An Analog Neuromorphic On-Chip Training System With IGZO TFT-Based 6T1C Synaptic Memory

This article proposes an analog synapse-based neuromorphic on-chip training system that uses emerging indium gallium zinc oxide (IGZO) thin film transistor (TFT) synapse cells to store multi-bit states for deep neural networks (DNNs). IGZO TFT demonstrates extremely low leakage currents, preserving the charge stored in capacitors during prolonged training periods. The 6 transistor 1 capacitor (6T1C) structure, characterized by its symmetrical design and current sources configuration, achieves an average of 367 distinct states with high linearity, reflected by an <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</i><sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> value of 0.99 through a neuron circuit. By adjusting currents and capacitor sizes, the system effectively integrates currents from both individual synapses and the overall array. Additionally, the neuron circuit, implemented separately from the IGZO TFT synapse array, demonstrates an 8.95 effective number of bits (ENOB) in overall performance measurements. The neuron circuit and IGZO TFT array have areas of 7.2 and 10.2 mm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>, respectively. Using the proposed neuromorphic system with the 6T1C memory structure, we successfully conducted the first analog on-chip training with the last layer, achieving an accuracy of 97.1% on the MNIST dataset.