You are on page 1of 7

www.advmat.

de
www.MaterialsViews.com

Reconfigurable Nonvolatile Logic Operations in Resistance


Communication

Switching Crossbar Array for Large-Scale Circuits


Peng Huang, Jinfeng Kang,* Yudi Zhao, Sijie Chen, Runze Han, Zheng Zhou, Zhe Chen,
Wenjia Ma, Mu Li, Lifeng Liu, and Xiaoyan Liu*

Conventional computer systems store information in memory functionality of circuit including RS devices, such as neuromor-
units and process information in central processing unit. This phic computing and stateful logic operations.[23] Based on the
leads to huge energy-hungry data transfer between memory highly nonlinear dynamics nature, implication logic operation
and computing units, which has been the limitation of compu- in RS devices was proposed in 2008.[16] ‘Stateful’ logic opera-
tation speed and energy efficiency—the so-called von Neumann tions were experimentally demonstrated in a row of RS devices
bottleneck.[1,2] To break the bottleneck, it is necessary to find via material implication in 2010.[10] RS devices are a simple
new materials, new devices, and new architectures to realize metal-insulator-metal sandwich structure, and their resistances
computing and memory functions.[3–11] Significant scientific can be switched between two digital states, low resistance state
progress has been made on resistance switching (RS) devices (LRS) and high resistance state (HRS), by two different appro-
in terms of speed, power, integration density, and reliability for priate voltage pulses.[24–26] Leveraging the conditional toggling
the application of memory.[12–15] Meanwhile, implication logic property of RS devices, logic operations can also be executed.
in RS devices was proposed[16] and demonstrated in a row of Therefore computing and memory units can be unified with
devices.[10,17] However, no viable method to construct large- RS devices, which may pave the way for breaking the bottleneck
scale logic circuits with RS devices has been presented. Here, of conventional computer systems.
we report a new computer unit built of RS crossbar array based Figure 1a shows a 16 × 8 RS crossbar array. Each junction
on resistance switching materials, where logic and memory between the 16 top wires (WL) and 8 bottom wires (BL) is an
functions are executed in the same devices. In the new com- RS device, whose structure is shown in the inset of Figure 1b.
puter unit, nonvolatile NAND and AND logic operations are The typical quasi-direct-current electrical characteristic of RS
executed by RS devices via resistance state interaction. By syn- devices is shown in Figure 1b. RS devices can be switched from
thesizing NAND and AND logic operations, resistance states HRS to LRS when the applied positive voltage across it is larger
stored in devices located in arbitrary positions of RS crossbar than program voltage VSET, and opposite operation is performed
array can perform various logic operations via logic cascading, by a negative voltage with magnitude larger than erase voltage
such as full adder. The functions can be reconfigured at run VRESET. Here, we define HRS as logical ‘1’ and LRS as logical
time simply by altering trigger signals. Leveraging the unity of ‘0’. Figure 1c shows the current response of an RS device under
logic and memory units, the huge energy-hungry data transfer voltage pulses with 100 ns width, where a +1.2 V voltage pulse
can be vastly mitigated in this new computer unit. The demand is applied to switch device to LRS from HRS, while a −1.8 V
for energy-efficient computer systems would be increasingly voltage pulse can switch it to HRS from LRS. Meanwhile, the
larger than ever due to the information revolutions in big data device can't be switched when the voltage pulse with magni-
processing, Internet of Things, and wearable devices. This tude smaller than VSET or VRESET. This conditional switching
reconfigurable nonvolatile logic operations in RS crossbar property is critical for the design of logic operation with RS
array offers a technology solution to support diverse societal devices. More basic electrical characteristics of our RS devices,
applications. such as SET voltage distribution and reliability behaviors, were
RS devices have been widely studied as one of the most measured and shown in Section S1 (Supporting Information).
promising candidates for future memory technologies by semi- NAND gate is a universal logic gate. Any binary logic func-
conductor research community since 2004[18–20] even though tion can be realized using a finite number of NAND gates.
RS phenomenon has been reported in 1960s.[21] In 2008, RS Figure 2a shows the subcircuit used to execute NAND opera-
devices were linked with memristor, which reveals that RS tion, where top electrodes of three RS devices are connected
devices are not just bistable switches, rather they are dynamic by a common WL and BL of output device Y is grounded. A
devices with highly nonlinear dynamics.[22] This extends the long and low voltage pulse (VR) is assigned to the BL of input
devices (A and B) and a relatively short and high voltage pulse
(VDD) is assigned to WL through a constant reference resistor
Dr. P. Huang, Prof. J. F. Kang, Y. D. Zhao, S. J. Chen, (RC) to trigger the NAND logic operation. Inputs are the resist-
R. Z. Han, Z. Zhou, Z. Chen, W. J. Ma,
M. Li, Prof. L. Liu, Prof. X. Y. Liu ance states of devices A and B, and the output logic state will be
Institute of Microelectronics stored into device Y after logic operation. The output device Y
Peking University is assigned to HRS before the logic operation. When the input
No.5 Yiheyuan Road Haidian District devices A and B are both in HRS (logical ‘1’), the potential of
Beijing 100871, P. R.China
E-mail: kangjf@pku.edu.cn; xyliu@ime.pku.edu.cn
common WL is approximate to VDD, and the output device Y
will be programmed to logical ‘0’ after logic operation. If any
DOI: 10.1002/adma.201602418 input device is in LRS (logical ‘0’), the potential of common WL

