(Color online) (A) Plateau and decrease in performance per power density as transistor gate length is decreased. Performance degradation is due to lower limits on threshold voltage that do not allow for power density reduction at low feature sizes. (Used with permission from W. Haensch, E. J. Nowak, R. H. Dennard, P. M. Solomon, A. Bryant, O. H. Dokumaci, A. Kumar, X. Wang, J. B. Johnson, and M. V. Fischetti, IBM J. Res. Dev. 50, 339 (2006).) (B) Minimum energy per switch operation as a function of switch size for arbitrary switch design. Energy requirement increases rapidly below 5 nm. (V. V. Zhirnov, R. K. Cavin, III, J. A. Hutchby, and G. I. Bourianoff, Proc. IEEE 91, 1934 (2003). © 2003 IEEE)
(Color online) Comparison of selected computational strengths of modern computers and human brains. Computers are composed of binary transistor switches interconnected to perform Boolean functions. Computers are generally superior in complex arithmetic and logic calculation, and they are designed to have perfect memory. On the other hand, an animal brain is composed of analog neurons with adjustable weight interconnections that have vast but poorly understood functionality. Brains more easily perform generalization, pattern recognition, and are more tolerant to faults through massive parallelism.
(Color online) Illustration of how adaptive oxides may emulate biological synapses. In biological systems, electrochemical pulses transmitted from adjacent neurons alter the weight of the conjoining synapse (circled oval). In a two-terminal device with a non-volatile, adjustable internal state, excitation pulses incident on the device modify the internal state, which may be resistance, polarization, magnetization, etc., in an analogous manner as a synapse.
(Color online) List of oxides discussed in this review organized by crystal structure and experimentally reported band gaps [Refs. 197–217].
(Color online) Diagram of adaptive oxide devices surveyed in this review. The levels of the diagram are switching mechanism (2nd), internal state (3rd), representative oxides (4th), and device structures (5th). We discuss devices that have internal state modified by ferroelectricity, ion diffusion, conductive filament formation, and ferromagnetism.
(Color online) Typical I-V characteristics of (A) bipolar and (B) unipolar resistance switching. In bipolar switching, a device in a high resistance state switches to a low resistance state at high positive voltage and remains in that state until a large negative voltage is applied. In unipolar switching, a device in a high resistance state switches to a low resistance state at high voltage of either polarity and switches back to a high resistance state at lower voltage, again at either polarity. Prospective models of the respective low resistance states are given in (C) and (D). In (C), oxygen vacancy accumulation lowers the Schottky barrier at one electrode. In (D), conductive filaments that span from one electrode to the other effectively shunt the oxide. (Adapted from Materials Today, A. Sawa, Resistive switching in transition metal oxides, 11, 28, © 2008, with permission from Elsevier.)
(Color online) Oxygen vacancy drift bipolar switching mechanism for representative n-type oxide. (A) In high resistance state, there is a lack of oxygen vacancies at the interface. Carriers must overcome Schottky barrier to contribute to current. (B) In low resistance state, oxygen vacancies accumulate at the interface, reducing depletion width such that tunneling is possible.
(Color online) Typical polarization (magnetization) curve of a ferroelectric (ferromagnetic) material. The large hysteresis window in ferroelectrics is caused by a spontaneous non-centrosymmetric distortion of the unit cell, for example, in a cubic perovskite (planar view in left inset), switchable by an applied electric field. The hysteresis window in ferromagnets is caused by spontaneous parallel alignment of electron spins (right inset), switchable by an applied magnetic field. P s (M s) is the saturation polarization (magnetization). The hysteresis is the basis for non-volatile memory. The smooth nature of the curve suggests multilevel switching is possible. The intersection of each curve with the ordinate axis defines the remnant polarization (magnetization) P r (M r).
(Color online) (A) Energy level diagram of ferroelectric tunnel junction. (From E. Y. Tsymbal and H. Kohlstedt, Science 313, 181 (2006). Reprinted with permission from AAAS.) (B) Tunneling barrier height dependence on remnant polarization direction. (Reprinted figure with permission from M. Y. Zhuravlev, R. F. Sabirianov, S. S. Jaswal, and E. Y. Tsymbal, Phys. Rev. Lett. 94, 246802 (2005). © 2005 American Physical Society.) Note that polarization directions reversed with respect to reference as per the erratum (Phys. Rev. Lett. 102, 169901(E) (2009)).
