10 September 2010—A device that was until recently dismissed as a laboratory novelty will be in commercial memory chips within three years, says South Korean memory maker Hynix Semiconductor. The technology at the heart of the new chip will be Hewlett-Packard’s memristor, a device that its proponents promise will revolutionize not only chip memory but also logic—and perhaps even artificial intelligence.
Last week, Palo Alto, Calif.–based HP announced a partnership with Hynix that aims to produce a cheaper, faster ultradense memory. Hynix says it is looking to HP’s architecture to form the basis of its next-generation memory strategy.
"This is huge progress," says Sung-Mo (Steve) Kang, the chancellor of the University of California, Merced, who has studied memristive behavior for the past 30 years. He adds that the HP-Hynix partnership is concrete evidence that developing memristor technology is a realistic pursuit.
Hynix is the world’s second-largest manufacturer of the type of dynamic random access memory (DRAM) found in computer main memory, trailing only behind fellow South Korean chipmaker Samsung. The company also has a thriving flash memory business for products such as digital camera cards, smartphones, and the chips in Apple’s iPod and iPhone. But Hynix says it is banking on a different kind of memory for the future: With the HP partnership, the company seeks to eventually replace both flash and DRAM with the new device. "We think that the most likely first product will be a competitor to flash memory," says HP senior fellow Stan Williams. Within a few years, however, Williams says he hopes to see the HP-Hynix components "compete successfully with [both] flash and DRAM and be the basis for solid-state drives."
Such a development would be unprecedented. So far, no single component has been able to replace both DRAM and flash. DRAM is cheap but volatile, meaning it loses data when the power is turned off. Nonvolatile flash memory, by contrast, doesn’t need power to retain its data but is too expensive for the volume necessary for main hard disks.
In both kinds of memory, data is stored as charge. The problem is that the envelope for the electrons that constitute a discrete charge can’t be made much smaller. That makes it difficult to squeeze more memory into the same size footprint for the same cost.
Researchers have worked for years to come up with an all-purpose memory that would have the strengths of both flash and DRAM but none of their weaknesses. Such a wish-list memory technology would be dense, scalable, rugged, nonvolatile, fast, and cheap. Many technologies have been proposed for this all-purpose memory, but most are just in the earliest research stages.
The device Hynix seeks to build with the memristor is called resistive random access memory, or RRAM. Instead of storing data as charge, RRAM stores it as resistance. Unlike DRAM—which leaks electrons and must be periodically infused with fresh charge—memristor-based RRAM is nonvolatile, and that translates into significant power savings. In addition, memristors can be created in ultradense configurations that would allow continued scaling. Therefore, the memristor’s architecture is an ideal foundation for RRAM.
The first physical implementation of a memristive device was demonstrated at HP in 2008. First dreamed up as a mathematical proof by professor Leon Chua of the University of California, Berkeley, the memristor is known as the fourth fundamental circuit element. It joins the nuclear family of resistor, capacitor, and inductor by completing the relationship matrix among charge, flux, voltage, and current. None of the other circuit design elements had ever provided a link between flux and charge. Physically, this link enables a device that acts like a resistor; its value can vary according to the current passing through it, and it can remember that value after the current stops. Those properties make it an attractive nonvolatile memory candidate.
The architecture of HP’s memristor circuit is a hash of perpendicular wires known as a crossbar array, in which the memristors are sandwiched between the crossing points. In HP’s memristor, the interstitial material is titanium dioxide. An RRAM cell formed by a similarly constructed sandwich can store a bit when an applied voltage makes the normally resistive material conductive. Hynix will not discuss the planned architecture of the new devices, but Williams points to a publication last December in which his HP team described a hybrid CMOS-memristor architecture that had a small number of CMOS transistors controlling a 3-D system made of stacked crossbar memristor arrays. Williams speculates that Hynix’s planned architecture could be "very similar."
By 2013, Williams says, "our goal is to have a product that will have at least twice as many bytes as a flash memory with the same chip size, since [both flash and the Hynix-HP chip] will be built using the same technology." Soon after that, Williams says, the partnership will put DRAM in its crosshairs. "Right now, about half of the total power used by DRAM is static power," he explains. "A memristor competitor to DRAM would not dissipate static power just to keep the bits refreshed. So we should have a factor of two total power advantage."
Hynix spokeswoman Seong-Ae Park confirms that the memristor is the backbone of its next-generation RRAM memory chip. The license is nonexclusive, and HP plans to release its own chips at some point. "[Hynix] will be free to sell any memory product that they want," says Williams. But Hynix is not putting all its eggs in HP’s basket. The DRAM firm is also working on other options, such as spin torque transfer RAM (which operates using electron spin measurement), phase-change memory (which stores data by alternating two material states), and Z-RAM, a capacitorless version of DRAM that the company licensed from Innovative Silicon in 2007.
In fact, some researchers say that eventually memristors won’t just revolutionize memory but also logic. "HP has demonstrated a stateful logic," says UC Merced’s Kang. "Many conference papers highlight unique circuit applications, including programmable voltage amplifiers [and] nonvolatile CMOS latches," he adds.
Indeed, the United States Department of Defense is trying to capitalize on this potential by funding HP to build a memristor-based chip that can reason and learn the way a mammalian brain can. The ability to remember the amount of current that has flowed through them makes memristors good approximations for the synapses in a brain.
Ennio Mingolla, a researcher under HP’s Defense Department grant at Boston University, says that the HP-Hynix partnership addresses the "dumbest" possible use of memristors—in the sense that they function as passive memory only, for data storage. But, he says, "any investment that accelerates the ’dumb’ use of memristors can only speed their use in smarter hybrid memristor-and-CMOS chips."
A correction to this article was made on 13 September 2010.