9758 wileyonlinelibrary.com © 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim Adv. Mater. 2016, 28, 9758–9764
www.advmat.de
www.MaterialsViews.com

charged by the voltage pulses VDD and VR

Communication
via different paths. Thus the transient poten-
tial of the common line may be larger than
the stable potential if they are assigned at
the same time, which would lead to incor-
rect switching of the output device and the
decrease of logic operations robustness. To
avoid it, VDD should be assigned later and
ended earlier than VR. When all input devices
are in HRS, RC should be selected as low as
possible compared to RH /2 in order to guar-
antee that the potential of common WL is
close to VDD (RH is the abbreviation of resist-
ance of HRS). RC should be selected as high
as possible compared to the resistance of LRS
(RL) to ensure that the potential of common
WL is close to VR when any input device is
in logical ‘0’. A further detailed discussion
about the selection of RC is presented in the
Section S2 (Supporting Information).
To synthesize other logic operations,
NAND is enough but it may cost many oper-
ation steps and RS devices. To simplify it,
another frequently used logic operation AND
is introduced. The circuit of AND operation
Figure 1.  RS crossbar array. a) Microscope micrograph of an integrated 16 × 8 array with a TiN/ is the same as the NAND operation except
HfOX/AlOY/Pt RS device at each crosspoint. b) Typical measured current voltage characteristics the exchange of VDD and VR as shown in
of an RS device. c) Transient current response under voltage pulse. Resistance of the device Figure 2b. So the potential of common WL is
can be switched between low resistance state (logical ‘0’) and high resistance state (logical ‘1’) opposite to the condition of NAND operation
with 100 ns width voltage pulses VP (programming to logical ‘0’) and VC (erasing to logical
under the same logic inputs. Therefore, the
‘1’). The amplitude of VP and VC is +1.2 and −1.8 V, respectively. State of the device cannot be
turned with voltage pulse when the amplitude is lower than VP or VC. The inset in b shows the logic operation results are contrary and the
device cross-section schematically. logic function is AND Y′ = AB. The states of
input devices are nondestructive in these two
logic operations, which is important to miti-
is close to VR thus the output device will be in logical ‘1’ after the gate data transfer.
logic operation. That means a NAND logic operation Y ′ = AB Various logic operations can be performed via resistance
is performed via the resistance state interaction between input state interaction among the common WL devices based on the
and output devices. The state interaction is triggered by the previously presented method. However, this method cannot
pulse signals VDD and VR. Output device can retain its state for be used in the common BL devices due to their different con-
more than 105 s under 150 °C temperature stress, which can figurations as shown in Figure 2a and Figure S3a (Supporting
be seen in Figure S1c (Supporting Information). Therefore, the Information). It will greatly decrease the computing capability
logic function is nonvolatile, which means static power dissipa- of RS crossbar array as well as hinder the design of large-scale
tion of this logic operation can be zero and the latches used in logic circuit with RS crossbar array. To overcome this obstacle,
conventional computer systems are not necessary. the method of logic executed in the common BL devices is
To correctly perform a NAND logic operation with RS introduced. The trigger signal VDD is directly applied to the
devices, the selections of VDD, VR, and RC are critical. To switch WL of output device Y and the WL of input devices A and B is
the device to LRS, the magnitude of VDD should be larger directly connected to VR as shown in Figure S3a (Supporting
than VSET due to the voltage drop across RC. Bias stress across Information). The common BL is connected to the ground
output device Y is about VR when any input device is in LRS. through RC. Y is initialized to HRS before logic operation. If
The output device Y should not be switched according to logic the devices for input are all in logical ‘1’, the electrical poten-
function of NAND, so VR should be selected as small as pos- tial of BL will be close to the ground. The voltage drop across
sible. However, VR should be selected as close as possible to output device is about VDD, which leads the transition of output
VDD to ensure unchanged logic inputs during the operation. device to LRS. If any one of the input devices is in logical ‘0’,
VR is chosen to VDD/2 based on synthetical consideration of the electrical potential of the common BL will be approximate
the robustness of this logic operation. Generally, there is a to VR. Therefore, the voltage drop across output device is about
certain variation of VSET,[27] which makes the choice of VDD VDD-VR and the output device remains in HRS. It implies that
and VR more delicate. Therefore, the selection of VDD and VR the logic operation NAND is executed under the control of
should base on the distribution of VSET, which is presented trigger signals with the resistance interaction between the input
in Figure S1a (Supporting Information). The common line is and output devices. The AND logic operation can be performed

