Device Codesign using Reinforcement Learning

We demonstrate device codesign using reinforcement learning for probabilistic computing applications. We use a spin orbit torque magnetic tunnel junction model (SOT-MTJ) as the device exemplar. We leverage reinforcement learning (RL) to vary key device and material properties of the SOT-MTJ device for stochastic operation. Our RL method generated different candidate devices capable of generating stochastic samples for a given exponential distribution.

I. Introduction

Device codesign for a given application is often a tedious process dependent on labor-intensive and time-consuming simulations, fabrication, and testing. However, there is tremendous opportunity to customize devices for particular applications in order to get the best performance possible – whether that be a particular capability, energy usage, latency or throughput, or some other metric or combination of metrics of interest. Emerging devices, such as magnetic tunnel junctions, often have simulation models that abstract device behavior based on key materials and device properties. To effectively leverage the properties of these devices for new computing capabilities (e.g., probabilistic computing), an automated codesign framework is critical to account for application and algorithm constraints while designing these devices. We propose a reinforcement learning (RL) guided framework for device codesign which can account for application and algorithm requirements while optimizing device parameters.