NbO2-based frequency storable coupled oscillators ...

15 downloads 0 Views 907KB Size Report
S. P. Levitan, Y. Fang, D. H. Dash, T. Shibata, D. E. Nikonov, and. G. I. Bourianoff ... N. Shukla, W. -T Tsai, M. Jerry, M. Barth, V. Narayanan, S. Datta,. “Ultra Low ...
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JEDS.2018.2793342, IEEE Journal of the Electron Devices Society

1

NbO2-based frequency storable coupled oscillators for associative memory application Donguk Lee, Euijun Cha, Jaehyuk Park, Changhyuck Sung, Kibong Moon, Solomon Amasalu Chekol, and Hyunsang Hwang1

(a)

(b)

VDD

AC measurement

RRAM1

RRAM2

Ccoupling IMT1

IMT2

C1

C2

Current [A]

1m

Ramping speed: 10s/V

Oscillation resistance range

100μ

10μ

(c)

Frequency storable OSC1

Frequency storable OSC2

100n 0.0

V

NbO2

1.0

1.5

Voltage [V]

1m

RESET

Linear resistance state

200μ 150μ

`

200uA C.C

100μ

100uA C.C

100μ

V 50μ

0.5

(d)

250μ

Vth

Vhold



Current [A]

Abstract—Oscillatory neural networks with nano-oscillators and synapse devices are a promising alternative to implement neuromorphic systems owing to its fast recognition speed and low power consumption. In this work, we demonstrate a compact frequency storable oscillator using nanoscale two-terminal NbO2 insulator-metal-transition (IMT) devices along with TaOx-based resistive switching memory (RRAM) devices. By controlling RRAM resistance, we realized a wide range of analog oscillation frequencies. The synchronization window of two coupled oscillators, which is a key parameter for determining pattern recognition, increases with the increasing coupling capacitance and decreasing RRAM resistance of the reference oscillator. The simple device structure (Metal-NbO2-Metal-TaOx-Metal), small device area (4F2), and frequency storability of NbO2-based coupled oscillator device show a strong potential for future integrated neuromorphic device application.

VDD

50uA C.C

TaOx

Current [A]



Vth, IMT SET

V

10μ NbO2



Oscillation voltage range

TaOx

0 -1.5

-1.0

-0.5

0.0

0.5

Voltage [V]

1.0

1.5

100n

-2

-1

0

1

2

Voltage [V]

I. INTRODUCTION

Fig. 1 (a) Schematic of the proposed coupled oscillators circuit consisting of two frequency storable oscillators and a coupled capacitor. (b) Current–voltage characteristic of the NbO2 IMT device. (c) Current–voltage characteristic of the TaOx-based RRAM device with linear multi-level resistances. (d) DC current– voltage characteristic of the frequency storable oscillator device consisting of NbO2 IMT and TaOx RRAM devices.

ynchronization phenomena of oscillators have been the focus of considerable interest in the field of mathematics, physics, and neurobiology. Specifically, many studies have reported that periodic firing and synchronization of neurons play an important role during the information processing in brain systems [1]. An oscillator-based neuromorphic system that simulates these biological characteristics in hardware has been reported to have advantages in terms of parallel processing and energy consumption [2]. In oscillator-based neuromornphic systems, the input information is stored as oscillator parameters such as phase or frequency. If the patterns stored in the oscillation parameters are similar to each other, synchronization between the oscillators occurs, and this is considered to be pattern recognition [3], [4]. Recently, associative memory characteristics of various coupled oscillator devices such as nanoscale spin-torque oscillator [5], VO2 insulator-metaltransition (IMT) device [6] - [8], NbO2 IMT device [9], [10] and TaOx-based volatile filamentary switching device [11] were reported. However, implementing associative memory hardware using coupled oscillators is complex since it requires additional circuitry to memorize the information of the oscillator [3], [4]. In addition to this, VO2 is not a practical