Adv. Mater. 2016, 28, 9758–9764 © 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim wileyonlinelibrary.com 9759
www.advmat.de
www.MaterialsViews.com
Communication

Figure 2.  NAND and AND logic operations. a) sub-circuit of NAND and b) AND logic operations executed by the common WL devices. Logic inputs
are stored in devices A and B, and Y will store logic output depending on the states of the devices A and B. Voltage pulses VDD and VR are applied to
trigger the resistance interaction between the input and output devices. c) States of output after NAND and d) AND logic operations versus the state
combinations of inputs A and B. States are read by a small voltage pulse of 0.25 V and the measured results verify the feasibility of nonvolatile NAND
and AND logic operations in RS devices.

by changing the WL of input devices to be directly connected to demonstrations are only used to verify the correctness of opera-
the ground and applying VR to the common BL through RC as tions, which are not necessary in the practical logic operations.
shown in Figure S3b (Supporting Information). If the states of output device of NAND and AND logic opera-
The fabricated RS crossbar array was used to experimentally tions are logical ‘0’ before operation, the output states are always
demonstrate the NAND and AND logic operations. Logic inputs logical ‘0’ no matter what the logic inputs are. Hence, if the state
are initialized to desired states by applying appropriate positive or of Y is treated as an input, the logic functions of circuits shown
negative voltage pulse across devices. Then the resistance states in Figure 2a,b should be revised as Y ′ = ABY and Y′ = ABY,
are read by a small pulse voltage Vread (+0.25 V, 200 ns) to verify respectively. These two logic operations can be experimentally
the initialized results. To mitigate the impact of charging and demonstrated as shown in Section S4 (Supporting Informa-
discharging current on the read signal, long rise and fall time of tion). Even though the input states may be destructive, these
applied pulses are used. To perform the NAND and AND logic two logic operations are suitable for conditions that the input
operations, trigger pulse signals VDD (+1.4V, 100 ns) and VR state Y would not be used in the subsequent logic operations
(+0.7 V, 1 μs) are applied to corresponding nodes to stimulate anymore, such as the intermediate variable in compound logic
the resistance state interaction among devices. After that, Vread is operations. Up to now, we have achieved four basic logic opera-
applied to device Y to read out the results of logic operations. The tions: AB, AB, ABY and ABY. Other frequently used logic oper-
logic operations executed in the common WL devices were meas- ations such as INV, OR, transfer, and XOR can be executed by
ured and summarized as a truth table in Figure 2c,d. The meas- synthesizing the four basic logic operations and are experimen-
ured results demonstrate that the logic operations of Y ′ = AB tally demonstrated in the Section S5 (Supporting Information).
and Y ′ = AB are successfully executed. The measured VSET dis- It can be found that the circuits of NAND and AND logic opera-
tribution and disturbance behaviors of RS devices (Figure S1a,b tions, transfer, and INV logic operations (Figures S5 and S6 in
in Supporting Information) indicate that the logic operations the Supporting Information), and OR and XOR logic opera-
executed by RS devices are robust with the selected voltage tions (Figures S7 and S8 in the Supporting Information) are the
pulses. The demonstrations of NAND and AND logic opera- same except the applied voltage pulses, which means that the
tions executed in the common BL devices are presented in the logic functions of the circuits constructed by RS devices can be
Section S3 (Supporting Information). It should be noted that the reconfigured simply by altering the trigger signals without any
initialization and readout processes employed in the experimental other additional circuit change. This results from the fact that

