Abstract
The growth of big data has enabled the proliferation of modern technology along with sophisticated gadgets, driven by the von Neumann architecture for the past 70 years. However, the von Neumann bottleneck that results from the current system limits data processing. Moreover, the continuation of Moore’s law is debatable and whether it ensues or not, will both create complications for future data processing limited by the bottleneck. The fast progress of artificial intelligence (AI) and machine learning requires highly efficient and powerful computational capabilities. To overcome this, brain-inspired systems, such as in-memory and neuromorphic computing have been designed that are not restrained by Moore’s law and von Neumann bottleneck. Nevertheless, the energy consumption is still large and there is a need to seek ways to further minimize energy consumption across various implementations.
In-memory computing localizes data storage and processing operations, and it offers a pathway to decrease the energy cost of computations. However, the task of designing appropriate material platforms is highly challenging, especially for next generation devices that can bridge the time gap between different conventional memories. This thesis first demonstrates an in-memory reprogrammable shift register using all four material (M) states of a doped phase-change Ge2Sb2Te5 material. Additional simulations showcase the strong potential of this material in the realm of neuromorphic computing. This offers a pathway towards the advancement of future low-power two-terminal electronic systems that can overcome the problem in the memory hierarchy.
Next, to address the concerns of modern hardware security, this thesis proposes leveraging the large degree of randomness in the amorphous (A) state variations within the Ge2Sb2Te5 material, to build a framework for in-memory physically unclonable function (PUF). Consequently, this creates an opportunity for an unbreakable anticounterfeiting system, which can be reconfigured in a rapid manner.
Nevertheless, the von Neumann architecture will stay relevant in years to come, which necessitates solutions to concurrently minimize the energy consumption on traditional computers. The modelling of compliance-current in resistive random-access memory has been performed in earlier studies via machine learning. However, to achieve better accuracy, more training data is needed, yet the energy efficiency from data movement is constrained by the bottleneck. This thesis thus explores a brain-inspired learning method that can bypass retraining the whole dataset and yet maintain a sufficiently high accuracy in compliance-current prediction. Moreover, this thesis seeks to convert the problematic resistance drift issue in a phase-change Ge2Sb2Te5 material into an energy reduction solution for reinforcement learning, via the decay-based proximal policy optimization (DP-PPO) method. This will be implemented for training a state-of-the-art fighting agent which presents excellent versatility. As a result, both methods pave the way towards energy-efficient von Neumann systems for solving modern problems.
In a nutshell, integrated memristive systems are accomplished by: 1) merging data storage and processing through an in-memory approach in shift registers and PUF, 2) harnessing the natural characteristics of phase change memristive system to support reinforcement learning on conventional computers and developing brain-inspired models that synergize with memristive systems in controlling compliance current. Most importantly, these discoveries offer a dual strategy to address energy consumption in both von Neumann and non-von Neumann computing for future applications.