Reconfigurable stochastic neurons based on strain engineered low barrier nanomagnets

Area of Science:

• Unconventional computing
• Materials science
• Nanotechnology

Background:

• Stochastic neurons, both binary (BSN) and analog (ASN), are hardware accelerators for complex computational problems.
• Low barrier nanomagnets (LBMs) can implement these neurons, with BSNs exhibiting a double-well potential and ASNs lacking a significant energy barrier.
• The key difference lies in the potential energy barrier height, dictating binary versus analog state fluctuations.

Purpose of the Study:

• To investigate strain-engineered barrier control in stochastic neurons.
• To explore the reconfiguration between binary and analog stochastic neurons using local strain.
• To demonstrate the potential for integrating diverse computational functionalities on a single hardware platform.

Main Methods:

• Utilizing magnetostrictive low barrier nanomagnets (LBMs) for stochastic neuron implementation.
• Employing electrically generated local strain to dynamically adjust the energy barrier of LBMs.
• Studying the effects of strain-mediated barrier control on neuron behavior and reconfiguration.

Main Results:

• Demonstrated the ability to reconfigure analog stochastic neurons (ASNs) to binary stochastic neurons (BSNs) and vice versa by modulating the energy barrier via strain.
• Showcased that magnetostrictive LBMs allow for low-energy, dynamic reconfiguration of neuron functionality.
• Identified applications beyond reconfiguration, including adaptive annealing and dynamic memory hierarchy emulation.

Conclusions:

• Strain-engineered barrier control offers a powerful method for creating field-programmable unconventional computing architectures.
• This approach enables the integration of diverse computational fabrics (e.g., optimization and sequence learning) on the same substrate.
• The technology paves the way for highly versatile and efficient computing systems with minimal energy cost for reconfiguration.