9760 wileyonlinelibrary.com © 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim Adv. Mater. 2016, 28, 9758–9764
www.advmat.de
www.MaterialsViews.com

Communication

Figure 3.  Full adder. a) Array configuration of one-bit full adder circuits, where four assisted RS devices (AS1–AS4) are used to perform the logic func-
tions of full adder. b) Computation procedure of one-bit full adder. Here formula ‘NAND(A,B) →Y’ is used to represent that states of devices A and B
are executed the NAND logic operation and the logic output is stored into device Y. c) The time sequence of the applied trigger signals to achieve the
logic operation of full adder. d) Experimental demonstration of logic operation of full adder for all eight input combinations. The states of the input
devices before and output and assisted devices after operations are read out as resistance values shown by the grayscale maps.

the logic functions of circuits constructed by RS devices depend processes of OR and XOR logic operations, we can find that the
on not only the structure and layout of circuit but also the time four basic logic operations executed by RS devices are arranged
sequence of trigger signals. Moreover, from the computational along the time dimension, not along a spatial dimension.[28]

Adv. Mater. 2016, 28, 9758–9764 © 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim wileyonlinelibrary.com 9761
www.advmat.de
www.MaterialsViews.com

An example of logic function reconfiguration and further The logic operations were performed in ten sequential
Communication

detailed discussion about the logic function reconfiguration is steps with nine devices including the input, output, and four
shown in the Section S6 (Supporting Information). assistant devices (AS1–AS4) based on Formula (2) as shown in
Adder circuits play a key role in computer system and other Figure 3a. The computation procedure is shown in Figure 3b.
complex computing tasks, such as multiply operation, can be Open circuit was used as input logical ‘1’ to execute NAND (X,
executed using it as basic unit. Here the functions of one-bit 1) operation. The time sequence of triggered signals applied to
full adder are experimentally demonstrated by assembling the each node is shown in Figure 3c. Detailed descriptions about
logic operations presented above as shown in Figure 3. For one the computational process of one bit full adder are shown in
bit full adder, there are three inputs (addend A, summand B, video SV1 (Supporting Information). The experimental results
and carry-in Ci) and two outputs (summary S and carry-out Co), are shown in Figure 3d for all eight possible input combina-
the logic functions can be expressed as tions. The current responses of output devices and resistance
evolution of all nine devices during the computational process
S = A ⊕ B ⊕ Ci (1) were measured and shown in Section S7 (Supporting Informa-
C o = ( A ⊕ B)C i + AB tion) for the typical input combination (1 + 0 + 1). The meas-
ured results demonstrate that the functions of full adder are
This set of formulas can be disassembled to the combination successfully executed. To execute complicated compound logic
of four basic operations, which are written as the form below operation like full adder with four basic logic operations, the
first step is to establish the logic function equations. Then the
S = A ⊕ BCi ⋅ ( A ⊕ B)Ci
logic function equations are disassembled to the combination
Co = ( A ⊕ B)Ci AB (2) of the four basic logic operations by logic synthesis, which can
A ⊕ B = ABAB be accomplished by circuit design software. The final is the

Figure 4.  Logic operation in devices of different rows and columns. a) Location of input devices in RS crossbar array. b) Computation procedure of logic
operation Y ′ = AB + C in RS crossbar. c) Time sequence of pulse signals applied to the different nodes of RS crossbar array to perform Y ′ = AB + C
logic operation. d) Measured truth table. States are represented by the resistance values.

