A Thousand State Superlattice(SL) FEFET Analog Weight Cell
Khandker Akif Aabrar, Sharadindu Gopal Kirtania, Anni Lu, Abhishek Khanna, Wriddhi Chakraborty, Matthew San Jose, Shimeng Yu, Suman Datta
Abstract
In-situ training of a deep neural network (DNN) requires high performance analog weight cell. In this work, we experimentally demonstrate a BEOL-compatible scaled-superlattice (FE-3nm/DE-4nm/FE-3nm) ferroelectric-FET analog weight cell with a maximum conductance level of 1,000 states, good linearity (α <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</inf> /α <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">d</inf> =-2/-1.58), excellent symmetry (α <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</inf> ~α <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">d</inf> =0.42), fast-switching speed (100ns) and large dynamic range (G <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">max</inf> /G <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">min</inf> =80) using identical weight update pulses. System-level benchmarking shows that the SL FEFET analog weight cell can improve the online training accuracy of DNN to 91.05% with record energy consumption and latency. These results demonstrate the feasibilty of superlattice(SL) FEFET for high performance and energy efficient neuromorphic computing.