Emerging Non-volatile Memory Devices for Neuromorphic Computing
Abundant data applications such as image recognition, voice activated assistance, and disease diagnostics require novel hardware platforms to efficiently run emerging machine learning algorithms. The conventional scaling of today's silicon-CMOS technology cannot satisfy the speed & energy-efficiency needs of such applications, as they require real-time analytics on enormous quantities of user data. Neuromorphic computing takes inspiration from the way the brain computes, with information processing featuring a co-localized memory and logic system where memory (1015 synapses) is distributed with the processing (1011 neurons), unlike computing systems based on von Neumann architecture where we have one or more central processing units physically separated from memory. Of particular interest, synapse is the key component that determines plasticity by varying the conductance level between the interacting neurons. Given the much larger number of synapses compared to neurons, developing nanoscale synaptic hardware elements with gray-scale or multi-level programming capability is essential to implement the high-density, energy-efficient neuromorphic computing system. In this work, spin-transfer-torque magnetic random-access memory (STT-MRAM) that features a superior lifetime compared to other emerging nonvolatile memories that use destructive write techniques, is investigated as a spintronic synaptic device by focusing on the novel characterization techniques for analog programming. Another emerging non-volatile memory candidate, resistive random-access memory (RRAM), is also fabricated and investigated for the purpose of best tuning its characteristics as highly stable synaptic elements.