Building Ultra-Energy-Efficient Computers Out of Tiny Bar Magnets
Nanomagnet computers could consume one-tenth the power of today’s microprocessors
Illustration: Emily Cooper
When it comes to computing, we can very easily divide the technologies we use into two distinct categories: speedy electronics and stable magnetics. Electrons, which move fast and interact strongly with one another, are ideal for performing computation. Magnets, by contrast, aren’t known for their speed, but they’re hard to perturb, making them the perfect medium for data storage.
But this division may soon disappear. Thanks to modern fabrication technology, we can now create nanometer-scale magnetic devices that can perform computations. These devices are not as speedy as state-of-the-art transistors, but they require far less energy to switch.
We’ll need devices like these because modern chips are consuming too much power. Today, the amount of electricity needed simply to maintain data in a circuit—called standby power—is fast approaching the amount that’s consumed when an actual computation is performed. Magnet-based devices, which require no power to save their state, could drastically cut down on this constant power drain, which is one of the main obstacles to continued progress in chips.
There are many ways to make logic devices using magnets. Our group at the University of Notre Dame and several others are working on one of the most straightforward approaches: building logic gates and wires out of small patches of magnetic material. These “nanomagnets” act just like tiny bar magnets. In circuits made with them, information isn’t carried from one place to another electronically. Instead it’s transmitted directly through the magnetic attractions and repulsions, flipping the polarity of north and south poles as it moves from one magnet to the next, in much the same way that a NOT gate reverses the logic state of a bit.
We have already demonstrated that we can build simple circuits, such as adders, with these nanomagnets. Now, with a new fabrication technique, we’re starting to contemplate what can be done to build fully integrated logic chips. Though relatively modest, these achievements make it clear that this technology could someday be used to make ultralow-power chips. In some cases we expect that nanomagnet circuits could use a tenth or maybe even a hundredth of the power with no sacrifice in performance. Such capabilities could be ideal for sensors and display electronics, hardware accelerators on multicore chips, and computationally demanding applications such as machine-vision systems in autonomous vehicles.
The idea of computing with magnetism isn’t new. Some early computers actually contained iron-based cores, which were used both to store information and perform computations. But that magnetic logic technology, which was based on stringing wires between ring-shaped magnets, was too bulky to compete with increasingly compact semiconductor technology.
Today’s magnetic logic is fabricated at much smaller scales, and the semiconductor industry, which funds university-based research through the Semiconductor Research Corp., has been supporting a variety of approaches.
Many of those efforts are part of a field called spintronics, in which researchers exploit the spin of electrons, rather than their charge, to transmit and manipulate information. Spin is a quantum-mechanical property associated with magnetism. One such approach, pursued by a group led by Kang Wang at the University of California, Los Angeles, transmits information in the form of waves of spins propagating through a magnetic layer of material (see “The Computer Chip That Never Forgets,” IEEE Spectrum, July 2015). Another scheme, being developed by Supriyo Datta and colleagues at Purdue University, uses similar pulse-carrying wires to connect magnetic dots.
Our strategy, which we expect to be more energy efficient than either of the above, is based on a simple premise: Make very small magnetic regions and allow them to interact with one another just as bar magnets would. These interactions can be harnessed to make wires that transmit information and to construct logic gates that can perform computations.
The key to this approach is miniaturization. Take a look at an ordinary piece of magnetic material under the microscope and you’ll find that it is naturally divided into a number of tiny patches. Within each of these patches, called domains, the spins of the electrons all point in the same direction. That alignment is what causes magnetization. However, a bulk sample of the material will not be magnetic if all of these domains are aligned at random—that is, if the north poles of the domains point in different directions. A patchwork of randomly oriented domains helps minimize the amount of energy that’s locked up in the magnetization of the material as a whole. But a lot of energy is required to pull all the domains into alignment to create a permanent magnet, and even more energy is needed to change the orientation of the poles of that magnet.
Of Magnets and Wires
The magnetic field of an elliptical nanomagnet naturally prefers to align along the longer, or “easy,” axis (top illustrations, left). Wires can be constructed in two ways. One is by arranging the magnets side by side (center), where their magnetizations prefer to alternate in direction. The other is end to end (right), where all magnetizations prefer to point in the same direction. All arrows point north.Images: Edit Varga/University of Notre Dame (5)
Make a magnet smaller than the material’s domain size—typically around 100 nanometers across—and you can make a tiny version of a permanent magnet. As with a bar magnet or a compass needle, the magnetic field of a nanomagnet naturally prefers to align itself on the longer axis, with a north pole on one end and a south pole on the other. But by fine-tuning the aspect ratio—the ratio of height to width—we can make it so that it requires relatively little energy to flip the direction of the nanomagnet’s north and south poles. In fact, if you make a nanomagnet too round—or too small, for that matter—little kicks from the thermal energy in the material can cause that flip to happen on its own at random.
