Advances in machine learning and artificial intelligence on various cognitive tasks, including computer vision and natural language processing, have been accompanied by a surge in hardware development to meet the high computational requirements. However, machine intelligence processed in conventional von-Neumann architectures are still orders of magnitude more inefficient than the biological brain. Moreover, building energy-efficient ML hardware accelerators for edge applications faces further challenges, due to constraints of power budget and on-chip memory. Hence, fundamentally new approaches are needed to sustain a continuous growth in the performance of computers beyond the end of CMOS technology roadmap. In order to achieve a better match between the hardware primitives and computational models, exploring new paradigms of computing necessitates a multi-disciplinary endeavor across the stack consisting of devices, circuits, hardware architectures, and learning algorithms. Specifically, such holistic endeavors will involve exploration of novel learning algorithms inspired by bio-plausible principles, design of hardware architectures best suited for data-intensive machine learning models, together with the creation and integration of novel device technologies (such as spintronic devices) that can efficiently mimic neuronal/synaptic operations in biological brains. In this talk, I will discuss our recent exploration of exploiting spintronic devices for emerging computing paradigms such as analog in-memory computing, systolic array, and neuromorphic computing in pursuit of robust and efficient ML hardware. I will first present sparsity-aware device circuit co-design of spin-orbit-torque MRAM for robust crossbar-based ML inference engine. Significant energy improvement with near-software accuracy is demonstrated leveraging robust crossbar arrays with low precision analog-to-digital conversion. We further investigate technology selection among various emerging non-volatile memory under realistic area budgets, and identified the scenarios where spin-orbit-torque MRAM may have advantages in the hardware performance compared to non-volatile memory. Moreover, towards the development of bio-plausible neuromorphic hardware I will introduce a multi-granular spintronic device that can mimic a leaky integrate-and-fire spiking neuron with compact footprints and high energy efficiency. Incorporation of such neuronal devices into the training of a deep convolution spiking neural network for image classification demonstrates improved robustness against various types of noise injection. We conclude with a brief discussion on potential opportunities and directions for future work.
|