Energy-Efficient In-Memory Architectures Leveraging Intrinsic Behaviors of Embedded MRAM Devices

For decades, innovations to surmount the processor versus memory gap and move beyond conventional von Neumann architectures continue to be sought and explored. Recent machine learning models still expend orders of magnitude more time and energy to access data in memory in addition to merely performing the computation itself. This phenomenon referred to as a memory-wall bottleneck, is addressed herein via a completely fresh perspective on logic and memory technology design. The specific solutions developed in this dissertation focus on utilizing intrinsic switching behaviors of embedded MRAM devices to design cross-layer and energy-efficient Compute-in-Memory (CiM) architectures, accelerate the computationally-intensive operations in various Artificial Neural Networks (ANNs), achieve higher density and reduce the power consumption as crucial requirements in future Internet of Things (IoT) devices. The first cross-layer platform developed herein is an Approximate Generative Adversarial Network (ApGAN) designed to accelerate the Generative Adversarial Networks from both algorithm and hardware implementation perspectives. In addition to binarizing the weights, further reduction in storage and computation resources is achieved by leveraging an in-memory addition scheme. Moreover, a memristor-based CiM accelerator for ApGAN is developed. The second design is a biologically-inspired memory architecture. The Short-Term Memory and Long-Term Memory features in biology are realized in hardware via a beyond-CMOS-based learning approach derived from the repeated input information and retrieval of the encoded data. The third cross-layer architecture is a programmable energy-efficient hardware implementation for Recurrent Neural Network with ultra-low power, area-efficient spin-based activation functions. A novel CiM architecture is proposed to leverage data-level parallelism during the evaluation phase. Specifically, we employ an MRAM-based Adjustable Probabilistic Activation Function (APAF) via a low-power tunable activation mechanism, providing adjustable accuracy levels to mimic ideal sigmoid and tanh thresholding along with a matching algorithm to regulate neuronal properties. Finally, the APAF design is utilized in the Long Short-Term Memory (LSTM) network to evaluate the network performance using binary and non-binary activation functions. The simulation results indicate up to 74.5 x 215; energy-efficiency, 35-fold speedup and ~11x area reduction compared with the similar baseline designs. These can form basis for future post-CMOS based non-Von Neumann architectures suitable for intermittently powered energy harvesting devices capable of pushing intelligence towards the edge of computing network