Today we use similar tiny, flippable nanomagnets to store information in ordinary hard disk drives. They can also be found in magnetic RAM technologies now under development. In both of these cases, we don’t want the magnetic bits to interact with one another; such interactions would risk corrupting the data that you’re trying to store. But when building nanomagnetic logic, those interactions are exactly what you need in order to connect the devices and make them perform computations.
Our work on nanomagnet logic grew out of some previous research we’d done in the 1990s on quantum dots. Our original idea was to try to make computers by using arrays of these dots, which are patches of semiconductor in which electrons are confined to a space so small that they exhibit the same quantum behavior that they would in an atom. Rather than using wires, neighboring dots would transmit information by the Coulomb forces, which attract opposite charges and repel like charges. We showed that these physical interactions could be used in an appropriately structured array to perform logic operations, but we ran into technical limitations, in part because it was hard to control size variations when constructing quantum dots.
We realized early on that using magnetic dots—which are much more stable, easier to fabricate, and capable of operating at room temperature—might be a good alternative. And in the early 2000s, following some initial work done by Russell Cowburn, now at the University of Cambridge, in England, we began simulating and then experimenting with the devices.
We started out by making magnetic dots from a mix of nickel and iron. We gave them an elliptical shape, stretching them out in one direction to give each magnet a preferred north-south axis. But we limited the elongation, thus minimizing the energy required to flip the orientation to represent the 0s and 1s in binary logic.
Like a bar magnet, each nanomagnet has “fringing fields”—magnetic fields that extend a long distance from the magnet. If we place the nanomagnets close enough, we can use the fields from one magnet to influence the state of the ones near it, and so build up logic circuits that can pass magnetic information from one magnet to the next.
To get a sense of how such nanomagnetic logic would work, it’s helpful to remember the most basic rules of magnetism: Like poles repel and unlike poles attract. So, just as bar magnets would, two nanomagnets placed end to end will naturally prefer to have their north poles point in the same direction, so the north pole of one is closest to the south pole of the other. If two nanomagnets are placed side by side, so their long axes are parallel, their north poles will point in opposite directions, so that each pole is closest to one with the opposite polarity. This second configuration, which is called antiferromagnetic coupling, can actually be considered an elementary circuit element: the inverter. One nanomagnet would be the input and another the output, flipping the state of the information.
Voting With Magnets: Four or five nanomagnets can be arranged to form a key circuit component: the majority logic gate. This gate outputs the state that the majority of the inputs are in—or in this case, the inverse of that state. Three 1s, for example, will yield a 0. Note that this gate is built from nanomagnets with an out-of-plane magnetization—more on this design in the final section of the article.Images: Stephan Breitkreutz-v. Gamm/Technical University of Munich (5)
The inverter is not a bad start, but to perform useful logic operations you also need AND and OR gates. And in 2006, we reported a proof-of-concept demonstration of a second logic element called a three-input majority-logic gate, which can be used to build both of those other gates. The majority gate takes on the state of the majority of its inputs. One way to build it is with a cross-shaped arrangement of five nanomagnets—one center dot surrounded by four others. Three of those dots act as inputs, and the center dot “calculates” the majority by naturally aligning its magnetization with the majority of the other spins. The fifth dot carries the result out of the device.
As for the wires that you’d need to connect logic gates, these can be constructed by simply lining up nanomagnets either end to end or side by side. Information can travel in either of these configurations, flipping spins as it goes—a little like falling dominoes. (In the side-by-side configuration, of course, you need to make sure you have an odd number of nanomagnets, so the information retains the same state at the end of the wire that it did at the beginning.)
The fundamentals are quite straightforward. But you might have noticed that there’s a catch. Unlike a line of dominoes, which all fall in the same direction after the first domino is tipped, there is nothing intrinsic in a line of nanomagnets that determines which way the information will flow. A nanomagnet in the middle of a line of nanomagnets will be equally affected by the nanomagnet to its left and the one to its right. If we can’t reliably control the direction of computation, our computer will have a high error rate. To make up for this intrinsic symmetry in computational direction, we need to add some clocking circuitry—more on that in a moment.