9762 wileyonlinelibrary.com © 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim Adv. Mater. 2016, 28, 9758–9764
www.advmat.de
www.MaterialsViews.com

design of time sequence of the applied trigger signals for this of output devices and resistance evolution of all seven devices

Communication
logic operation based on the disassembled equations. during the computational process Y ′ = 10 + 0 were measured
Devices located in arbitrary positions of RS crossbar array and shown in Section S8 (Supporting Information). To alle-
can execute the user-defined computing task based on the viate the impact of sneak current on the logic operations, all the
four basic logic operations presented above. Figure 4 shows an other cells that did not contribute to the actual circuit operation
example that computation ( Y ′ = AB + C ) was performed with were remaining in pristine state or HRS for the measurement.
three inputs of different rows and columns. Figure 4a sche- The WL and BL resistance would lead to inhomogeneous logic
matically shows the RS crossbar used to execute logic opera- operation in the RS crossbar due to the voltage drop on line
tion. The WL1R, WL2R … WL16R (BL1U, BL2U … BL8U) are the resistance. The measured BL and WL resistance of our RS array
ends of WL (BL) with connected RC, while WL1L, WL2L … WL16L is smaller than 50 Ω due to the small array size and wide line
(BL1D, BL2D … BL8D) represent the ends of WL (BL) without width (10 μm). Therefore, there is little impact of line resist-
connected RC. According to the subcircuits of logic operations ance on the homogeneity of logic operation (Figure S12, Sup-
executed by common WL devices and common BL devices, the porting Information). Further detailed discussion about the
voltage pulses are applied to WL via RC and directly applied to impact of line resistance on the homogeneity of logic operation
BL when performing logic operation by common WL devices, is presented in Section S9 (Supporting Information).
while the voltage pulses are directly applied to WL and applied With appropriate peripheral complementary metal oxide
to BL via RC when performing logic operation by common semiconductor (CMOS) control circuits, RS crossbar array can
BL devices. Therefore, these two groups of ends must be dis- be used to construct a unit which can realize the functions of
cerned. The computation procedure is shown in Figure 4b. information processing and data storage on the same physical
The time sequence of trigger signals applied to each node location. The circuit architecture of this new computer unit is
of RS crossbar array is shown in Figure 4c. Detailed illustra- shown in Figure 5. To guarantee the correctness of logic opera-
tions about the computational process are presented in video tion, reading, and writing processes, RS devices should be inte-
SV2 (Supporting Information). The measured truth table is grated with selective device to mitigate sneak current.[29,30] The
shown in Figure 4d, in which logic states are represented by amplitudes of voltage pulses should be increased accordingly
resistance values measured at 0.25 V. The current responses to offset the voltage drop on the selective device. The type of

Figure 5.  Circuit architecture of the new computer unit. The unit is constructed by RS crossbar array together with peripheral CMOS control circuit.
The RS crossbar array size is m × n and is divided into two parts B1 and B2 as shown. Cells in B1 are the assisted devices which are programmed to ‘1’
during the power off time, while cells in B2 are used to compute and store information. Peripheral CMOS control circuits include four demultiplexers,
two pulse generator modules, and write/read module. Pulse generator modules generate four paths pulse signals controlled by the clock and enable
signals. Demultiplexers deliver the pulse signals or write/read signals to the selected ports of BL/WL and assign read feedback signals to write/read
module, which are controlled by address and enable signals.

