The Persistence of Memory: Looking into MRAM

Many in the electronics industry want to change how memories are made. So PML scientists are exploring the physics and developing the metrology required to do just that -- including an evaluative scan technique that is an order of magnitude faster than any comparably accurate method.

Industry planners are concerned that the ubiquitous dynamic random access memory (DRAM) chip, a key component of computers for more than three decades, may not be adequate to the demands of future computing, which include increased speed and reduced power consumption. DRAM chips store information in the form of electrical charges in individual capacitors, one for each bit. Because the charge dissipates rapidly, the system has to be refreshed with new current many times per second, consuming power and generating heat. Moreover, DRAM is "volatile" – that is, it disappears when the current is turned off – and cannot offer "instant-on" memory of its prior configuration. 

Not surprisingly, DRAM alternatives are the subject of worldwide research efforts. One of the most promising technologies is magnetic random-access memory (MRAM), which stores information as magnetic states rather than charges and is inherently non-volatile. One very recently developed version, called spin transfer torque MRAM (STT-MRAM), requires only about 1/50th as much electrical power as DRAM. STT-MRAM relies on transferring the spin angular momentum of electrons to switch the relative magnetic alignment of two adjacent ferromagnetic layers, each about 2 nm thick and tens of nanometers wide. The process has been measured to take as little as 2 ns (so far) and the device could in theory execute an infinite number of cycles without wearing out.

"A lot of companies are trying to make prototypes," says Ron Goldfarb, leader of the Magnetics Group in PML's Electromagnetics Division, "but often they lack adequate measurement techniques and can't be certain their devices are switching reliably. The underlying physics of the process is still incompletely understood."

That's where the Magnetics Group's Bill Rippard and colleagues Eric Evarts and Matt Pufall come in. "What we're doing is looking at new ways to manipulate magnetism in these nanoscale devices and measure the results," Rippard says. "Up until about 10 years ago, the only way you could control magnetism in such structures was by using applied electromagnetic effects. But then researchers came up with an entirely new method: injecting electron spins into a ferromagnet. Our goal is to understand spin-transfer switching effects at a fundamental level, and help companies exploit what we learn to develop new products."

A typical MRAM unit used to store a single bit of information is a sandwich containing two cobalt-iron ferromagnetic layers ("fixed" and "free" respectively) separated by a 1 nm magnesium oxide non-magnetic layer, which both acts as a tunnel barrier and serves to break the exchange coupling between the ferromagnetic layers so that they can be independently switched.

The fixed layer has a strong permanent magnetic orientation. So when a direct current is routed into the fixed layer, only those electrons with spins aligned to the fixed-layer field can pass through it to the barrier layer and then into the weaker free layer.

Those "polarized" electrons exert a torque on the magnetic orientation of the free layer. If the torque is large enough, it can align the free layer so that it is parallel with the fixed layer. That is the informational equivalent of writing a "1." Conversely, if the current is passed through the structure in the opposite direction the layers are forced into an antiparallel state, writing a "0." The contents of the bit can be read by measuring the resistance across the MRAM sandwich unit.¹ Low resistance signals a 1, higher resistance a 0.

That's the physics. But in real materials, how reliable is the switching process?

"Companies that intend to fab a billion of these things on a chip will tend to follow whatever path will take them to lower error rates – without necessarily knowing the underlying physics or the specific physical problems associated with defects," Rippard says. "So one of our major goals is to fill that in, look at the physics that's going on, and find an informed way for fabricators to engineer the problems out."

Recently the PML team found that about 10 percent of the devices in an average run are prone to write errors for reasons that are still unclear. If MRAM chips are to be viable for mass production, fabricators will need to know how to minimize the number of anomalous devices. Unfortunately, the conventional test method is to switch the devices through hundreds of thousands of cycles of setting, reading, resetting, re-reading, and so forth, and then plotting the data. "Measuring a million operations that way takes around six hours," says Rippard's co-worker Eric Evarts. "But we have devised a method that takes a matter of minutes."

The method employs a variation on ferromagnetic resonance (FMR) spectroscopy, which makes use of the fact that the magnetization of a ferromagnetic material placed in an applied magnetic field will precess at a larger amplitude when the perturbation is applied at its resonance frequency, which depends on its orientation relative to the applied field, field strength, and other factors. Determining that frequency at different electron currents reveals much about the MRAM's behavior, both in single devices and in continuous-film arrays.

The researchers, utilizing a fast field-swept, spin-torque FMR technique, found that devices prone to write errors could be readily identified by the telltale presence of two resonant frequency peaks instead of one, apparently indicative of multiple modes in the free layer. Detecting that signature alone, they determined, is as effective as much more time-consuming measurement methods in identifying anomalous devices.

The specific cause or causes of the write errors are unknown, but are most probably due to thermal perturbations and their interactions with impurities, and/or non-uniformity in fabrication. Continuing research may narrow that list. But new effects could emerge as device dimensions shrink. "At some point," Rippard says, "these things are going to be 10 nm wide, but manufacturers will still need them to behave exactly the way they want."

Meanwhile, Evarts says, techniques such as the FMR analysis are helping to "tie these measurements back to things almost anyone in the field can do. We want to try to find comparatively easy measurements that can give nearly the same results as really complicated measurements suited to a PML lab."

---

¹ The discovery of this "giant magnetoresistance" phenomenon -- in which the resistance between two thin films varies substantially depending on the relative alignment of their magnetizations -- earned the 2007 Nobel Prize in Physics for Albert Fert and Peter Grünberg.