MRAM-based In-Memory Computing for Efficient Acceleration of Generative Adversarial Networks

Generative Adversarial Networks (GANs) have displayed favorable outcomes in generating high-quality samples for various applications. Nevertheless, the computational complexity associated with GANs poses a significant challenge to their practical implementation. To resolve this issue, this work proposes an implementation for GAN acceleration employing Processing-in-Memory (PIM) architecture. Magnetic Random Access Memory (MRAM) is a non-volatile memory technology that enhances the potential of PIM architecture due to its low-power consumption, high-density, and fast access times. The proposed work introduces an MRAM-based PIM architecture to accelerate matrix multiplication operations in GANs. The study aims to assess and compare the effectiveness of the MRAMs for Processing-in-Memory (PIM) based architecture for implementing GAN models by analyzing their performance. The computationally intensive GAN models such as WGAN demonstrate a competitive edge in generating images with high correlation to ground truth images compared to simpler architectures such as VanillaGAN. PIM architecture with Spin-Orbit Torque (SOT) MRAM has demonstrated substantial improvements in the order of 10x, 200x, 1.4x as compared to previously proposed architectures using Resistive Random Access Memory (RRAM) crossbars in terms of energy, power, and latency, respectively, in implementing WGANs inference operations.