candidate material for oscillator neuron because of its intrinsic low transition temperature (~340K). Also, the threshold switching mechanism of the TaOx-based device has not yet been clearly understood, it needs additional research to guarantee device reliability and uniformity. In this paper, we demonstrate the nanoscale frequency storable oscillator using NbO2 with TaOx resistive switching memory (RRAM) device having multi-level linear conductance to store information as an oscillation frequency [12]. NbO2 has a relatively high transition temperature (~1080K), and lower leakage current due to its large band gap (1.0eV) compared to that of VO2 (0.67eV) [13]. Owing to the above reason, NbO2 has the following benefits in IMT based oscillator: 1) wide range temperature operation; 2) reduced power consumption [14]; and 3) increased oscillation frequency due to the reduced charging time [15]. Through the combination of IMT and RRAM devices, we presented a compact and frequency storable oscillator device with a wide range of frequency modulation as a potential candidate for associative memory application. Furthermore, we evaluated the synchronization window of two coupled oscillators for associative memory application.

Index Terms— insulator-metal-transition, system, oscillatory neuron, coupled oscillators

neuromorphic

S

1

This work was supported by the Pioneer Research Center Program through the National Research Foundation of Korea funded by the Ministry of Science, ICT & Future Planning (2012-0009460)

The authors are with the Department of Materials Science and Engineering, Pohang University of Science and Technology, Pohang, 790-784, Republic of Korea (e-mail: [email protected]).

2168-6734 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JEDS.2018.2793342, IEEE Journal of the Electron Devices Society

2

(a) Measurement Simulation Vth of IMT

VDD: 1.3 V RRRAM: 5 kΩ

1.5

Voltage [V]

Fig. 1(a) shows a schematic of the oscillation neuron circuit consisting of two frequency storable oscillators and a capacitor for capacitive coupling. Each oscillator consists of an IMT device and a RRAM device with multi-level conductance states to store analog input values. To configure the circuit as shown in Fig. 1(a), we developed an IMT and a RRAM device with a TiN/NbO2/W [16] and TiN/TaOx/Pt [17] structure, respectively. Both IMT and RRAM devices were fabricated using simple metal-insulator-metal (MIM) structure with an effective device diameter of a 100 nm. Fig 1(b) shows the current-voltage characteristics of the NbO2 IMT device measured by a triangular voltage pulse. Through the AC measurement, we extracted important parameters needed to evaluate the oscillation characteristics of the NbO2 IMT device such as the threshold voltage (Vth), hold voltage (Vhold), resistance in the insulating state (Roff), and resistance in the metallic state (Ron). The Vth and Vhold parameters determine the voltage window of oscillation. The Roff and Ron parameters, on the other hand, represent the allowable range of the external resistor for oscillation. The NbO2 IMT device transforms from the insulating state to the metallic state when the applied voltage exceeds Vth. In contrast, if the applied voltage is smaller than Vhold, the NbO2 IMT device goes back to the initial insulating state. Considering the thickness of the NbO2 layer (~30 nm), it was confirmed that an electric field of 0.4 MV/cm is required for the insulator-metal-transition at room temperature. Fig 1(c) shows the current-voltage characteristics of TaOx-based RRAM device. By adjusting the compliance current during SET operation (from high resistance state to low resistance state), we confirmed the multi-level resistance characteristics of the RRAM device. The RRAM device also exhibited linear resistance regardless of the compliance current. After evaluation of the electrical characteristics of the IMT and RRAM devices, we investigated the DC I–V characteristics of the proposed frequency storable oscillator by connecting the IMT device and RRAM device as shown in Fig. 1(d). The oscillator shows the same SET and RESET switching as RRAM, and the current suppressed region at low voltage. By controlling the compliance current during SET or the voltage amplitude during RESET, we confirmed the modulation of the resistance states of the RRAM in the oscillator. The oscillation characteristics of the frequency storable oscillator device were evaluated through experiments and SPICE simulations, as shown in Fig 2(a). For the SPICE simulation, we used experimentally obtained IMT parameters and a parasitic capacitance value (C1,2 in Fig. 1) of 50 pF. The result of the oscillation obtained from the electrical measurement was in a good agreement with the SPICE simulation result. In addition, the linear conductance of the RRAM was helpful in predicting the oscillation characteristics depending on the voltage. The oscillation behavior is the consequence of a repetitive change in the voltage drop between the IMT and RRAM device. When supply voltage VDD (>Vth) is applied to the oscillator, most of the voltage is dropped on the IMT device, taking into consideration the relative resistance difference (Roff of the IMT device > Resistance of RRAM), and

