Spintronic Memories to Revolutionize Data Storage

Electronic technology has evolved enormously over the past century, but in the most fundamental way it has not changed at all. From the earliest vacuum tube amplifiers to today's billion-transistor processors, all electronic devices work by moving electrical charges around. The countless discoveries and innovations that made the digital age what it is today were all made possible by our ever-improving mastery over electrons.

But those electrons are now beginning to rebel.

As we build transistors and other components with nanoscale dimensions, processors and memories are becoming so dense that even their infinitesimal individual currents are combining to produce scorching heat. Furthermore, quantum effects that were negligible before are now so pronounced that they're threatening to render circuits inoperable. The upshot is that we're fast approaching the point when moving charge is not going to be enough to keep Moore's Law chugging along.

Click on the image for the full illustration view.

In anticipation of that day, researchers all over the world are already working on a promising alternative. We have set our sights on a different property of electrons, which we hope to exploit for storing and processing data. This property is spin.

Spin is a fundamental yet elusive quantum attribute of electrons and other subatomic particles. It is often regarded as a bizarre form of nanoworld angular momentum, and it underlies permanent magnetism. What makes spin interesting for electronics is that it can assume one of two states relative to a magnetic field, typically referred to as up and down, and you can use these two states to represent the two values of binary logic—to store a bit, in other words.

The development of spin-based electronics, or spintronics, promises to open up remarkable possibilities. In principle, manipulating spin is faster and requires far less energy than pushing charge around, and it can take place at smaller scales. The holy grail in the field is a spin transistor. Chips built out of spin transistors would be faster and more powerful than traditional ones and, farther down the road, may feature such new and remarkable properties as the ability to change their logic functions on the fly.

We're still decades away from being able to build such a thing. But chips that exploit spin in a more modest way are already available. At least one company, Everspin Technologies, of Chandler, Ariz., is now selling magnetoresistive random access memory, or MRAM, a kind of spintronic memory. And many others—including Freescale, Honeywell, IBM, Infineon, Micron, and Toshiba, as well as start-ups and university research groups—are busy investigating MRAM technology.

The reason for all this interest is clear. Today's computers often use four kinds of storage. Dynamic random access memory, or DRAM, has high density but needs to be constantly refreshed and consumes lots of power. Static random access memory, or SRAM, used in caches, is fast to read and write but takes up considerable space on a chip. Flash, unlike SRAM and DRAM, is nonvolatile but is quite slow to write to. And then there are hard disk drives; these have high density but rely on moving parts, which impose size and speed limitations. MRAM is attractive because it could, in principle, replace all other kinds of memory.

Rather than representing a bit as charge in a capacitor or as the state of an interconnected set of transistors, MRAM stores data using the spin of electrons in a ferromagnetic substance, which is to say it stores data by creating a magnetic alignment in one direction or the other. In a tiny region of that material, spin up means 0, and spin down means 1. Proponents say that as MRAM improves, it could combine all the advantages of SRAM, DRAM, flash, and hard disks—with none of their shortcomings. It would be a compact, speedy, low-power, and nonvolatile "universal memory." With MRAM, a computer wouldn't have to juggle data between main memory, cache, and disk; instead, it could load all data into its working memory. This capability would make possible instant-on systems and maybe even change the way we think about computer architecture.

At the moment, however, MRAM suffers from two problems: The density of bits is low, and the cost of chips is high. The early MRAM designs needed lots of current to change a 1 to a 0 or vice versa. This requirement prevented their further miniaturization. Improved designs might overcome that hurdle using novel techniques and materials, but they would operate at only liquid-nitrogen temperatures. This is not going to work for your iPod.

This problem—the need for cryogenic temperatures to reduce the write current of MRAM—has been the focus of our work at North Carolina State University. It's a major challenge, but we've recently made a significant breakthrough: We demonstrated a device that shows potential as an MRAM memory cell. It can be written to using conventional voltage levels and almost no current at all. The key is a material called gallium manganese nitride, a semiconductor whose magnetic properties we can manipulate electrically. And here's the best part: It works at room temperature.

MEMORY VS. MEMORY:
Future MRAM chips could combine all the advantages of existing memories with none of their shortcomings

Illustrations: Emily Cooper
 

HARD DRIVE:
+ High density; very low cost per byte stored
– Moderate read and write speeds; bulky moving parts
 

SRAM:
+ Superfast read and write speeds; low power
– Large memory cells take up considerable space; volatile
 

DRAM:
+ High density; low cost; fast read and write speeds
– Volatile; constant refreshing of data drains power
 

FLASH:
+ Nonvolatile; high density; fast read speed
– Power consuming; write operation is slow and has limited endurance
 

MRAM:
+ Nonvolatile; high density; fast read and write speeds; low power; unlimited write endurance

Spintronic technology is already in your computer, at least in a primordial incarnation. Modern hard disk drives have a read head that relies on an effect known as giant magnetoresistance, or GMR, which was discovered by French and German researchers in the late 1980s. Basically, when the spins of electrons in the read head point in the same direction as those creating the small magnetic domains on the disk, the head's electrical resistance decreases. When the spins are in opposite directions, the resistance increases slightly. More recently, engineers have developed even better read heads that rely on tunnel magnetoresistance, a kind of enhanced GMR. It is this ability to sense very feeble magnetic fields that has allowed hard-disk makers to keep doubling the capacities of hard-disk drives on a schedule that's even outpaced Moore's Law.

