Spansion Seeks DRAM Replacement in Servers
David Lammers, News Editor -- Semiconductor International, 6/18/2008 8:56:00 AM
“DRAM has been the technology of choice for the search engine server farms, largely because of the fast read times. But DRAM must be refreshed. With MirrorBit flash, we can multiply the packing density by at least four times for the same amount of power,” Cambou said.
Later this month, Spansion plans to provide more details of the strategy, announcing search engine partners, publishing white papers and inviting press to demonstrations of the approach.
In recent months, more attention has been paid to the energy consumed by server farms, which requires large amounts of cooling. Cambou said flash consumes about one-fourth the power of DRAM, which must be refreshed. He said Spansion's two bit per cell nitride-based flash can provide 4-8× the packing density of DRAM-based memory modules used today.
The server farm strategy comes six months after Spansion decided not to pursue the solid-state drive (SSD) market, largely because of sharp price declines by NAND flash vendors. “Six to nine months ago, we decided to pull out of the SSD market. We realized that we don’t need to do that to stay on our growth path. Competing with hard disk drives, which are very cheap, is going to be tough. And we saw Samsung and Toshiba offering dirt-cheap prices for NAND and building mega factories to get ready for that market. We didn’t see the merits in a 'me too' strategy, competing against NAND,” he said.
Instead, Spansion will pursue the DRAM replacement strategy using its 300 mm SP1 factory now ramping in northeast Japan to reduce bit costs. Cambou said Spansion is likely to remain the only NOR flash vendor with a 300 mm fab for some time, claiming that Numonyx (Geneva) — the joint venture formed by Intel Corp. (Santa Clara, Calif.) and STMicroelectronics (Geneva) — does not have the cash to equip the empty 300 mm shell in Catania, Italy.
Spansion is pursuing “a node a year” strategy, sampling 45 nm technology at the end of 2008 and 32 nm technology a year later. “As we ramp up production in Japan and at SMIC, our costs per wafer go down. At 2000 wafers per week, we already have a fundamental cost advantage, and we can fund capacity expansions from cash from operations.”
With a claimed smaller die size, Cambou said the MirrorBit flash can compete with DRAM on a cost-per-bit basis. “Our die size is one half that of DRAM. Our wafer cost is 25% higher, which means we can be 40% cheaper than DRAM at the 2 Gb and 4 Gb density. What our partners in the server farm business are telling us is that if we can match the DRAM price, they would be happy because of the power advantage. We don’t have to give up those margins.”
Also, Cambou said DRAM vendors face scaling challenges, including building the tall stacked capacitors. “The aspect ratio is horrible, because they must stack so tall,” he said. The 2 bit per cell MirrorBit at 32 nm has been proven to work, he said, adding that, “We need some innovation at the 25 nm generation, but we can go there. At 18 nm, we still have a lot of work to do.”