CMOS compatible Materials and Devices for beyond von Neumann

Abstract

Brain inspired computing is a promising paradigm of artificial intelligence (AI) systems that aims at developing an efficient computing architecture that resembles the biological brains. The development of novel materials and devices with neural and synaptic functions incorporated into unique architectures will allow the implementation of a computing system that can efficiently perform the heavy vector - matrix manipulation inherent to AI workloads with O(1) time complexity[1]. Memristors are key building blocks for the realization of the artificial neural and synaptic function in neuromorphic computing. [2] These devices rely on diverse physical mechanisms and materials and their understanding via experimental and theoretical means is pivotal to the device optimization and coupling to the higher layers of the computer architecture. We will see examples of how material and device engineering can lead to breakthroughs in device performance. Nonetheless, device variability among other non-idealities still hinders the hardware scaling to large networks required to solve more complex AI tasks. Challenges at device and hardware level may also be overcome through a complementary research effort to develop more robust, hardware-friendly algorithms and computational models that could compensate for the variability issues of devices[3].