Adv. Mater. 2016, 28, 9758–9764 © 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim wileyonlinelibrary.com 9763
www.advmat.de
www.MaterialsViews.com

logic operation executed by RS crossbar is not defined, which [1] J. V. Neumann, IEEE Ann. Hist. Comput. 1993, 15, 27.
Communication

can be programmed according to the requirements of specific [2] S. Borkar, A. Chien, Commun. ACM 2011, 54, 67.
computing task at run-time by altering the trigger signals. In [3] A. Ney, C. Pampuch, R. Koch, K. H. Ploog, Nature 2003, 425,
this new computer unit, logic inputs are stored in input devices 485.
[4] D. A. Allwood, G. Xiong, C. C. Faulkner, D. Atkinson, D. Petit,
before operation and logic outputs are stored into output
R. P. Cowburn, Science 2005, 309, 1688.
devices during the logic operation process. So there is no addi- [5] H. Dery, P. Dalal, L. Cywinski, L. J. Sham, Nature 2007, 447, 573.
tional data transfer operation in this new computer unit except [6] M. D. Ventra, Y. V. Pershin, Nat. Phys. 2013, 9, 200.
the data input and output, which has the potential to increase [7] M. Cassinerio, N. Ciocchini, D. Ielmini, Adv. Mater. 2013, 25,
the computation energy-efficiency. 5975.
The logic operations discussed above only consider two inputs [8] P. A. Merolla, J. V. Arthur, R. A.-Icaza, A. S. Cassidy, J. Sawada,
logic operation. Actually, N inputs NAND and AND logic opera- F. Akopyan, B. L. Jackson, N. Imam, C. Guo, Y. Nakamura,
tions can also be executed by the same theory. Although bipolar B. Brezzo, I. Vo, S. K. Esser, R. Appuswamy, B. Taba, A. Amir,
metal oxide based RS devices are discussed in this study, it is M. D. Flickner, W. P. Risk, R. Manohar, D. S. Modha, Science 2014,
believed that the proposed method to execute logic operations 345, 668.
[9] M. Prezioso, F. Merrikh-Bayat, B. D. Hoskins, G. C. Adam,
and construct large-scale logic circuit is also suitable for unipolar
K. K. Likharev, D. B. Strukov, Nature 2015, 521, 61.
RS devices[25,26] and any other devices which have conditional tog- [10] J. Borghetti, G. S. Snider, P. J. Kuekes, J. J. Yang, D. R. Stewart,
gling property as shown in Figure 1c. The conventional computer R. S. Williams, Nature 2010, 464, 873.
system based on transistor with voltage as physical variable is [11] E. Linn, R. Rosezin, S. Tappertzhofen, U. Bottger, R. Waser, Nano-
beyond its ability to keep up with recent increasing need of infor- technology 2012, 23, 305205.
mation technology area. The nonvolatile logic operations in high [12] M.-J. Lee, C. B. Lee, D. Lee, S. R. Lee, M. Chang, J. H. Hur, Y.-B. Kim,
density integration RS crossbar array with resistance as physical C.-J. Kim, D. H. Seo, S. Seo, U-I. Chung, I.-K. Yoo, K. K. Lee, Nat.
variable offer a new technology solution. This technology may be Mater. 2011, 10, 625.
viable for new computer system applications such as the distrib- [13] H. Y. Lee, Y. S. Chen, P. S. Chen, P. Y. Gu, Y. Y. Hsu, S. M. Wang,
uted computation framework to offer an alternative architecture W. H. Liu, C. H. Tsai, S. S. Sheu, P. C. Chiang, W. P. Lin, C. H. Lin,
W. S. Chen, F. T. Chen, C. H. Lien, M.-J. Tsai, IEEE Int. Elec. Dev.
for computation (Section S10, Supporting Information).
Meet. Tech. Dig. 2010, 460.
[14] R. Fackenthal, M. Kitagawa, W. Otsuka, K. Prall, D. Mills, K. Tsutsui,
J. Javanifard, K. Tedrow, T. Tsushima, Y. Shibahara, G. Hush, IEEE
Experimental Section ISSCC Tech. Dig. 2014, 338.
