Impact of RTN on Pattern Recognition Accuracy of RRAM-Based Synaptic Neural Network

Abstract

The version presented here may differ from the published version or from the version of the record. Please see the repository URL above for details on accessing the published version and note that access may require a subscription.

Committee machines—a universal method to deal with non-idealities in memristor-based neural networks

“… First author (year) Non-ideality Device type Proposed solution C. Sung (2018) 31 Current/voltage nonlinearity TaO x RRAM Hot-forming step is adopted C. Li (2018) 15 Current/voltage nonlinearity Ta/HfO 2 RRAM 1T1R architecture is adopted Y. Fang (2018) 32 Device-to-device variability HfO x RRAM Ultra-thin ALD-TiN buffer layer is introduced B. Govoreanu (2013) 33 Device-to-device variability Al 2 O 3 /TiO 2 (VMCO) RRAM Non-filamentary RRAM is adopted A. J. Kenyon (2019) 34 Device-to-device variability SiO x RRAM The roughness of bottom electrodes is increased L. Xia (2017) 14 Faulty devices – A modified mapping algorithm and redundancy schemes are used S. Ambrogio (2018) 7 Limited dynamic range PCM Two pairs of conductance of varying significance for every synaptic weight are used M. Hu (2016) 17 Line resistance – Advanced mapping algorithms are used to compensate for line resistance effects W. Wu (2018) 35 Programming nonlinearity HfO x RRAM Electro-thermal modulation layer is deposited on the switching layer J. Woo (2016) 9 Programming nonlinearity HfO 2 RRAM Bilayer structure is adopted S. Ambrogio (2018) 7 Programming nonlinearity PCM PCM devices are used together with CMOS transistors Z. Chai (2018) 36 Random telegraph noise TiO 2 /a-Si (aVMCO) RRAM …”

Exploring the Impact of Random Telegraph Noise-Induced Accuracy Loss on Resistive RAM-Based Deep Neural Network

“…DNNs were trained with the gradient descent backpropagation until the accuracy reached a level similar to their reported value. To carry out the simulation, the well-trained weights in each layer were mapped to two simulated RRAM arrays which handle the positive and negative weights separately [4]. We use conductance between 1.25μS (800kΩ) to 12.5 μS (80kΩ), which is the range for our measured data.…”

“…By performing matrix-vector multiplication in cross-bar arrays, resistive-switching memories (RRAM) have successfully lowered down the power consumption to nW level [2][3]. Therefore, the study of the interaction between non-ideal characteristics of RRAMs and the inference accuracy becomes essential and has attracted attention from both industry and academic in recent years [4][5]. Among all the non-ideal characteristics, the RRAM cell variation and the Random Telegraph Noise (RTN) can cause the deviation of the weights away from their pre-set values.…”

GeSe-Based Ovonic Threshold Switching Volatile True Random Number Generator

“…Emerging memory devices are promising for a broad spectrum of applications such as storage class memory [11] and neuromorphic computing [12]. Several TRNGs based on emerging memory devices have been proposed, utilizing RTN [8], cycle-to-cycle variability [13] or the stochastic delay time Manuscript received dd/mm/yyyy; revised dd/mm/yyyy; accepted dd/mm/yyyy.…”