1.0

Vhold of IMT

0

(b) Oscillation frequency [MHz]

II. RESULTS AND DISCUSSION

100

1

2

3

4

Time [s]

RRAM resistance [] 2k 3.3k 5k

1 kΩ 2 kΩ 3.3 kΩ 5 kΩ

10

1

30 kΩ

X30 0.1

5

1.2

1.4

1.6

1.8

2.0

Supply voltage [V]

Fig. 2 (a) Oscillation behavior of the proposed frequency storable oscillator device. (Voltage indicates the value applied only to the IMT device) (b) Oscillation frequency dependent on the input voltage and resistance of RRAM. (Dot: experimental data, line: simulation data)

the IMT device begins to change from an insulating state to a metallic state while charging a capacitor connected in parallel. When the transition of the IMT device is completed, most of the voltage will dropped on the RRAM device because Ron of the IMT device is smaller than the resistance of RRAM. Following this, the IMT device quickly changes from a metallic state to an insulating state while discharging the capacitor. Through this process, the voltage applied to the oscillator can oscillate with a constant period, and the oscillation frequency is determined by the charging and discharging time [15]. Fig. 2(b) shows the oscillation frequency dependency on the resistance of RRAM device and the supply voltage. As shown in the figure, the oscillator frequency increases as the RRAM resistance decreases and supply voltage increases. Small RRAM resistance and high supply voltage reduce charging time [15], which in turn increases oscillation frequency. Compared to 1T1IMT based oscillators [7], [8], in which the oscillation frequency is tuned by controlling the gate voltage, our proposed frequency storable oscillator has advantates due to novolatility and tunability of RRAM resistance. The oscillator can store the desired oscillation frequency just by controlling the resistance of the RRAM. Stable oscillation characteristics are realized only in a specific resistance and voltage ranges as shown in Fig. 1(b) and Fig. 1(d), respectively. The available supply voltage of the oscillator with a stable frequency is determined in the following manner. When supply voltage to the oscillator is smaller than the threshold voltage of the IMT device, the oscillation behavior does not occur owing to the absence of a repetitive change of voltage drop of the IMT and the RRAM device. Further, when the supply voltage to the oscillator is larger than the SET voltage, oscillator does not show a stable oscillation behavior owing to the change of the resistance state of RRAM in the oscillator.

2168-6734 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JEDS.2018.2793342, IEEE Journal of the Electron Devices Society

3 Using the SPICE simulation, it was confirmed that the oscillator frequency can be modulated up to 30 times with varying RRAM resistance as shown in Fig. 2(b). Finally, we evaluated the synchronization characteristics of a coupled oscillator consisting of two oscillators and a capacitor by SPICE simulation. Fig. 3(a) shows the process of synchronizing two oscillators with different RRAM resistance. We observed that the synchronization of the coupled oscillators

(a) Voltage [V]

1.0

Syncing Locking freq. & phase

0.5

R1

R2

Cc

VOUT,1

0.0

0

1

(b)

2

VOUT,2

3

4

Time [s] R1: 5k. R2: 5.1k

5

when the coupling capacitance is 10 pF and the resistance of the reference RRAM is 2 kΩ, the synchronization of the two oscillators is realized when the resistance of the other RRAM is within 100% of the resistance difference (∆RRRAM < 4 kΩ). Therefore, the synchronization window of the coupled oscillator increases with increasing coupling capacitance and decreasing RRAM resistance of the reference oscillator. Synchronization phenomenon in the coupled oscillator can be used for associative memory application by using ‘degree of match’ in the voltage periodic orbits or common coupling node [4] - [8]. In case of pattern recognition application using associative memory, template patterns need to be stored for evaluating ‘degree of match’ between template patterns and input pattern. According to associative memory researches using coupled oscillator, it is necessary to have an additional memory to store the template pattern [3], [4]. Due to the nonvolatility and tunability of RRAM resistance, template patterns can be memorized in our proposed frequency storable oscillator for associative memory application.

