We are in the midst of a computing revolution, as cutting-edge research is fundamentally re-imagining how computers think and process information to enable and improve next-generation applications. This computational re-programming brings with it a litany of ongoing challenges for engineers and scientists.
Researchers at The University of Texas at Austin are studying a magnetic device that breaks down key barriers for how information is stored and processed as part of the quest to make computers think more like humans. Jean Anne Incorvia, an assistant professor in the Cockrell School's Department of Electrical and Computer Engineering, recently published a trio of papers seeking to solve challenges facing these new computing paradigms.
“There are a lot of new physics and materials that we are always discovering, and how you translate those into applications could give us the ability to do a whole new breadth of things in computing,” Incorvia said.
Incorvia noted that we are beginning to reach the limits of how small and powerful silicon chips can be. This has initiated a race to develop new building blocks for computers that can think like human brains to accomplish complex tasks.
The push toward neuromorphic computing can enable dynamic applications that require processing devices able to shift and adapt to changing environments on the fly. Some examples include circuits for self-driving vehicles, pattern and image recognition systems and Internet of Things devices.
"Why stick with silicon when we can create devices with bio-inspired behaviors that could be used for furthering and expanding what computing can do?" Incorvia said.
A New Device to Solve Next-Gen Computing Issues
In a new paper, Incorvia and her students solved challenges related to a new in-memory computing device. These magnetic devices can simultaneously store information and process it, providing a huge speed and energy improvement over traditional computing devices.
The new device, called a domain wall-magnetic tunnel junction and described in Applied Physics Letters, was a fabrication triumph. It was built in such a way that it achieves a remarkably high on/off ratio, given its more complex structure compared to traditional magnetic memories, which is important for translating information stored in the magnetic component to ones and zeros for computing purposes. The higher the on/off ratio, called the tunnel magnetoresistance, the easier it is to distinguish the ones and zeros. A low on/off ratio means the ones and zeros start to blur together, making it harder to compute with the devices.
Incorvia says groups at Intel and IMEC have made similar devices with an on/off ratio of about 15%. An MIT group made one with a 40% ratio. Incorvia's device boasts an on/off ratio between 170% and 200%.
The device is made of a stack of nano-meter-thin layers, grown in collaboration with Applied Materials. Patterning the layers into nanostructured devices without damaging them was key to the on/off achievement, a process led by Thomas Leonard, a graduate student in Incorvia's lab.
The devices described in the paper represent the fifth iteration. Now Leonard is on the ninth version. The fabrication was done using UT Austin’s Microelectronics Research Center, a world-class clean room facility for processing nano-devices, and the research was supported by Sandia National Laboratories’ Laboratory Directed Research and Development (LDRD) Program.
The new devices use a new method of controlling the magnetization with applied current, called spin orbit torque, which enables the devices to switch at lower energies and more reliably. With the combination of high tunnel magnetoresistance and reliable switching behavior, they were able to show, for the first time, that the domain wall-magnetic tunnel junctions can repeatably operate in a circuit. This opens up the ability to build in-memory computing circuits with magnetism.
Computers With Muscle Memory
We all do things every day that feel like second nature. These repetitive actions become so simple because the brain and body are used to performing them.
Incorvia and her team have applied this principle of muscle memory to their magnetic device by showing it has an "edgy-relaxed" behavioral capability. As described in IEEE Magnetics Letters, this allows the neurons to fire faster when doing a repetitive task, speeding up the processing abilities.
Incorvia’s team showed using device and circuit modeling that the concept comes in handy when doing tasks like image recognition. The magnetic devices could recognize similar types of images at a faster rate, while not taking up as much computing resources, leaving plenty of bandwidth for harder tasks. The method could be useful in a number of other applications, including speech recognition where there are many frequently repeated sounds and phrases.
The researchers showed the devices are adaptable. Depending on the repetitiveness of the dataset, the “edgy-relaxed” behavior can be altered to most efficiently process the data. This is a step closer to using new hardware for context-aware computing. This work was supported by a National Science Foundation CAREER award.
Magnetic devices like the ones Incorvia's team creates tend to hold up well under high-radiation conditions, in a way silicon chips don't. This is because radiation can create electrical charges that affect electrical-based devices much more than magnetic-based devices. This is particularly relevant to space applications, where high levels of radiation are a challenge.
In the third paper, published in IEEE Transactions on Nuclear Science, the team tested the thin film stack detailed in Applied Science Letters for resistance to radiation. The researchers worried that the structural differences between the new device and other magnetic devices would nullify its radiation resistance. That wasn't the case, and the new device stood up well to radiation exposure.
The team applied high doses of radiation and used resources at the Texas Materials Institute to analyze what happens when breakdown does occur. They found that high levels of radiation affect the nanometer-thick layers at the bottom of the multi-layer stack more than layers at the top of the stack. This finding will help inform design considerations when adapting these magnetic devices to high radiation environments.
“Using nanomagnetism in computing combines cutting-edge ideas in electrical engineering, computer engineering, physics, materials science and neuroscience,” Incorvia said. “We are just starting out on what can be done, and these results provide some clear directions on where to go next.”