Reliability-Ensured and Fast (< 100 ns) Analog Synapse for Training Accelerators: All-Sputtered HfOy/HfO_xRRAM

We reveal that sputtered HfO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">x</inf>, which has been often neglected in resistive random access memory (RRAM) design, can be an alternative for highly reliable and fast (< 100 ns) analog synapses to accelerate training algorithms. The sub-stoichiometric HfO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">x</inf>, enables rapid current update owing to the abundant oxygen vacancies (V<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0</inf>). Additional HfOy layer with increased stoichiometry makes uniformly distributed Vo profile across the entire HfOy/HfO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">x</inf>RRAM stack, enhancing reliability in dynamic synaptic operation, where multiple pulses are continuously addressed. This not only results in robust device-to-device/cycle-to-cycle uniformity, but also ensures multilevel retention, which is further validated by real-time transient current analysis.