A possible four-way collaboration on 1-Gbit DRAM, now in the early stages of negotiations, could link IBM and Siemens' work on trench technology with Toshiba's and Fujitsu's recently announced intention to pursue stacked capacitors. A new collaboration commitment with Apple Computer Inc. and Motorola Inc. could resolve the split over the PowerPC architecture that emerged last spring when IBM shunned the enhanced instruction set of Motorola's Altivec processor.

A merger being sought in the world of wired communications could parallel IBM's acquisition of CommQuest Technologies Corp. for wireless chip sets earlier this year, perhaps pushing IBM's silicon-germanium technology further toward commercial markets. And the search is on for backers for an IBM-defined on-chip bus scheme that the company believes could be a key enabler for the core-based design-reuse movement.

The $7 billion-plus microelectronics unit repeatedly grabbed headlines in 1998 for its technical advances in copper interconnects, silicon-on-insulator, silicon germanium and embedded dynamic RAM. Copper interconnects for processors and ASICs went into production, as did an SiGe process for wireless and Sonet markets.

Next year, IBM will push its silicon-on-insulator (SOI) capabilities out to the merchant market and will combine SOI and copper in a PowerPC processor. Late in 1999 or early in 2000, a fourth technology — an embedded DRAM capability at 0.18-micron design rules — will go into production, initially for high-performance networking, in which the on-chip bandwidth possible between embedded DRAM and logic can speed transaction processing.

1999 also presents IBM with a number of sticky challenges. The DRAM-development cooperative formed in 1992 among IBM, Siemens A.G. and Toshiba Corp. must be restructured now that Toshiba has joined Fujitsu to work on trench technology. And IBM must push down the costs inherent in its new process technologies if it is to find a ready market for such products as cellular-phone ICs and the SOI digital-still-camera chips being codesigned with Sanyo Electric Corp. (Osaka, Japan).

United front
Attardo said IBM and Motorola's Semiconductor Products Sector soon will be able to present Apple Computer with a united front for the PowerPC. He declared that the two companies "will come to a solution" on how to adopt the Altivec instruction set and architecture, developed by Motorola but initially shunned by IBM. Altivec adds a vector-processing unit and significant instruction-set enhancements to the PowerPC architecture.

Attardo said he and Motorola Semiconductor Products Sector president Hector Ruiz "are good friends" and predicted that the matter will be resolved. "The three companies are motivated to cooperate, and we will cooperate," he said.

Similarly, IBM is seeking to mend fences on its 1-Gbit DRAM development road map. Harry Calhoun, director of IBM's Semiconductor Research and Development Center, said IBM managers are in the early stages of discussion with Toshiba and are in very early discussions with Fujitsu. The goal, he said, is to cooperate on the basic process technologies: "lithography, etch, the back end of the line, where the directions are not clear."

While IBM remains committed to trench-capacitor technology through the 0.12-micron generation, Calhoun said, it is willing to explore stacked-capacitor technologies with its research partners.

"We still may work together on 1-Gbit-and-beyond technologies," Calhoun said. "The four companies have an intention to tie the work at the two development centers together, and we are in the very preliminary, explorative stages about how to do that. There is an opportunity to share costs and risk in 1 Gbit and beyond, where the development costs and expenses are astronomical."

Toshiba still wants to work with IBM, a spokesman said, but it has not received any formal overtures. If one were made, Toshiba would be willing to discuss a trilateral relationship with Fujitsu. "At this moment, there is no official proposal from IBM," the spokesman said. "We are willing to exchange information on 1-Gbit DRAM development with IBM, and we want to continue cooperating with them."

IBM, meanwhile, is keeping its options open. "We are looking at any number of different potential partners, but nothing is cooked [decided]," Attardo said. While there may be some sharing with Toshiba, he said, "they are going down a stacked path, and we [IBM and Siemens] are going down a trench path."

Attardo said he remains assured by his counterparts in Munich that Siemens will stick by its partnership with IBM in both 1-Gbit-DRAM development and in the 16-Mbit-manufacturing joint venture in Essonnes, France. He said IBM will continue to look for a partner to replace Toshiba in a three-way alliance, adding that more than three partners increases management complexity exponentially.

IBM will stay in commodity DRAMs but will continue to scale back (DRAMs account for about 25 percent of division revenues), Attardo added. As the DRAM industry continues to consolidate, IBM won't be one of the largest five or six vendors, he said. But he noted that DRAMs and logic share about 75 percent of their process steps and that DRAMs remain an important process driver at IBM.

Nevertheless, ASICs, not DRAMs or microprocessors, drive the division's product strategy. IBM had $1.541 billion in ASIC revenue last year, and growth in 1997 was a phenomenal 70 percent. That rocketed IBM past many of its competitors in the ASIC industry. According to Dataquest Inc. (San Jose, Calif.), IBM is now the second-largest ASIC vendor worldwide, behind NEC Corp., and is about equal to Lucent Microelectronics in cell-based ICs.

