Top tier locked in race to 90nm production
By Jack Robertson
EBN
(08/19/02 10:34 a.m. EST)
The next-generation 90nm semiconductor node coming late next year could determine the winners and losers among processor and logic chip manufacturers and foundries.
Vendors that jump ahead of rivals in the race to 90nm (0.09-micron) processing will leave their competitors in the dust, according to many analysts and chipmakers.
“The companies that get into 90nm production first will get a tremendous advantage in lower cost due to higher yields. The die shrink can also lead to much higher-performing devices,” said Dan Hutcheson, an analyst at VLSI Research Inc., San Jose. “Rivals who are late in 90nm process technology will fall behind and may not be able to catch up.”
And processor and logic ICs as a class may reach the next technology node before DRAM, according to Hutcheson.
Leading the 90nm pack are perennial processor rivals, Intel Corp. and Advanced Micro Devices Inc. Intel, Santa Clara, Calif., last week claimed it has made test chips with 90nm processing, achieving its smallest transistor gate length yet of 50nm. The small size means higher speeds and greater chip densities-as many as 330 million transistors on a die. Intel expects to be in full 90nm production in the second half of 2003.
AMD, Sunnyvale, Calif., claimed it will have prototype chips with 50nm transistor gate length on existing 130nm processes in the first quarter of 2003. Craig Sander, vice president of technology development, said by the time AMD transitions to 90nm processing in the second half of next year, transistor gate length will be less than 50nm.
Intel's first 90nm chip will be its highest-performance Pentium 4 processor yet, code-named Prescott, the company said. It will be initially produced at Intel's D1C development fab in Hillsboro, Ore., and later in planned 300mm-wafer fabs in Albuquerque, N.M., and Ireland. Mark Bohr, Intel fellow and director of processor architecture and integration, said the 90nm process will only be used in 300mm-wafer fabs.
AMD will launch 90nm with an advanced version of its Athlon Hammer desktop processor. AMD's first 90nm production will be at its existing 200mm-wafer fab in Dresden, Germany. The company's foundry partner, United Microelectronics Corp., will also use 90nm technology co-developed with AMD to make processors in one of its Taiwan fabs.
Texas Instruments Inc., Dallas, expects to qualify its 90nm process in the third quarter of 2003 and ramp production later in the year, a spokesman said. One of the first chips to use 90nm will be a new-generation UltraSparc processor that TI makes for Sun Microsystems Inc. TI is planning to deploy 90nm processes in its existing 300mm-wafer fab and in some 200mm fabs.
Infineon Technologies A.G.'s logic division expects to be in 90nm production by the end of 2003. Franz Neppl, senior vice president of core logic development at the Munich, Germany, company, said the UMCi 300mm joint venture fab coming on line in mid-2003 with 130nm production will quickly transition to 90nm.
Unlike MPU companies with large production runs of selected devices, broad portfolio logic vendors such as Infineon face greater 90nm challenges in making limited volumes of a wide variety of devices.
The low volumes typical of logic chips can actually increase costs and lead to higher prices at 90nm, Hutcheson said. That's because the process requires more costly lithography photomasks. Hutcheson estimated some 90nm photomask sets may cost up to $1 million each, twice the cost for a 0.13-micron mask set.
The higher cost is ultimately passed onto customers. Some ASICs processed at 0.13-micron have up to a $50 price adder simply due to costly photomasks, Hutcheson said.
Bill McClean, an analyst at IC Insights Inc., Scottsdale, Ariz., questioned the need for any rush to reach 90nm production, when the move to 300mm-wafer fabs alone will increase yields, leading to lower prices.
“With OEMs under severe price pressure [on their own products], can they really afford the initial high chip prices that will come with 90nm?” McClean asked.
One solution for logic vendors has been to outsource next-generation IC designs to foundries, which ostensibly have the production volume to offset the extra processing cost. But the success of this strategy depends on how fast even the biggest foundries can transition to 90nm -- and whether they will have capacity ready to meet demand.
Taiwan Semiconductor Manufacturing Co. Ltd. will launch 90nm production at the end of 2003 at its new 300mm Fab 12 in Hsinchu. A spokesman said it hasn't yet been decided whether to convert existing 200mm-wafer fabs to the new process.
UMC will enter 90nm pilot production at its 200mm-wafer development fab in the first half of next year. Its 300mm Fab 12A will add 90nm capability next year.
Steve Pelayo, an analyst at Morgan Stanley in San Francisco, wondered how well foundries will fare at 90nm, considering the difficulties many are having in ramping up 0.13-micron. Pelayo and VLSI Research's Hutcheson said foundries are getting only 20% to 50% yields on many 0.13-micron IC designs.
“Yields above 70% are the norm for effective mass production,” Pelayo said.
A TSMC spokesman said lower yields are to be expected with a new- generation manufacturing process. He said the relatively smooth transition from 0.25- to 0.18-micron created unrealistic expectations in the shift to 0.13-micron processing. TSMC is beginning to achieve effective yields at 0.13-micron and processing problems are being overcome, he added.
Infineon's Neppl agreed that 0.13-micron has been a major challenge for logic chipmakers, which unlike DRAM makers have a wide variety of IC designs in their portfolios. But, he said the learning- curve experience will benefit suppliers in the 90nm transition, since issues with larger wafers and new processing materials will have been solved.
However, a new lithography technology will come into play at 90nm, using argon fluoride 193nm-wavelength exposure tools -- mostly for the first time -- to pattern critical layers of the chip, Pelayo said.
Any unexpected trouble in ramping up the new lithography could repeat the yield troubles of the 0.13-micron node. |
