For computer manufacturers, Moore's Law has been about getting something for nothing, and not just in processor chips.
For 40 years, systems designers have seen memory densities double every 18 to 24 months while memory chip prices have remained essentially flat, cutting the cost per bit in half each time. As the technical challenges of building ever-smaller memory cells in silicon have increased, however, some memory manufacturers have predicted that the cost curve will start to swing the other way before the end of the decade.
Researchers are working on several alternate technologies that could eventually replace those in the three memory types commonly used today: low-cost dynamic RAM (DRAM), used in PCs and servers; fast static RAM (SRAM), used for processor caches and mobile devices; and nonvolatile flash memory, used in everything from computer BIOSs to cell phones.
Researchers at IBM, Intel Corp. and other companies envision the development of a universal memory technology that could someday replace all three. "That's the Holy Grail," says William Gallagher, senior manager of exploratory and nonvolatile memories at IBM.
A universal memory technology could change how computers are designed, Gallagher says. For example, nonvolatile RAM could allow computers to boot up and power down instantly because stored information wouldn't be lost when power was. But the emergence of a universal memory technology is probably at least 10 to 15 years away, Gallagher says.
Others argue that a universal memory isn't possible because one memory type can't satisfy all needs. For example, nothing could be the fastest and cheapest at the same time.
Most research today is focused on addressing the limitations of one memory technology at a time, such as flash. But some attributes of today's technologies will be hard to beat.
"Most technologies will probably not be able to compete on lowest cost per bit [against DRAM] or with the fastest SRAM. So they will fall into that space in between," says Bob Merritt, vice president of memory research at Semico Research Corp. in Phoenix.
Ferroelectric RAM (FRAM) and magnetoresistive RAM (MRAM) are the best-funded and most-evolved of the emerging memory technologies. FRAM is a nonvolatile RAM that was developed by Ramtron International Corp. in Colorado Springs. It's licensed by Texas Instruments Inc. and others. More than 30 million products have already shipped using FRAM, including metering, radio frequency identification and smart-card devices, according to Ramtron.
FRAM, which is based on nanoscale "quantum dots," uses less power and writes faster than DRAM or flash, and it has a long life span. But the technology remains 20 to 50 times more expensive per bit than DRAM, and chip density is far lower. Ramtron is prototyping 1Mbit parts today and hopes to push the technology to 4Mbit or 8Mbit in 2006. Until MRAM is ready for the market, however, FRAM is the only game in town for nonvolatile DRAM.
MRAM is closer to production than most other experimental memories, and IBM and Freescale Semiconductor Inc. are the leading MRAM developers. Gallagher calls the technology "magnetic storage on a chip" because it adapts the magnetic polarization techniques used in disk drives to silicon.
"It's fast, nonvolatile memory that offers a nice combination of high speed, high endurance and reasonable density," Gallagher says.
Chip samples can be produced at about the same density and cost per bit as flash. But while promoted as universal memory, the density of MRAM doesn't approach that of DRAM or SRAM. Most interest today is focused on embedded applications.
Several other challengers to conventional memory technologies look promising but are much earlier in their development cycles than FRAM or MRAM.
Phase-change memory (PCM) is a fast, nonvolatile memory that proponents claim could become a universal memory. IBM and Intel have each partnered with other companies to develop the technology.
A transistor in a PCM cell applies energy to either heat or cool the material, forcing it to change between an amorphous (high-resistance) and a crystalline (low-resistance) form. A current is then applied to measure the resistance and establish the state of the memory cell as a 0 or 1.
While PCM technology is much faster than flash, it's slower than SRAM. To be competitive with DRAM, it would also have to support unlimited writes. IBM's research shows that PCM can match flash's 100,000-write limit, but endurance beyond that hasn't been proved, says Gallagher. Ovonyx Inc. in Santa Clara, Calif., however, claims that its technology, called Ovonics, can be written to 10 trillion times, making it useful for DRAM as well as for flash.
Carbon nanotubes—hollow, tube-shaped lattices of carbon atoms—can be used to make a mechanical memory that works by bending the carbon filament up or down to make or break a connection between two electrodes.
"We can combine the nonvolatility of flash with the speed of SRAM and the density of DRAM," says Greg Schmergel, co-founder, president and CEO of carbon nanotube pioneer Nantero Inc. in Woburn, Mass.
The technology can theoretically scale as small as 5 nanometers and has achieved speeds a few times faster than today's DRAM chips. Nanotube chips could be 10 to 15 times smaller than today's DRAM and could offer a tenfold reduction in power consumption. Initial products from partners such as LSI Logic Corp. are still two to five years away, Schmergel says.
Molecular memory, developed by ZettaCore Inc., uses a chemical process to create DRAM memory cells with a molecular capacitor. The "chemically self-assembling" molecules work by adding and removing electrons. This changes the voltage, which is then measured to determine the state (0 or 1). The technique supports four states and can store 2 bits per memory cell.
Molecular memories also require 70% less power than a standard DRAM memory cell because the capacitor can hold 100 times the charge and therefore needs to refresh memory less frequently, says Subodh Toprani, CEO of ZettaCore.
Molecular memory will allow manufacturers to double or quadruple capacity without increasing costs, Toprani says. He says he hopes to have a product available by 2007 or 2008.
Axon Technologies Inc.'s Programmable Metallization Cell memory (PMCm) is a DRAM alternative that's nonvolatile, uses less power and offers higher density than DRAM. In it, tiny quantities of metal self-assemble into a filament as electrons are added to the metal ions. Resistance is then detected to determine the state of the memory cell.
"We don't store information as a charge; we store it as atoms," says Michael Kozicki, co-founder and chief technology officer at Scottsdale, Ariz.-based Axon.
Kozicki says he hopes to have the first "real design" available by 2007. "It looks like we won't cost any more than current DRAM," he says.
Ultimately, the ability to cost-effectively manufacture new memory technologies using existing fabrication facilities may separate the winners from the losers, says Semico's Merritt. To succeed, emerging technology vendors will initially focus on niche markets where they can coexist, rather than compete, with established vendors, and thereby continue to evolve.
"At the moment, all of those [technologies] look very powerful and doable," Merritt says. "But it's always the question of, Do they look so valid because they are, or because we haven't peeled enough layers of the onion yet?" The answers aren't likely to come before the end of the decade.
 |
|
|
|
|||||||||||||||||||||||||||||||||||||||||||
|