Nanomagnets have a few nice properties. They’re inherently insensitive to radiation, they can switch pretty much indefinitely without degradation, and they’re nonvolatile, requiring no energy to retain data when they’re not switching.
At the same time, they’re very slow by modern transistor standards, maxing out at about one-hundredth the speed of a traditional transistor. That means nanomagnet logic will likely never reach gigahertz speeds. But the potential energy savings still make it an attractive alternative for the many applications that don’t require such speeds.
Much of the energy advantage comes in at the circuit level. Because of the way nanomagnets interact to perform logic operations, it can take as few as five magnets to add two 1-bit numbers together. For comparison, it can take 20 to 30 transistors to construct a similar adder in silicon.
In 2011, we demonstrated that we could combine majority logic gates, inverters, and nanomagnet-based wires to create the first full nanomagnet-based circuit: a functional 1-bit full adder. And thanks in part to work done in collaboration with IBM and with funding from the U.S. Defense Advanced Research Projects Agency, we’re confident that we can use magnetic RAM technology to connect nanomagnet circuits to the outside world. That’s because these nanomagnets, other than being shaped differently, aren’t much different from the magnetic bits used on such memory chips, which have already found their way into commercial production.
To start a chain of spin flips, first use a strong magnetic field to flip magnet 1. Magnet 2, originally in a down configuration, must flip up to propagate the information to the right. The magnetic field of magnet 1 isn’t strong enough to do this on its own. So, the magnetic field from a wire, or clock, is first used to draw magnet 2’s magnetization into the metastable “hard” axis. Magnet 2 can now relax into a new stable state. But it’s not clear which way it will fall; the field from magnet 1 would have it point up, and the field from magnet 3 would have it return to a down state. To make sure that magnet 2 feels a stronger influence from magnet 1, magnet 3 can be clocked as well. When the clock signal is now removed from magnet 2, it will relax into an up state, advancing information to the right. Magnet 4 can now be clocked to keep the flow of information going. Controlling every nanomagnet would require clocking circuitry for each device, which would be a miniaturization challenge. In practice, a line of nanomagnets might be clocked in groups. A wire, for example, might control half a dozen nanomagnets at once. The result will be a little more error prone but easier to manufacture.Illustrations: Erik Vrielink
But even early on, we realized that any nanomagnet computer we build will be only as good as its clock. As we mentioned before, we need clocking circuitry to keep computations moving in the right direction. We also need a clock to get the nanomagnets to switch reliably. Because magnets are so stable—the very reason they’re commonly used for data storage—switching them is often the tricky part. On its own, the fringing field surrounding one dot isn’t strong enough to reliably induce a 180-degree switch in a neighboring magnet.
A clock signal can help a nanomagnet switch. It could be created with something as simple as a nearby wire, which would generate an extra magnetic field near a nanomagnet when it’s carrying a current. To see how such a clock would work, imagine an elliptical nanomagnet. As noted before, its natural state is to have its magnetization along its longer axis. If we use the added wire to apply an additional magnetic field, called a switching field, we can rotate the nanomagnet’s magnetization by 90 degrees into what’s called the “hard” axis, the shorter axis of the two. This is an unstable state for the nanomagnet, and when the switching field is removed, the magnetization will begin to snap back into either one of two directions along its longer, “easy” axis. When it starts to do so, the fringing fields from the neighboring magnets will determine which direction it falls into.
To make sure that the nanomagnet is influenced only by the correct neighbor when it makes the transition, the clocking circuitry can also be used to pull nearby nanomagnets that are not supposed to influence the magnet—the ones that are downstream of the flow of information—into the same 90-degree state. In that orientation, they can’t influence which direction a neighboring nanomagnet will switch to.
In 2012, we built such a clock system out of copper wires clad with ferromagnetic material on the sides and bottom to help concentrate the magnetic field. We demonstrated that nanomagnets can indeed be switched using the magnetic fields created when current was run down those wires.
The downside of this scheme is that wires are a relatively energy-intensive approach to clocking. They dissipate quite a bit of heat and emit unfocused magnetic fields, even with cladding. But it turns out that a single clock line could control many parallel ensembles. If a nanomagnetic circuit has enough devices—say 100,000 or so—the clock energy needed per device will be acceptably low. As always, though, there will be a trade-off. The more devices you have that are governed by the same clock, the less control you’ll have over each individual device, which will translate to a higher error rate.
