] 970FX have been greater than some of our competitors from 130 nm to 90 nm. Were getting a larger frequency jump, and were facing the challenges better than most.”
John Kelly, senior vice president and group executive for technology in IBMs Systems & Technology Group, acknowledged the defect issue in a conference call with reporters in May. “As our CFO John Joyce said when we reported our first-quarter earnings, our 200-millimeter yields are basically at or above our plan while our 300-millimeter yields have been improving but are not quite yet where we want them to be,” he said.
Kelly said the companys 130-nm, 300-mm defect densities are “showing rapid improvement.” About 50 percent of the companys manufacturing is at 130-nm processes or smaller, he said. “Finally, as John Joyce suggested, we expect to do a better job of meeting our customer demand in the second quarter,” Kelly said.
Since then, IBM has remained mum on the issue, according to analysts who follow the semiconductor manufacturing industry. Confidential defect-density reports shown to Risto Puhakka, vice president of VLSI Research Corp. in Santa Clara, Calif., has led him to believe that “IBMs problems are behind them,” Puhakka said.
“Weve talked a couple times since but havent disclosed it further,” Puhakka said. “The problem is that to nail it down youd have to go pretty deep on them.”
In some sense, aggressively ramping process technologies is out of the hands of an IBM or an Intel; its the responsibility of the manufacturers of lithography equipment. At an International Electronics Forum meeting in the Czech Republic in May, IBM executives seemed to pooh-pooh the idea that aggressively ramping process technology nodes was the secret to getting ahead in the chip industry.
“The real road maps going forward are going to be innovation road maps, not lithography road maps,” Bernie Meyerson, IBMs chief technology officer, said, as reported by EE Times. “Its not just lithography that is driving progress.”
IBMs 90-nm process technology combines silicon-on-insulator, copper interconnects and “strained silicon,” which lines up silicon atoms with a substrate whose atoms are placed farther apart. In the strained silicon, electrons experience less resistance and flow up to 70 percent faster, according to IBM. All three methods are designed to decrease electrical resistance within the chip and allow them to run faster. IBM has yet to combine these with low-k dielectrics, designed to reduce resistance further.
The problem low-k introduces is one of stability, Puhakka said. “Low-k dielectric presents a major integration issue in terms of the softness of that material,” he said. “All that upside becomes much more difficult. … SOI doesnt affect [the structure] that much; copper has some effect on it, placing the two materials side-by-side in that structure.”
“What I really expect is for the recipe to become more complex, and for them to talk about [adding] high-k metal gates,” designed to electrically separate the source from the drain in an electrical transistor, Puhakka said. High-k gates are most often used to minimize leakage current, a critical element in developing chips for mobile applications. Norman Rohrer, the chief architect of the PowerPC 970FX told eWEEK.com that the company is trying to massage its process technologies to push the current G5 chip into notebooks.
For its part, Intel spokesman Manny Vara challenged the idea that the scaling of process technologies from 130nm to 90nm and beyond is dead. “They [IBM] have been talking a bit lately about CMOS not scaling moving forward and how that can slow Moores Law, etc.,” he said from the companys Santa Clara headquarters. “We do not agree with their take. We agree that it is more difficult to continue to scale as you move to smaller sizes, but thats no different from what weve faced in the past.”
To date, IBM has filled a special niche in the manufacturing industry. In the late 1990s, IBM served as a key foundry for National Semiconductor Corp.s Cyrix X86 processor division, keeping half of the Cyrix MII and 686MX wafers it produced for Cyrix and reselling them on the open market as IBM-branded chips. A year later, after National and Integrated Device Technology Ltd. sold off their X86 chip divisions to Via Technology Inc., IBM pulled back from the foundry business, concentrating on manufacturing wafers based on its own chip designs.
In January 2003, however, IBM delicately re-entered the merchant chip world by announcing an agreement with AMD to co-develop process technology at the 65-nm and 45-nm manufacturing nodes. The deal did not involve a foundry component, unlike the agreement struck with graphics chip maker Nvidia Corp. two months later, in March. The two companies agreed that IBM would produce the Nvidia GeForce 6800 graphics chip at IBMs new $1.5 billion, 300-mm wafer fab in East Fishkill, N.Y. After being delayed for about two months, Nvidia representatives said that add-in cards based on the chips would begin shipping this month.
A senior executive at one electronics distributor said the shortages were being used to advantage by customers. To date, the majority of the Nvidia cards announced have been from the “Ultra” series, aggressively clocked to appeal to the high-end gaming market. “Their 6800 Ultra is what gamers want,” the executive said. “Theyre willing to pay anything for extreme anything. Its bragging rights. Theyre obsessed with the high end of the market. Thats where the money is.”
The PowerPC 970FX and Nvidia 6800 chips are manufactured in the same 300-mm IBM wafer fab in East Fishkill, IBM spokesman Christopher Andrews confirmed.
Taiwans Via Technologies has also agreed to use IBMs fabs to produce its low-power C5J “Esther” X86 microprocessor core. In January, when the manufacturing relationship was disclosed, Via said Esther was expected in the second half of this year.
In May, Via disclosed more details about the chip, including the expected 2GHz clock speed, removing the chips expected ship date from its announcement.
However, officials at the chipmaker said theyre still on track. “As far as Im aware, were not experiencing any particular issues moving to IBM and are pretty much on schedule,” said Richard Brown, Vias director of worldwide marketing.
“IBM really hasnt had a tradition of heavy emphasis on high-volume manufacturing,” Brookwood said. “Theyve always emphasized advanced technology first, to be first to market in advanced technology rather than high volume. Intel or Texas Instruments has been much better in ramping high-volume products. Its a different discipline; theres some learning to be done. As a result, some of the people that went to [IBM] on 90 nm and asked them to ramp their products quickly have been disappointed.”