Many advances in spintronics resulted from two big research programs that the U.S. Defense Advanced Research Projects Agency, or DARPA, funded in the 1990s. The first one produced the earliest MRAM prototypes. These devices used memory cells consisting of magnetic tunnel junctions: two layers of a ferromagnetic material like iron separated by an extremely thin, nonconductive barrier of magnesium oxide. When the spins of the electrons in the two ferromagnetic layers point in the same direction—in other words, when their magnetizations are aligned—the electrical resistance across the junction decreases; when the spins point in different directions, the junction becomes more resistant to current. The prototypes used this change in resistance to sense whether a 1 or a 0 was stored.

Some MRAM chips built at the time contained millions of memory cells, each with dimensions of about 150 nanometers, an impressive achievement back then. But the researchers soon discovered that going below 100 nm was not going to be easy. The problem had to do with the method they used to change bits, which was to drive currents through electrodes connected to each memory cell, creating a magnetic field that oriented the spin state of the cell. This method required currents that proved quite high, draining lots of power. Worse still, the magnetic fields affected not only the desired bit but also others nearby, resulting in errors.

Researchers are now trying to improve on this scheme. The most promising alternative is called spin-torque-transfer, or STT. The idea is to send electrons through a magnetic layer of cobalt, which tends to orient their spins in the same direction. The resulting spin-polarized current then flows into another layer of cobalt material. There, by virtue of one of the many mysteries of quantum mechanics, the incoming spin-polarized electrons transfer their spin orientation to the electrons on this second layer, thus magnetizing it.

So instead of writing a bit by applying a magnetic field, as early MRAM designs do, STT uses a spin-polarized current of electrons. To be commercially viable, the magnetic region where the bit is stored has to be quite small, of course. Researchers believe STT should work down to at least 65 nm and possibly even smaller dimensions. Last year, engineers at Hitachi and Tohoku University demonstrated a prototype capable of storing 32 megabits this way. But that's not all that much. For comparison, a modern DRAM chip can hold 128 times that amount. And though in theory such memories should require very small currents to change a bit, in practice the currents are still too high for most commercial applications.

For such reasons, our group and several others are betting on a different approach entirely. Forget about current-induced magnetic fields and spin-polarized currents. Instead, find a storage medium with a permanent magnetism that you can control by applying small voltages. These materials exist: They are called dilute magnetic semiconductors. As their name suggests, they are semiconductors that are also somewhat magnetic. Their magnetism stems from certain metal atoms added in a process similar to doping. What's interesting about these materials is that the presence of charge carriers—electrons and holes (vacancies left when electrons are missing in places where they'd normally be found)—can alter their magnetic properties.

As part of DARPA's second MRAM-research program, initiated in 1999, researchers investigated several dilute magnetic semiconductors, in particular gallium manganese arsenide and indium manganese arsenide. Both proved to be good candidates. There was just one problem: A material is magnetic only up to a given temperature—in this case about 200 kelvin, or –73 °C. That's colder than nighttime in most parts of Mars! Go above that level—known as the Curie temperature—and atomic vibrations cause the spins to lose the orderly arrangement that makes the material a permanent magnet. If this was a memory chip, you'd lose your data.

Our first breakthrough came in December 2001. At the time we were seeking a dilute magnetic semiconductor with a Curie temperature higher than room temperature. Following what were then just theoretical results, we decided to add some manganese to gallium nitride—about two to five manganese atoms for every 100 gallium atoms—to see what would happen.

The resulting gallium manganese nitride turned out to be very promising. When you apply a magnetic field to this substance, it becomes permanently magnetized. That is to say, when you remove the field, the magnetization doesn't go away, so it can be used to store data.

Our next major step, which we reported last year, was the ability to manipulate the magnetic properties of this semiconductor electrically. We started with ordinary gallium nitride. We then applied a thin layer of gallium nitride that contained a little added silicon, a dopant that donates electrons, thereby creating an n-type semiconductor. (The n stands for "negative," reflecting the addition of negative charges—electrons.) Next we added another gallium nitride layer, this time using magnesium as a dopant to remove electrons from the lattice of atoms, creating a p-type layer (p stands for positive). Finally, we deposited a very thin veneer of gallium manganese nitride on top of all this.