Tale of Two Adders: Nanomagnet circuits, such as the adders shown here, can be made in two ways. One method (top) builds nanomagnets in two dimensions, with their magnetization in the plane of the substrate. Another approach (bottom), which might prove easier to implement in large circuit designs, builds nanomagnets from many thinly deposited layers. In this case, the magnetization points out of the plane of the substrate.Images: Top: Edit Varga/University of Notre Dame/IEEE (2); Bottom: Stephan Breitkreutz-v. Gamm/Technical University of Munich
Our initial work at Notre Dame has been based on dots patterned in a thin film of magnetic material. Each nanomagnet is a two-dimensional ellipsoid, in the plane of the chip. But we encountered a tricky circuit-design limitation with this approach. To take full advantage of the space on a chip, you must find a way to build circuits that take up the full two-dimensional surface and so can pass information in both the x and y directions.
The most straightforward way to do this is to make all the magnets on the chip the same, so they all have their magnetization pointing along the same axis. For the sake of argument, let’s pick the y-axis. All the nanomagnets will be elongated in that direction. But the magnets that pass information along the y-axis will have to be strung end to end, while the ones that pass information in the x direction will have to be arranged side by side.
This dual arrangement makes circuit design quite complicated. The switching behavior of these two configurations of nanomagnets is different. As a result, signals will travel at different speeds depending on whether they’re traveling in the x or the y direction, making synchronization of signals much more complicated than it is in traditional silicon circuits.
To get around this limitation, we began collaborating in 2009 with groups led by Doris Schmitt-Landsiedel and Paolo Lugli at the Technical University of Munich. Schmitt-Landsiedel’s team has pioneered the construction of nanomagnet logic devices that have all their easy axes aligned out of the plane of the chip. Instead of building the magnets from small patches of material, the team constructs them out of multiple, alternating layers of elements such as cobalt and platinum. The interface between these layers is magnetized vertically, creating what is essentially a bar magnet standing up on one end, perpendicular to the plane of the chip.
The process begins by depositing thin films and then patterning them with a focused-ion-beam instrument. The ion beam destroys the clean interface between materials where it hits, so it can be used to pattern the smooth film into individual islands small enough to exhibit single-domain behavior. The Munich groups recently built the same sorts of logic gates and adders that our group demonstrated with in-plane nanomagnets.
Encouragingly, the out-of-plane design offers the opportunity for new ways of clocking. In late 2013, a team led by Sayeef Salahuddin at the University of California, Berkeley, showed that out-of-plane nanomagnets could be clocked using a sheet of magnetic material placed beneath them. Owing to a principle called the spin Hall effect, electrons with a particular spin can be made to concentrate underneath a nanomagnet, creating a magnetic field that can alter its orientation. At the circuit level, this approach could be hundreds or even thousands of times more energy efficient than clocking by using current-
And that’s very good news indeed for the prospect of a highly efficient computer. One application where we expect nanomagnets to be particularly useful is in data-intensive, high-throughput applications, such as filtering, polynomial evaluation, and discrete Fourier transforms. These sorts of computations, which constitute the backbone of image and signal processing, can be greatly speeded up by being executed in a pipeline. This is a strategy where computations run in parallel as much as possible, and pieces of data that have already been computed are saved until the entire computation is ready to move on to the next step. In ordinary computers, this approach requires adding extra circuits to actively hold onto that data, adding to the energy consumed by the chip. But nanomagnets, which naturally retain states until they are changed, would passively hold the data until it’s ready to be used.
When will we see this technology in chips? Those of us pursuing novel computing devices are in an awkward position now. Any such technology has to compete with silicon systems that have been endlessly optimized for decades. The success of nanomagnet logic—or any future logic, for that matter—will depend on all kinds of factors, not just technological ones.
At some point though, it’s quite likely that the need for more-energy-efficient circuitry will trump the convenience of silicon. When that happens, magnetic logic will finally have its moment.
This article originally appeared in print as “Better Computing With Magnets.”
About the Authors
Wolfgang Porod is a professor of electrical engineering, and Michael Niemier is a professor of computer science and engineering. Both are based at the University of Notre Dame, in Indiana. Porod, who used to focus on semiconductor devices, says working with electric charge is more straightforward: “I actually promised myself early on that I would never work on magnetism. But it turned out to have fantastic properties.”