IBM's ASIC operation did 106 designs this year, a number that may double next year. Chris King, vice president of s marketing and field engineering, said 300 engineers now work in IBM design centers worldwide.

King said IBM's average gate count per design was 1.3 million in 1998, compared with the industry average of about 500,000. Dataquest predicts the industrywide average won't hit a million until 2001.

A key driver for the ASIC business could be IBM's on-chip buses for linking intellectual-property cores: a high-performance data bus and a peripheral bus for I/O circuits. Bruce Beers, director of ASIC products, said IBM hopes to win support for its buses among other companies and, perhaps, from the Virtual Socket Interface (VSI) alliance, which to date has supported a more generic, non-proprietary bus "wrapper."

Though Beers said he recently had a "good meeting" with VSI president Howard Sachs, IBM and the other members of the ASIC Industry Council have remained relatively distant from the alliance, largely because of VSI's intellectual-property sharing requirements.

The issue is important largely because success in stitching together IP blocks is essential to the ability to consume the much larger gate counts possible with 0.25-micron and 0.18-micron process technologies, where IBM wants to compete.

"We are using the old, ugly methodology," said Beers. "We use the same methodology for an 8 million-gate design, which we did recently for a data-processing customer, as we did for a half-a-million-gate design a few years ago. The industry is not making enough progress in terms of design and cosimulation tools."

SiGe heads for big time
IBM's plan to use its silicon-germanium technology to penetrate the wireless space, with both global-positioning-satellite (GPS) and cellular chip sets, is perhaps the most intriguing aspect of its OEM thrust. Much of IBM's growth in ASICs has been in communications. Silicon germanium heterojunction bipolar transistors (HBTs), used in a BiCMOS process, present IBM with a potent weapon in the fast-growing communications sector.

Attardo said IBM's February acquisition of CommQuest Technologies (San Diego) was done to "exploit silicon germanium" in the wireless-phone market. IBM is contemplating an acquisition in the wired-communications sector as well, he said, where it would marry its SiGe technology with the design skills of the acquired company.

Former Motorolan Ron Soicher, now an IBM vice president of RF communications and storage products, said the SiGe business plan calls for $1 billion in revenue by 2002.

"There are a lot of opportunities for silicon germanium," Soicher said. "To do 10-Gbit Sonet driver ICs, you need to do a 40-GHz process, and we can do that with a 0.5-micron 5HP process that is running in the 970 fab — the older of the two fabs in Burlington — which is a completely depreciated facility.

"We are running 600 wafers a day through that fab, and by adding just a few tools we can bring up a SiGe process there," Soicher said.

Next year, SiGe HBTs will be added to the 0.35-micron 6SF CMOS process, with four times as many CMOS transistors, giving IBM the ability to "do a lot of CMOS gates with an on-chip RF capability," Soicher said.

The next step is to combine a newer bipolar transistor with the 7SF CMOS process, a BiCMOS process that can handle 40-Gbit/second Sonet "with some tricks."

The highest potential volumes for SiGe parts, of course, are in cellular handsets, and here IBM is taking what Soicher called a "bowling-pin strategy." CommQuest is developing its Tri-band chip sets with SiGe and will offer its turnkey solution to cell-phone vendors that focus on retailing and support rather than on in-house technical development.

"We have more than 10 engagements with customers who want to use SiGe to replace GaAs. We have more than five companies that want to use their baseband CMOS architecture on one chip and SiGe for the RF portion, and one or two customers are working on complete radios-on-a-chip," Soicher said.

The first goal, he said, "is to replace the traditional radio with an RF/IF capability in SiGe. Where the pure CMOS portion of the design goes depends on the customer, and whether we absorb the CMOS depends on the customer partitioning. But there are a lot of applications that can use silicon germanium."

Moving copper, SOI and SiGe out of IBM — where design engineers are familiar with a proprietary set of design tools, device models and verification and test techniques — could prove to be one of IBM's biggest challenges next year.

In silicon germanium, for example, technically savvy customers like Northern Telecom, Alcatel and Motorola worked with IBM to create their own SiGe-based circuits. Getting customers to work with SiGe, SOI, copper and other new technologies leaves IBM with a major support task, so the microelectronics division has been beefed up to support commercial customers. SiGe designs can be done either at Waltham, Mass., or in New York, for example.

Calhoun said much of the underlying innovation in process technology can be tracked back to early efforts to develop fail-safe 390 mainframes. "No other company has anything like" the Semiconductor Research and Development Center, he said. "We run our R&D fab 24 hours a day, seven days a week, with 150 wafer starts a day, developing three generations of logic at the same time. That is where we learned to do SOI and SiGe at high yields, and the whole emphasis now is on developing technologies that can be brought to manufacture.

"Sure, there is a lot of debate about copper and SOI and SiGe," Calhoun said. "But we think we were right."

— Richard Wallace, Rick Boyd-Merritt and Anthony Cataldo contributed to this report.