A year ago, it was believed that flash would hit a wall at the 65nm node, but that has now shifted downward.
"Scaling is getting to be difficult, but it hasn't yet stopped," said Stefan Lai, vice president of the technology and manufacturing group at Intel Corp., Santa Clara, Calif. "We have a very good feeling [about current flash technology] until the 45nm node, which is in 2008."
The prospect that flash memory's technological limitations will soon be reached has caused a step-up in research activity at major flash vendors, which are searching for replacements.
Stakes are high. With the global market for flash expected to jump from $11.2 billion in 2003 to $43 billion in 2007, according to Semico Research Corp., companies are already placing their bets.
They differ widely in their choice of the most commercially viable flash replacement. Each hopes to find what Intel's Lai refers to as the holy grail of memory--cheap like DRAM, fast like SRAM, and nonvolatile like flash.
Intel, which derives about $2 billion, or roughly 7%, of its revenue from flash, has closely examined 30 possible replacement technologies for high-volume nonvolatile memory (NVM) applications such as cell phones, Lai said.
Of those, Intel believes ovonics unified memory (OUM), which uses the same materials as rewritable CDs and DVDs, is the most promising.
OUM-based NVM can read and write in hundreds of nanoseconds, which also makes it attractive for embedded systems. "The memory is fast enough to talk to the CPU directly," Lai said.
OUM requires the same chipmaking equipment as silicon, just a different recipe. To avoid contamination with silicon, the ovonic process runs on a separate line from silicon and the same area where Intel today runs copper technology.
The company is working on a 4Mbit test chip based on OUM, although it has no timeline for implementation. "We know we can make a memory out of it," Lai said. "The problem we have is how to lower the cost."
OUM is also supported by ST-Microelectronics N.V., Geneva. "We see OUM as the only potential candidate to address the broad market addressed by flash memory," said Philippe Berge, director of marketing for ST's memory products group.
Berge said ST's evaluation of OUM technology shows that its mask steps are comparable in number with those for NOR flash. Scalability allows OUM to go down to small geometries, and material used for it is more user friendly than that of competing technologies.
The challenge is integration into the CMOS process. Although the OUM modules are relatively CMOS friendly, "in MOS history, the introduction of new materials has always required significant effort and time," Berge said.
Motorola Inc. finds magnetoresistive RAM (MRAM) technology more advantageous. Saied Tehrani, director of MRAM technology at Motorola Labs in Chandler, Ariz., said his company never pursued OUM.
"OUM has to go to very high temperatures from one phase to another, creating reliability issues," Tehrani said. "None of that has been resolved."
In October, Motorola produced sample 4Mbit MRAM-based chips for embedded applications. MRAM provides read/write speeds below 10ns, far faster than OUM, unlimited endurance on programming (unlike flash, which has 100,000 to 1 million cycles), and better scaling than other memory technologies, Tehrani asserted.
Motorola plans to increase volume production of MRAM, with an initial focus on embedded use. "The MRAM can replace SRAM and flash in a SoC," Tehrani said, adding that aside from a few tool changes, MRAM can use existing fab equipment and is being produced at the company's MOS 12 facility in Chandler, Ariz.
Some votes for FRAM
Another potential replacement technology for flash, ferroelectric RAM (FRAM), has been around for a long time but is now being commercialized by Texas Instruments Inc. TI last year produced a 64Mbit FRAM within a standard CMOS logic process at 130nm.
FRAM technology involves using an electric field to shift the position of atoms in ferroelectric crystals. Very low power is a key advantage.
Ted Moise, FRAM integration manager at TI in Dallas, said that FRAM is formed with the addition of only two masks, and the FRAM module does not disturb the logic. Moreover, only a few new tools related to depositing and etching materials are needed for making FRAM modules.
While Motorola is targeting the embedded-NVM market with MRAM, FRAM is being evaluated as a potential replacement for commodity flash. For example, Ramtron International Corp., which has worked with TI on FRAM development, will focus on stand-alone memory products.