The junction between n- and p-type layers was key. That's because you can control the concentration of electrons and holes around a p-n junction by applying a voltage across it. And that's exactly what we did next. We connected electrodes to the n-type and p-type layers and applied a few volts. Then we turned our attention to the upper layer of gallium manganese nitride, using very sensitive instruments to measure extremely weak magnetic fields within it.

When we applied a voltage of –5 volts across the p-n junction, the magnetization of that upper layer approached 0. When we removed the voltage, the magnetization shot up. It was a faint magnetization to be sure, but enough for storing bits.

Now, you might ask, why does the voltage on a p-n junction change the magnetization nearby? To understand that, you have to first think about what goes on at a p-n junction when no voltage is applied across it (or break out the textbook you used in your introductory electrical engineering class in college).

First, recall that the n-type material has an abundance of negative charge carriers—electrons—which are free to move around. In the p-type material, the charge carriers are holes, spots in the atomic lattice that are lacking in electrons. When you put one of these materials against the other, electrons move from the n-type material into the p-type material, filling what were vacancies, or holes. So you end up depleting both types of charge carriers in the vicinity of the p-n junction, which is called, naturally enough, the depletion zone. This process is self-limiting, though. The loss of electrons from the n-type material leaves it with a positive charge, while the gain of electrons in the p-type material makes it negatively charged. This sets up an electric field that opposes the migration of any more electrons across the junction.

As with an ordinary diode, if the p-type material is made positive with respect to the n-type material, the applied voltage can overcome this electric field, sending holes and electrons racing toward the junction, reducing the thickness of the depletion zone. A voltage of the opposite sense boosts the internal electric field and makes the depletion zone wider.

What makes our device different is that the p-type material is very thin and is positioned right next to the magnetic layer of gallium manganese nitride. So by adjusting the voltage across the p-n junction, we can control the concentration of holes in the p-type layer at the interface with this magnetic material. That's important because the pervasive quantum-mechanical weirdness that arises at these scales allows these holes to interact with the manganese atoms sitting a few hundred angstroms away. Though there is a debate in our community, we believe the quantum phenomenon at work here is what is known as carrier-mediated ferromagnetism. It's as though the holes told some of the electrons around these manganese atoms to align their spins and start acting like a refrigerator magnet.

By the same token, when we apply a negative voltage across the p-n junction, we increase the width of the depletion zone enough to diminish the number of holes at the interface with the magnetic material. That then allows the spins of the electrons in these manganese atoms to revert to random directions. The device's magnetization vanishes.

This was the first demonstration that ferromagnetism can be controlled by applying voltages to a p-n junction without relying on ultracold temperatures. We hope this discovery will help turn spintronics into a hot topic again, so to speak.

The initial prototype we built can't be readily used as a memory cell. First, we need a major improvement on our design. The problem is that, although you can control the magnetization of our device using voltages, when you remove the voltages the magnetization returns to a baseline level. For a device to work as a memory, you need to be able to switch back and forth between two stable states.

One idea we're currently considering is making our device's layers even thinner and adding a barrier of nonmagnetic material, also very thin, between the p-type and magnetic layers. We're hoping that, by applying a voltage across these two layers, we can change the concentration of holes in the p-type region and also force some of the holes to cross the newly added barrier and migrate into the magnetic section of the device. The barrier would then play a key role: After the voltage is removed, it would prevent the holes from migrating back to the p-type region, thereby maintaining the magnetization of the device even when it's not powered on.

Now, if you take the device in this magnetized state and apply a voltage in the reverse direction, the holes would cross the barrier back into to the p-type region. The holes would remain trapped there, and the magnetization would disappear. This approach would provide the two stable states we need to use the device as a memory.

If this design is successful, the next step would be miniaturization. In fact, our initial prototype is rather big—each memory cell is about the size of a fingernail. To build smaller memory cells, we're investigating two approaches: One is using conventional photolithography, which we believe could lead to cells about 50 nm in size. Another idea is to grow the cell structures as nanowires, which we speculate might shrink them as small as 20 nm.

Such reduced dimensions would lead to another challenge: reading the bits in these tiny cells. As we proceed to nanoscale dimensions, the strengths of the magnetic fields will become even smaller. How to detect them remains an open question. We might have to equip each memory cell with a tiny magnetic sensor, similar to a read head of a hard drive but etched as a series of layers in the semiconductor. It's a possibility, but we don't know how it will perform and whether the resulting device would be economically viable.

Finally, another issue crucial to the commercial success of our MRAM proposal is its compatibility with conventional semiconductor technology. In theory, because MRAM would be programmed and interrogated electrically, it could be integrated with ordinary chipmaking processes. Then the MRAM devices could be made part of multifunctional integrated circuits, which would be able to perform all the processing, storage, and communication tasks that today require separate chips.

Clearly, overcoming these hurdles will take a lot of work. But if all goes well, our electrically controlled magnetic material may help engineers to ensure their continued mastery over electrons—and their spins.

This article originally appeared in print as "A Spin to Remember."