Like the weary traveler wondering "are we there yet," the DRAM industry is anxiously awaiting signs that the trail of attrition it has followed for much of the past year will finally cross into more sound economic terrain.
The numbers are sobering. In 1995, there were 28 DRAM vendors; by 2000 there were 18. And with Toshiba Corp. reportedly courting buyers for its memory business and cash-strapped Hynix Semiconductor Inc. struggling to meet its debt burden, the likelihood of further consolidation is strong, said Sherry Garber, an analyst at Semico Research Corp., Scottsdale, Ariz.
Even mighty Micron Technology Inc., arguably the most aggressive DRAM manufacturer in the industry, has fallen on hard times. The Boise, Idaho, company last week posted a net loss from continuing operations of $521 million on $3.9 billion in net sales for the fiscal year ended Aug. 30. That compares with net income of $1.5 billion on net sales of $6.3 billion the previous year.
Where does it go from here? Semico predicts 21% growth in DRAM revenue in 2002, to $14.9 billion, compared with the $12.3 billion expected this year. That's among the more optimistic of recent forecasts, and most analysts agree that the Sept. 11 terrorist attacks on the United States will have a negative impact that is still being measured.
"As disastrous as it was, we think it will have a short-term impact and the U.S. will return to business," said Garber, whose prediction for a DRAM rebound in 2002 is contingent upon a fourth-quarter economic recovery driven by improved consumer confidence, healthy holiday spending, and the successful launch of Micro-soft Corp.'s XP software.
PCs define the game
Yet in many ways, things remain the same. DRAM technology issues like latency, bus width, clock speeds, configuration, density, and cost will continue to exert their influences-and the PC market will still be the biggest game in town.
"The PC, midrange and high-end servers, and workstations are more than 65% of total DRAM usage," said Nam Kim, an analyst at iSuppli Corp., El Segundo, Calif. "In 2005 the major user of DRAM will still be PC applications, including server applications."
The electronics market in the coming year also will be faced with the choice of whether to use Direct Rambus DRAM, double-data-rate SDRAM, or both as the venerable single-data-rate SDRAM begins to give ground to higher-speed devices.
Arun Kamat, director of strategic marketing at Hynix Semiconductor in San Jose, predicted that DDR will command 40% to 50% of the DRAM market in 2002. However, if Direct RDRAM can penetrate the corporate desktop by the end of next year, it could lay claim to as much as 20% of the overall market, said Tom Quinn, vice president of marketing and business development at Samsung Semiconductor Inc., San Jose.
The goal of both approaches is to catch up with runaway microprocessor speeds, which have created a performance gap between the CPU and main memory. The chief issue, according to Kamat, is that DRAM is an inherently slow architecture, storing bits as an electric charge that must be refreshed every few cycles to retain data.
"There's no way DRAM will achieve parity with processor operating speeds," he said. "Of course, it makes up for that with wide buses, so the data coming out is still adequate for the processors."
Direct RDRAM and DDR use different interfaces, each of which is designed to move data off-chip at ever higher rates. Separately, DRAM developers are looking to solve another major bandwidth bottleneck-latency, the delay experienced when memory is slow to execute its first data access command.
Suppliers are refining their DRAM cores using virtual channel (VCDRAM), fast-cycle (FCRAM), reduced latency (RLDRAM), and other technologies. Combined, these specialty memories are expected to capture, at most, 10% of the DRAM market in 2002, but are becoming increasingly important to networking system developers.
"Latency is an important factor for any kind of numerical computing where the designer is interested in how quickly you can open pages in DRAM," said Jim Sogas, vice president of sales at Elpida Memory (USA) Inc., Santa Clara, Calif. Elpida is promoting its VCDRAM approach, which uses virtual channels or registers between the memory cell array and the I/O bus to enhance speed and efficiency. "We're all attacking the same problem from different directions," Sogas said.
Fujitsu Ltd., which developed FCRAM in the mid-90s, said its chips offer data access speeds of 25 to 30ns, twice as fast as standard SDRAM. The redesigned core enables a charge to be preloaded on the chip while it simultaneously writes data to or reads data from a cell. The precharge activates the cell immediately so it can accept a new instruction.
Infineon Technologies AG, Munich, Germany, and Micron are co-developing RLDRAM, an architecture that uses internal logic on the memory chip to reduce the time between access and the transmission of first data.
"RLDRAM is for the networking area and will be a niche product," said Peter Schaefer, vice president of memory products at Infineon Technologies North America Corp. "The customers will decide which architecture fits best into their roadmap." OR