Experimental Arrays: The RS crossbar arrays comprised two resistive [15] S. Yu, H.-Y. Chen, B. Gao, J. Kang, H.-S. P. Wong, ACS Nano 2013,
switching layers, HfOX and Al2O3, sandwiched between top electrodes 7, 2320.
(WL) TiN and bottom electrodes (BL) Pt. First, a 20 nm Ti adhesion [16] P. Kuekes, Memristor and Memristive Systems Symp. 2008.
layer and a 100 nm Pt bottom electrode were prepared on a 4 inch [17] H. Li, B. Gao, Z. Chen, Y. Zhao, P. Huang, H. Ye, L. Liu, X. Liu,
silicon substrate by electron beam evaporation. Then, 3 nm HfOX was J. Kang, Sci. Rep. 2015, 5, 13330.
deposited by atomic layer deposition (ALD). After that, 2 nm Al2O3 was [18] I. G. Baek, M. S. Lee, S. Seo, M. J. Lee, D. H. Seo, D.-S. Suh,
deposited by ALD. After a 40 nm TiN top electrode was deposited by J. C. Park, S. O. Park, H. S. Kim, I. K. Yoo, U.-I. Chung, I. T. Moon,
reactive sputtering in high vacuum and patterned with lithography, dry IEEE Int. Elec. Dev. Meet. Tech. Dig. 2004, 587.
etch was performed to form the square-shaped devices. Finally, 50 nm Al [19] R. Waser, M. Aono, Nat. Mater. 2007, 6, 833.
metallization was used to complete the device fabrication. [20] H. S. P. Wong, H.-Y. Lee, S. Yu, Y.-S. Chen, Y. Wu, P.-S. Chen, B. Lee,
Experimental Characterization: Device electrical measurements were F. T. Chen, M.-J. Tsai, Proc. IEEE 2012, 100, 1951.
carried out using a Cascade Probe Station with an Agilent B1500A [21] T. W. Hickmott, J. Appl. Phys. 1962, 33, 2669.
semiconductor parameter analyzer. Pulse Measurements were performed [22] D. B. Strukov, G. S. Snider, D. R. Stewart, R. S. Williams, Nature
with device measurements instruments together with an Agilent 81160A 2008, 453, 80.
pulse function generator and an Agilent Infiniium MSO9404A Series [23] J. J. Yang, D. R. Stewart, R. S. Williams, Nat. Nanotechnol. 2013, 8,
Oscilloscopes. Bit lines and Word lines of the array were connected
13.
through an Agilent 34980A switching matrix to the instrument, ground, a
[24] J. J. Yang, M. D. Pickett, X. Li, D. A. A. Ohlberg, D. R. Stewart,
high impedance or an external constant reference resistor (3.1 kΩ).
R. S. Williams, Nat. Nanotechnol. 2008, 3, 429.
[25] K. M. Kim, D. S. Jeong, C. S. Hwang, Nanotechnology 2011, 22,
Supporting Information 254002.
Supporting Information is available from the Wiley Online Library or [26] B. J. Choi, D. S. Jeong, S. K. Kim, C. Rohde, S. Choi, J. H. Oh,
from the author. H. J. Kim, C. S. Hwang, K. Szot, R. Waser, B. Reichenberg, S. Tiedke,
J. Appl. Phys. 2005, 98, 033715.
[27] J. H. Yoon, S. J. Song, I. H. Yoo, J. Y. Seok, K. J. Yoon, D. E. Kwon,
Acknowledgements T. H. Park, C. S. Hwang, Adv. Funct. Mater. 2014, 24,
This work was supported in part by NSFC (Grant Nos. 61421005 and 5086.
61334007), Beijing Municipal Science, and Technology Plan Projects. The [28] C. S. Hwang, Adv. Electron. Mater. 2015, 1, 1400056.
authors acknowledge Y. Y. Wang for guidance and fruitful discussions. [29] Y. Yang, J. Lee, S. Lee, C.-H. Liu, Z. Zhong, W. Lu, Adv. Mater. 2014,
26, 3693.
Received: May 6, 2016 [30] H. Yang, M. Li, W. He, Y. Jiang, K. G. Lim, W. Song, V. Y.-Q. Zhuo,
Revised: August 1, 2016 C. C. Tan, E. K. Chua, W. Wang, Y. Yang, R. Ji, in Proc. of Symp. On
Published online: September 22, 2016 VLSI Technology, IEEE, New York 2015, 130.

9764 wileyonlinelibrary.com © 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim Adv. Mater. 2016, 28, 9758–9764

You might also like