(Color online) Principles of tunneling magnetoresistance, (A) and (B), and of spin-transfer torque switching, (C) and (D). (A) In TMR, when magnetizations parallel, electrons of one spin transmitted and opposite spin scattered. (B) When magnetizations antiparallel, electrons of both spins scattered in different layers. At high current densities, STT switching can occur. (C) When unpolarized electrons flow from fixed to free layer, electrons become polarized by fixed layer and transfer spin angular momentum to free layer, aligning spins parallel. (D) When unpolarized electrons flow from free layer to fixed layer, reflected electrons with spin antiparallel to fixed layer transfer angular momentum, aligning spins antiparallel.
(Color online) Circuit implementation and operation of standard magnetic tunnel junction. Current through two wires, not through MTJ, generates sufficiently large magnetic field for switching only at intersection. Resistance state is then measured through MTJ with separate contacts. (Used with permission from T. M. Maffitt, J. K. DeBrosse, J. A. Gabric, E. T. Gow, M. C. Lamorey, J. S. Parenteau, D. R. Willmott, M. A. Wood, and W. J. Gallagher, IBM J. Res. Dev. 50, 25 (2006).)
(Color online) Selected theorized adaptive electronic devices. (A) Conventional CMOS- (left) and memristor-based (right) content addressable memory. Memristor design requires less area, no V DD line, and less power. (K. Eshraghian, K. R. Cho, O. Kavehei, S. K. Kang, D. Abbott, and S. M. S. Kang, IEEE Trans. Very Large Scale Integr. (VLSI) Syst. 19, 1407 (2011). © 2011 IEEE.) (B) Associative learning. In initial “probing” phase, output only fires when Input 1 fires. In “learning” phase, weight of synapse S 2 is adjusted such that, in final “probing” phase, output fires when either Input 1 or Input 2 fires. (Reprinted from Neural Networks, Y. V. Pershin and M. Di Ventra, Experimental demonstration of associative memory with memristive neural networks, 23, 881, © 2010, with permission from Elsevier.) (C) Neuromorphic circuit with simulated resistive switches as synapses for detecting human figure orientation. Neural network has > 86% accuracy. (J. A. Pérez-Carrasco, C. Zamarreño-Ramos, T. Serrano-Gotarredona, and B. Linares-Barranco, in Proceedings of 2010 IEEE International Symposium on Circuits and Systems (ISCAS), 2010, p. 1659. © 2010 IEEE.)
Output waveforms of theorized adaptive electronic devices utilizing idealized TiO2 memristive model. (A) Passive tunable electromagnetic wave switch. Initial and final excitations at high frequency show high reflectivity. When low frequency excitation is superimposed, resistance state of switch responds such that reflectivity is significantly reduced. (Reprinted with permission from M. G. Bray and D. H. Werner, Appl. Phys. Lett. 96, 073504 (2010). © 2010 American Institute of Physics.) (B) Programmable amplifier with continuously tunable gain. Again, a low frequency (0.5 Hz) excitation modifies the output voltage response of the amplifier. (T. A. Wey and W. D. Jemison, IET Circuits Devices Syst. 5, 59 (2011). With permission from IET.) (C) Input and output waveforms of least mean square adaptive filter. Input signal is a sinusoid corrupted by additive noise. Output initially is quite noisy, but filter quickly adapts with time such that clean sinusoid is well reproduced. (With kind permission from Springer Science + Business Media: Analog Integr. Circuits Signal Process., F. Merrikh-Bayat and S. Bagheri-Shouraki, Mixed analog-digital crossbar-based hardware implementation of sign–sign LMS adaptive filter, 66, 41, 2011, Figs. 9 and 10(b).)
Performance characteristics of theorized neuromorphic circuits utilizing adaptive components. (A) Spatial filter array for edge detection and pattern recognition. Edge detection schematically shown in top panel. Bar graphs in bottom panel show that network selectivity is only weakly dependent on the rate of defective devices. (G. S. Snider, Nanotechnology 18, 365202 (2007). With permission from IOP.) (B) Table comparing operating characteristics of a CMOS and memristor-MOS technology (MMOST) position detector arrays. The MMOST implementation requires less dynamic power and chip area. (I. Ebong and P. Mazumder, in 2010 International Conference on Microelectronics (ICM), 2010, p. 292. © 2010 IEEE.)