Moise said TI views FRAM as playing in several areas, including slow SRAM cache, embedded DRAM, and some applications where flash currently holds sway. The company expects its FRAM-based devices to become prevalent in a wide range of applications like wireless and broadband by 2005.
Lai agreed that FRAM and MRAM are viable technologies, but fit into different segments than the low-cost, low-power NVM that Intel is focusing on.
MRAM needs large amounts of power to switch magnetic material, and needs a material that is not really compatible with the silicon process, he said.
As for FRAM, its cell size is large compared to DRAM and flash, and it has destructive read. "The number of cycles is limited because every time you read from the cell, you also write to the cell," Lai said.
Still up in the air
Whether any of these technologies will become the future commodity NVM remains to be seen. Jim Handy, director of nonvolatile memory services at Semico in Phoenix, noted the fast read/write and low-power-consumption advantages of MRAM and FRAM, but stressed that cost remains a major stumbling block.
"If they could make [FRAM or MRAM] with a cost something like that of DRAM, they could wipe out the market for all other memories," Handy said. "But the major FRAM proponent, Ramtron, has said for 19 years that the emergence of FRAM is just around the corner."
Most companies are prudently putting their R&D dollars into more than one potential flash replacement.
Intel, for example, is working with Swedish company Thin Film Electronics ASA to develop polymer memory. Polymer is doped with a special composition so that when an electrical field is applied, it demonstrates ferroelectric properties. And polymers combine high storage capacity and low power consumption.
The real allure of polymers is low cost, according to Intel's Lai, who estimated they could come in at less than 5% of silicon-based nonvolatile memories.
Flash still has legs
Samsung Semiconductor Inc. is evaluating all memory technologies considered a possible successor to NOR and NAND flash, said Steffen Hellmold, director of flash marketing at Samsung in San Jose. But despite research progress, Hellmold said the new technologies are far from commercial viability, and in fact are not urgently needed.
Existing flash technology has at least five more years of life, he said.
"With NAND flash, we have the strongest growth ahead of us over the next two or three years," Hellmold said. "This year we've seen 150% flash bit growth and next year we're expecting a similar number."
Samsung, he added, possesses a firm technology roadmap to 45nm flash over the remainder of the decade using existing technology.
Along the same lines, some companies are addressing the scaling issues of flash technology in an effort to extend its life. Motorola is developing a thin-film storage technique called nanocrystal flash. The company views nanocrystals as successors to floating-gate-based flash, which many believe will not continue to scale to smaller geometries.
Silicon Storage Technology Inc., Sunnyvale, Calif., believes its current SuperFlash technology is viable for perhaps 15 more years, said senior vice president Mike Briner. The main way to erase flash is through tunneling, and SST's flash technology tunnels through a thicker oxide than rival flash memories, making it more scalable, he said.
SuperFlash technology is NOR-based, but SST thinks it can be used with a NAND interface to replace high-volume NAND. "We're transitioning manufacturing to second-generation [SuperFlash] technology, which is a self-aligned cell that will allow us to become competitive at the densities cell phones are using," Briner said.
References to technology walls remind Samsung's Hellmold of past proclamations that 100nm can't be mastered with flash technology. Samsung, he noted, has already announced NAND flash at 70nm and 40Gbytes for the first quarter of next year.
Intel's Lai said the problem with forecasting a NVM technology wall lies in the lack of an international technology roadmap like the one projecting limitations for logic, and which energizes the industry to work on solutions. "No one in the industry is saying, 'Let's work together [on NVM],' " he said. "We're all very competitive and tend to keep information secret."
NAND and NOR have 10 times more R&D dollars overall going into them than does OUM technology, Lai said. "It's very difficult to overcome an incumbent technology."
The wall pops up when NVM companies don't invest enough in overcoming the technical hurdles. "The moment we put our attention to it, that wall tends to come down," he said.