The so-called DDR400 specification, if it reaches the light of day, may get limited penetration in specialty applications like graphics cards. But for PC makers, a more likely scenario will be to leap from first-generation DDR to next-generation DDR-2 devices when they're ready to go beyond 333 MHz, according to several memory makers at the recent Platform Conference here.

The DDR-2 specification will allow higher data transfer speeds by writing to more parts of a memory array in a parallel fashion for every clock cycle.

Mike Seibert, strategic marketing manager for Micron Technology Inc. (Boise, Idaho), said DDR400 may find a home in some applications, but "certainly I wouldn't expect it to be in mainstream desktops. We all know Intel is going to 533-MHz [front-side bus] and we're never going to get DDR-1 there. It's a dead-end investment."

Elpida Memory Inc. also sees 333 MHz as the cut-off point for DDR-1 in PCs. "DDR400 is possible. But to put it out there as a main memory is a bit of a stretch," said Jim Sogas, vice president of sales for Elpida (Santa Clara, Calif.). "Even for SDRAM at PC100, going to PC133 was a stretch."

DRAM makers say it would require extraordinary effort to design, test and produce modules that support DDR400. Graphics applications can often get away with using faster speeds because the memory load is fixed and is much smaller than PC main memory.

PCs, on the other hand, must be able to work reliably using anywhere from one to six modules. That kind of variation makes it difficult to analyze and test simulate the boards without making costly upgrades on design tools, said Sogas.

Indeed, Denali Software Inc., which provides memory simulation models and controller intellectual property, said more and more customers are asking for help designing fast memory subsystems.

"For ASIC manufacturers, the amount of engineering support we have to apply goes up very rapidly as clock rates go up," said Lane Mason, memory market analyst for Denali (Palo Alto, Calif.)

DRAM makers are also concerned about the cost of producing ever faster versions of DDR-1. "400 MHz DDR-1 specs are too fast and won't yield well," Micron's Seibert said.

To be cost effective enough to sell, manufacturers must be able to reach 30 percent yield per wafer lot for any given speed bin. It would take nothing less than a redesign of the DDR core itself to get there, said Chee Ho, director of product marketing with Infineon Technologies in San Jose, Calif.

"If we actively pursue DDR doing 400, we would have to recreate the memory core, and that's a two year effort," Ho said.

Mortal concerns

Moreover, the life expectancy of DDR400 would likely be cut short by the imminent introduction of DDR-2 architectures. Such specification creep has proved to be a drag on DRAM makers. In the last four years, for example, Infineon has introduced five architectures, Ho said.

"We're not getting the R&D payback because the lifespan is too short. It takes 18 months for a memory to reach its peak," he said.

DDR333 also may not be around long because DDR-2 is right around the corner. DDR333 chip sets are starting to trickle out of Taiwanese core logic vendors this year, but high volume demand won't come until next year. DDR-2 chip sets are expected to hit the market too next year, followed by volume production of DDR-2 DRAMs in 2004 or 2005. "It's going to be a short life for DDR333," said Ho.

Seibert said Micron will start selling DDR333 in TSOP and chip-scale packages later this year. Sogas said Elpida's DDR333 design activity will pick up late this year, but said volume production won't come until later next year.

"A large percentage of the production yield has to be at that speed grade," Sogas said. "To see any kind of volume will be well into '03."