An efficient deep neural network accelerator using controlled ferroelectric domain dynamics

Abstract

The current work reports an efficient deep neural network (DNN) accelerator where analog synaptic weight elements are controlled by ferroelectric domain dynamics. An integrated device-to-algorithm framework for benchmarking novel synaptic devices is used. In P(VDF-TrFE) based ferroelectric tunnel junctions, analog conductance states are measured using a custom pulsing protocol and associated control circuits and array architectures for DNN training is simulated. Our results show precise control of polarization switching dynamics in multi-domain, polycrystalline ferroelectric thin films can produce considerable weight update linearity in metal-ferroelectric-semiconductor (MFS) tunnel junctions. Ultrafast switching and low junction current in these devices offer extremely energy efficient operation. Through an integrated platform of hardware development, characterization and modelling, we predict the available conductance range where linearity is expected under identical potentiating and depressing pulses for efficient DNN training and inference tasks. As an example, an analog crossbar based DNN accelerator with MFS junctions as synaptic weight elements showed > 93% training accuracy on large MNIST handwritten digit dataset while for cropped images, a > 95% accuracy is achieved. One observed challenge is rather limited dynamic conductance range while operating under identical potentiating and depressing pulses below 1V. Investigation is underway for improving the FTJ dynamic conductance range maintaining the weight update linearity under identical pulse scheme.

Back‐End and Flexible Substrate Compatible Analog Ferroelectric Field‐Effect Transistors for Accurate Online Training in Deep Neural Network Accelerators

“…[ 5 ] In our previous work, we have shown that it is challenging to obtain a large dynamic conductance range ( G max /G min ) in FTJ devices with NLS domain switching characteristics, with programming pulses of 1 V magnitude and ns durations. [ 14 ] In comparison, FeFETs show a much higher G max /G min ratio due to the large On/Off ratio of the FETs. However, the entire dynamic range cannot be used for training in most cases due to a lack of linearity and symmetry in response to potentiating and depressing pulses or due to the challenge of reading out too low current due to interfering read noise.…”

“…International Technology Roadmap for Semiconductors (ITRS) [34] predicts that MoS 2 is a potential semiconductor material for integration into nanometer-scale structures due to outstanding electrical properties exhibited by one or few monolayer thick films. [14,15] MoS 2 or other 2D semiconductor-based circuits can provide a solution to this high operating voltage problem of the FeFETs due to an indirect bandgap of 1.2 eV of few layer MoS 2 , that resulted in transistors with a high On/Off current ratio of %10 6 -10 8 under a low operating voltage. Also, MoS 2 MOSFETs exhibited a very low SS value of 74 mV decade À1 , benefiting from the absence of dangling bonds in the MoS 2 layers.…”

“…In our recent work, we have shown that precise control over ferroelectric domain dynamics in polymer ferroelectric P(VDF‐TrFE) can result in multiple distinguishable states in ferroelectric tunnel junctions (FTJs), [ 12 ] that lead to ultrafast operation, [ 13 ] conductance linearity, and efficient DNN training. [ 14 ] However, there is a compromise between the maximum and minimum conductance ratio ( G max /G min ) and the linearity of the weight update in these devices. For higher efficiency in DNN online training in larger networks, G max /G min ratio and linearity need to improve.…”

Editorial: Focus on Neuromorphic Circuits and Systems using Emerging Devices

“…Majumdar [5]: Majumdar presents an efficient deep neural network (DNN) accelerator using analog synaptic weights controlled by ferroelectric (FE) domain dynamics. They benchmark novel synaptic devices, specifically poly(vinylidene fluoride-trifluoroethylene)-based ferroelectric tunnel junctions, demonstrating their linearity in weight updates.…”

Toward Memristive Phase‐Change Neural Network with High‐Quality Ultra‐Effective Highly‐Self‐Adjustable Online Learning

“…[19][20][21][22][23][24] Experiments have demonstrated the maximum number of distinct conductance levels utilized for neural-network (NN) learning for different memristive elements, e.g., resistive-switching memory (RSM) elements, magnetic-tunnelling memory (MTM) elements, ferroelectric memory (FM) elements, and other memory elements. [25][26][27] A maximum number of distinguished conductance levels used for NN learning of 2-64 has been The T material state for in situ learning. a) PCM elements, based on the reversible switching between the crystallized state and the glassy state of a chalcogenide layer, exhibiting a marked contrast in the optical reflectivity and electrical conductivity, can reveal T-material states to develop new types of neural networks.…”