Read Noise Analysis in Analog Conductive-Metal-Oxide/HfO𝑥 ReRAM Devices

Abstract

Introduction Analog in-memory computing with resistive memory devices is a compelling alternative to conventional digital vonNeumann computers, enhancing both deep neural network inference and training [1]. Recent advancements in learning algorithms and hardware optimizations have enabled the utilization of Conductive-Metal-Oxide (CMO)/HfO𝑥 ReRAM technology for training purposes [2]. The enhanced analog conductive properties and diminished switching stochasticity, ascribed to a defect concentration change in a dome within the CMO layer, contrast with conventional filamentary Metal/HfO𝑥/Metal ReRAM devices where such changes typically occur in a nanoscale tunneling gap in the HfO𝑥 layer [3]. Furthermore, the demonstrated attributes of high endurance and prolonged retention pertaining to multilevel conductive states instill renewed interest in leveraging CMO/HfO𝑥 technology for inference applications as well [4]. To effectively harness the potential of this technology for both applications, a comprehensive understanding of the intrinsic sources of noise is required. Prior research on nanometer-scale devices has demonstrated noise properties being contingent on the active device volume (and associated resistance), frequency, and applied voltage [5]. Accordingly, low-frequency noise measurements emerge as a comprehensive indicator, offering valuable insights into the transport and noise-generating mechanisms within the investigated nanoelectronic systems [5]. This study unveils the inaugural investigation into read noise in CMO/HfO𝑥 ReRAM devices and compares it with other systems. Methods: The top view and cross-section of the 1T1R device and the ReRAM material stack are depicted in Fig. 1, respectively [2]. The ReRAMs were first formed and repeatedly switched between their high resistive state and low resistive state, as shown in Fig. 2. To evaluate read noise, the ReRAMs were potentiated (depressed) with 60 ns long identical positive (negative) pulses to the target conductance. Subsequently, a read pulse train of amplitude𝑉read = 0.2V was applied to the ReRAMs’ top electrode. Two-Level and Multi-Level Random Telegraph Noise (RTN) and 1/ 𝑓 noise were statistically examined. To study trap properties, such as the capture and emission time constants 𝜏𝑐 and 𝜏𝑒, a Gaussian Mixture Model was fitted to the distribution of read current samples; RTN signals were then separated from the background 1/ 𝑓 according to [6] (see Fig. 3). Fig. 4 presents the results of the RTN analysis from the ∼200 current time traces where RTN, rather than 1/ 𝑓 , was the dominant noise source (∼5% incidence), for various sampling frequencies and device conductance. Fig. 5 and Fig. 6 instead show multi-device statistics of the full unfiltered dataset. Results: To ensure that read noise in the 1T1R cell originates from the ReRAM devices, we measured the 1/f noise on a standalone transistor and observed that it was considerably lower than in our analog ReRAM (Fig. 3b). The relationship between 𝜏𝑒 and 𝜏𝑐 can reveal information about the alignment of the trap energy level (𝐸𝑇) with respect to the Fermi energy level (𝐸𝐹) of the switching electrode/interface. From Fig. 4a, no apparent skew towards either time constant extracted from RTN is observed, but numerous data points lie close to the line representing 𝜏𝑐 = 𝜏𝑒 (implying 𝐸𝑇 ≈ 𝐸𝐹), which can be used to estimate the spatial position of the defect in the oxide layer using the modified McWhorter model [7]. A complete picture of read noise in the devices is then presented by evaluating ∼5000 current measurements and studying their standard deviation against read duration (Fig. 5) and device conductance (Fig. 6). Fig. 5 shows the average standard deviation of read noise E [𝜎𝐺(𝑡)] for 4 bits. We observe a general log(𝑡) trend corresponding to a 1/ 𝑓 average Power Spectral Density (PSD), as expected, but it stabilizes within 5 seconds. These results can be directly used in analog AI simulator toolkits, such as IBM’s AIHWKit [8], to evaluate our ReRAM’s inference performance. In Fig. 6 we compare the saturated 𝜎𝐺(𝐺) of our CMO/HfO𝑥 ReRAM devices with those of other HfO2 based ReRAM devices and PCM technology reported in literature. The comparatively low read noise in our analog ReRAMs makes them a promising choice not only for neural network training, but also inference applications. Conclusions: In this work, we present a comprehensive read noise analysis and provide physical information extracted from RTN that will help in further developing device models for analog filamentary CMO/HfO𝑥 ReRAMs. Overall noise performance is competitive with comparable technologies. Noise saturation at seconds timescale and good retention show promise for inference and training applications alike. [1] G.W. Burr et al., https://spectrum.ieee.org/analog-ai (2021) [2] N. Gong et al., IEEE-IEDM p. 33.7.1 (2022) [3] D.F. Falcone et al., IEEE-IMW, p. 1 (2023) [4] T. Stecconi et al., Nano Letters 24(3), p. 866 (2024) [5] S. Ambrogio, et al., IEEE-IDEM p. 31.5.1, (2013) [6] G. Guisi et al., Rev. Sci. Instrum. 79, 024701 (2008) [7] Yu et al., Phys. Rev. B 85, 045324 (2012) [8] M. Le Gallo et al., APL Machine Learning 1, 041102 (2023) [9] F. M. Puglisi et al., 2014 IEEE-IRPS, p. MY.5.1 (2014) [10] X. Sun et al., IEEE Tran. El. Dev. 68(11), p. 5585 (2021) [11] D. Lee et al., IEEE El. Dev. Lett. 32(7), p. 964 (2011)