SAN JOSE, Calif. - Eyeing the growing market for cellular phones, DRAM vendors are seeking ways to fold more non-volatile memory and advanced packaging techniques into their game plans.
Though the use of DRAM in cell phones is still rare, some chip makers are betting that phone manufacturers are ready to switch from SRAM to lower-cost DRAM as they ship more handsets with multimedia capabilities and color screens, all of which require more memory.
"Phones are needing higher bit densities, and SRAMs are much more expensive per bit than DRAM. As the telephone requires more memory space it becomes cost-prohibitive to stay with SRAM technology," said Jim Sogas, vice president of sales at Elpida Memory (USA) Inc.
The move into cell phones began several years ago when DRAM makers started offering special low-voltage devices to reduce power consumption. Going forward, memory chip vendors say they will have to provide more than discrete DRAM devices.
Several top-tier DRAM vendors are putting into place plans to offer multichip packages that combine low-power DRAM or pseudostatic RAM with flash memory or even digital baseband processors. Some vendors are seeking to form alliances with companies that have more packaging or flash technology expertise. Other vendors are cultivating such capabilities in-house.
Hynix Semiconductor Inc., for one, is working on ways to stack two die in the same package to keep component size to a minimum for cell phones. For the low end, the company will provide either SRAM or pseudostatic RAM plus flash memory. At the high end, it will offer low-power SDRAM plus flash. The products are expected to ship in the second half of 2003, said Farhad Tabrizi, vice president of worldwide marketing at Hynix.
Such a move would directly challenge leading memory vendor Samsung Electronics Co. Ltd., which got the ball rolling earlier this year when it announced it was shipping a multichip device that stacked 128 Mbits of SDRAM with 256 Mbits of NAND flash or 32 Mbits of its UtRAM.
Second-ranking memory chip maker Micron Technology Inc. is also drawing up a plan to stack multiple memory chips inside the same package, a spokeswoman said. She declined to provide details.
In Hynix's case, the company will need to take two crucial steps before it can compete. One is to develop a package in which the I/O bond pads are at the edge of the DRAM, rather than at the center, which will allow connections between the two stacked die.
These edge-bonded DRAMs were common until the 16-Mbit DRAM generation, when chip makers switched to bonding the die across the center as a way to reduce die size and improve electrical characteristics. But that method will no longer do when stacking die, Tabrizi said.
Hynix will start offering these "semi-custom" packages by the second half of the year. "I estimate that 80 percent of the mobile market will use it by 2005 or 2006," Tabrizi said.
Hynix also may pair low-power DRAM or pseudostatic RAM with NAND flash technology, known for its high bit density. Today, Hynix provides NOR flash devices, but not NAND-type flash memory. Samsung, SanDisk Corp. and Toshiba Corp. are currently the largest providers of NAND flash chips.
Tabrizi, however, did not rule out the possibility that Hynix would enter the NAND market, perhaps with the help of a partner. "We are looking for ways to provide complete solutions to customers," he said.
Elpida too is working to fill gaps in its product line as it seeks a greater presence in cell phones. Like most DRAM companies, Elpida has offered low-power SDRAM to cell phone customers as an alternative to SRAM. Beyond that, the Japanese chip vendor is seeking ways to stack flash memory and baseband processor chips with its mobile DRAM.
Sogas said Elpida is negotiating separately with undisclosed chip vendors specializing in flash and baseband processors, neither of which Elpida makes itself. "We're in heavy discussions now. It could be happening by the middle of ," he said.
Like Hynix and Samsung, Elpida is leaning toward flash memory geared for data storage technology in combination with its low-power DRAM. One option is to use AND-type flash memory technology developed by one of Elpida's parents, Hitachi Ltd., though Sogas said no decision has been made. "We won't make a decision based on who our parents are; we'll make a decision on what will be most successful," he said.
Grounds for appeal
NAND-type flash appeals to DRAM vendors for manufacturing and performance reasons. "NAND flash is a lot lower-cost to manufacture than NOR and has better electrical characteristics in terms of writing. You can write in blocks of data; it's not bit by bit," said Hynix's Tabrizi.
As for packaging technology, Elpida is looking at ways to combine different die either laterally or vertically. There's also a chance Elpida could work with other companies specializing in multi-chip packages, Sogas said.
When it comes to combining different memory types, memory chip makers typically have soldered one package on top of another, but ideally they would like to find a way to stack bare die within one package to minimize height.
Today, for example, Hynix can stack two separate chip-scale packages - one with NOR-type flash, the second with pseudostatic RAM - and keep the height to 1.2 millimeters, the same as a standard thin small-outline package. Stacking the bare die within the same package, however, would yield a device that is "smaller than TSOP," Tabrizi said.
DRAM makers are moving headlong into cell phones for the same reason they are taking a keen interest in graphics subsystems and networking. The devices are specialized, which tends to boost profit margins. They also give memory chip vendors an opportunity to work closely with a few important customers in defining specifications in the early phases of design, giving those customers an edge when volume shipments begin.
Another reason is that memory chip companies are seeing the proportion of bits consumed by desktop PCs - still the largest market for DRAMs - shrink as profit margins dwindle. By 2006, the amount of DRAM bits consumed by desktop PCs will fall to 35 percent of the total bits produced. That will be in sharp contrast to five years ago, when 70 percent of the bits produced found their way into desktop systems, Tabrizi said.