(Color online) Illustration of FPGA-like hybrid CMOS-memristor IC. Dashed lines are programmed memristor connections. Dots are tungsten vias for connections to CMOS logic layers. A variety of logic operations can be wired by switching different combinations of memristors. (Reprinted with permission from Q. Xia, W. Robinett, M. W. Cumbie, N. Banerjee, T. J. Cardinali, J. J. Yang, W. Wu, X. Li, W. M. Tong, D. B. Strukov, G. S. Snider, G. Medeiros-Ribeiro, and R. S. Williams, Nano Lett. 9, 3640 (2009). © 2009 American Chemical Society.)
(Color online) Switchable Pt/TiO x /Pt rectifier. Opposite polarity voltage pulses control location of oxygen vacancies, which determines which contact is rectifying and which is Ohmic. (Reprinted with permission from H. Shima, N. Zhong, and H. Akinaga, Appl. Phys. Lett. 94, 082905 (2009). © 2009 American Insitute of Physics.)
(Color online) Demonstration of weighted sum operation in GdO x /Cu:MoO x bipolar resistive switch crossbar array. Adjacent cells are electrically programmed with a 1:2:4 (Cell A:B:C) resistance ratio, and input voltage is applied to each cell or combination of cells, as noted in the figure. The output varies depending on which cells are active, demonstrating a simple digital-to-analog convertor with adaptive oxide materials. (H. Choi, H. Jung, J. Lee, J. Yoon, J. Park, D.-J. Seong, W. Lee, M. Hasan, G.-Y. Jung, and H. Hwang, Nanotechnology 20, 345201 (2009). With permission from IOP.)
(Color online) Experimental demonstration of spike-timing dependent plasticity (STDP) in Pt/Cu2O/W device. (A) I-V curves of MIM device showing bipolar resistive switching. (B) For Δt > 0 (pre-synaptic pulse fires before post-synaptic pulse), the synaptic weight Δw = (I after – I before)/I o increases, while for Δt < 0, the synaptic weight decreases, in accordance with STDP. (With kind permission from Springer Science + Business Media: Appl. Phys. A, S.-J. Choi, G.-B. Kim, K. Lee, K.-H. Kim, W.-Y. Yang, S. Cho, H.-J. Bae, D.-S. Seo, S.-I. Kim, and K.-J. Lee, Synaptic behaviors of a single metal–oxide–metal resistive device, 102, 1019, 2011, Figs. 2(a), 3(d).)
(Color online) Adjustment of on/off threshold voltage of VO2 electrically-driven metal-insulator transition (MIT) in VO2-SmNiO3 heterostructure. VO2 MIT manifests as abrupt current jumps in I-V (State 1). Threshold voltage of MIT shifted by application of large voltage (e.g., line labeled for State 2 to 3), which increases resistance state of SmNiO3 underlayer in a non-volatile manner. SmNiO3 resistance increase redistributes electric field in heterostructure such that larger voltage is needed to trigger VO2 MIT. States 2 and 3 are I-V after such large voltage is applied. (Reprinted with permission from S. D. Ha, G. H. Aydogdu, B. Viswanath, and S. Ramanathan, J. Appl. Phys. 110, 026110 (2011). © 2011 American Institute of Physics.)
(Color online) Transfer function of adaptive RLC filter with VO2 memristor as resistor component as-fabricated (cross marks). After off-resonance pulse train, there is no change in filter output (circles). After on-resonance pulse train, there is enhanced Q factor (triangles). (Reprinted with permission from T. Driscoll, J. Quinn, S. Klein, H. T. Kim, B. J. Kim, Y. V. Pershin, M. Di Ventra, and D. N. Basov, Appl. Phys. Lett. 97, 093502 (2010). © 2011 American Institute of Physics.)
List of functional oxides used in redox resistive switching devices that may be suitable for adaptive electronics applications. Relevant device properties are specified in the header. Dashes denote that data was not specified in publication. Endurance and retention values do not necessarily reflect device failure limit, only the extent to which respective devices were tested. TE = top electrode and BE = bottom electrode.
List of functional oxides used in ferroelectric and ferromagnetic devices that may be suitable for adaptive electronics applications. Dashes denote that data was not specified in publication. Endurance and retention values do not necessarily reflect device failure limit, only the extent to which respective devices were tested. Internal state range of ferroelectric capacitors is defined as 2|P r|.
Article metrics loading...
Full text loading...