Neuromorphic Computing

Neuromorphic computing is a promising strategy to overcome fundamental limitations, such as enormous power consumption, by massive parallel data processing, similar to the brain. The idea behind neuromorphic computing is the idea to process information by millions of small processing units in a parallel fashion. These processing units can be rather simple and could be implemented by a wide range of approaches. Thus, a great implementation needs to be great in terms of energy efficiency, footprint, and processing speed. In addition, it should be scalable to networks of millions of devices.

The other properties that we are looking for are non-volatility of the memories and non-linearity of the processing component. Essentially, the memories should keep their memory over long time periods and the processing units should activate above a certain threshold value.

Here we investigate one of the top candidates for this solution based on spintronic nanodevices called magnetic tunnel junctions. Spintronics simply describes that this nanodevice makes use of not just the electrical properties of the electrons, but also the more exotic spin property of the electrons.

The proof-of-principle of the Weighted Spin Torque Nano-Oscillator

In this work we want to present the concept and proof-of-principle of the weighted spin torque nano-oscillator (WSTNO) system, which combines multiple spintronic devices with different functions to create a basic programmable computing unit. We implement an artificial neuron as a network of two memories (synapses) and one oscillator (neuron) nanofabricated from the same multifunctional material stack and only varying in lateral dimension. This allows the fabrication of dense magnetic tunnel junction nanopillar networks.

Two smaller elliptical (75 nm x 225 nm) magnetic tunnel junctions are used as memories and one larger circular (300 nm) magnetic tunnel junction is used as processing unit, as shown in Fig. 1. The memories each have two states, called parallel state and anti-parallel state describing the direction in which the magnetization of so-called free layer is pointing. Depending on the direction, the resistance of the memory is either 1000 Ohm or 1700 Ohm. Although the processing unit is made of the same material it functions quite differently. Due to the different shape, the magnetization forms a circular magnetization state that is called vortex state. Then we inject a spin polarized electric current in the vortex structure, meaning the current has excess of one spin direction, and this injection of the excess spin into the vortex structure leads to a motion of the vortex core. A large enough electric current is able to continuously rotate the vortex core around the center of the nanopillar. Thus, this device is called spin torque nano-oscillator. Due to the physics behind this spin transfer torque effect, the oscillation power follows naturally a threshold function behavior that we use in this processing unit.

We measure an integrated oscillator output power above 1 µW and frequency around 245 MHz in nanopillars of 300 nm diameter, as shown in Fig. 2. Due to this high output power and low oscillation frequency, monolithic integration with CMOS technology is less challenging. The required footprint is about 0.017 µm² per memory and 0.09 µm² per oscillator, which is very small compared to CMOS-based artificial neurons that are usually more than 100 times larger.

Our Long-Term Vision

The weighted spin torque nano-oscillator is a basic programmable computing unit for neuromorphic computing systems that combines the non-volatility and non-linearity properties of magnetic tunnel junctions. This circuit is fabricated in a single fabrication process from a multifunctional material stack, which paves ways to fabricate more complex neuromorphic computing systems. Our long-term vision is to develop a unique platform for next-generation complex neuromorphic computation system with improved performance compared to existing CMOS computing systems, filling the gap between the capability of current computers and the brain.