R1: 5k, R2: 5.15k

III. CONCLUSION

Vout,2 [V]

1.2

In summary, we have demonstrated a compact frequency storable oscillator device comprising a two-terminal NbO2– based IMT and TaOx–based RRAM device. The proposed device showed nonvolatile frequency storing function and analog oscillation behavior with wide frequency tuning range by controlling the resistance state of RRAM. In addition, the synchronization phenomenon of the coupled oscillator system was analyzed. The synchronization window of the coupled oscillator was evaluated by coupling the capacitance and resistance of RRAM in the oscillator. Considering various advantages such as compact device structure, frequency storable, analog oscillatory behavior, and synchronization characteristics, the NbO2 based frequency storable oscillator has a strong potential for neuromorphic pattern recognition applications.

1.0

0.8

SPICE Simulation 0.8

1.0

(c) Sync. windowR [%]

1.2

Vout,1 [V]

150

Rreference [] 2k 3k 5k 10k 15k

100

50

0

SPICE Simulation 0

5

10

15

20

Coupling capacitance [pF]

Fig. 3 (a) Synchronization of coupled oscillators consisting of the two proposed oscillators with different RRAM resistances (R1: 5 kΩ, R2: 5.1 kΩ, C1,2: 50 pF, coupling capacitance: 1 pF). (b) Periodic orbits of the coupled oscillator in steady state (black line: synchronized, red line: not synchronized). (c) Synchronization window of the coupled oscillator depending on coupling capacitance and difference of RRAM resistance in oscillators. All figures are simulation result by SPICE.

occurs within a few cycles of locking the phase and frequency simultaneously. Comparing the voltage across the two ends of the coupled capacitor, we can confirm that the two oscillators synchronized as shown in Fig. 3(b). If the two oscillators fully synchronized, the periodic orbit of both voltages shows a constant trace regardless of time. Otherwise, the periodic orbit of the two voltages shows random orbit over time. We identified the parameters that affect synchronization in the coupled oscillator. Fig. 3(c) shows the synchronization window of the coupled oscillator with varying the coupling capacitance and the RRAM resistances in two oscillators. For example,

REFERENCES [1]

[2]

[3]

[4]

C. M. Gray, “Synchronous oscillations in neuronal systems: mechanisms and functions,” Journal of computational neuroscience, vol. 1, pp. 11–38, 1994. N. Shukla, A. Parihar, M. Cotter, M. Barth, X. Li, N. Chandramoorthy, et al., “Pairwise coupled hybrid vanadium dioxide-MOSFET (HVFET) oscillators for non-boolean associative computing,” in Electron Devices Meeting (IEDM), 2014 IEEE International, 2014, pp. 28.7. 1–28.7. 4. S. P. Levitan, Y. Fang, D. H. Dash, T. Shibata, D. E. Nikonov, and G. I. Bourianoff, “Non-Boolean associative architectures based on nano-oscillators,” in Cellular Nanoscale Networks and Their Applications (CNNA), 2012 13th International Workshop on, 2012, pp. 1–6. T. Shibata, R. Zhang, S. P. Levitan, D. Nikonove and G. Bourianoff, “CMOS Supporting Circuitries for Nano-Oscillator-Based Associative Memories,” in Cellular Nanoscale Networks and Their Applications (CNNA), 2012 13th International Workshop on, 2012, pp. 29-31.

2168-6734 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JEDS.2018.2793342, IEEE Journal of the Electron Devices Society

4 [5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]

D. E. Nikonov, G. Csaba, W. Porod, T. Shibata, D. Voils, D. Hammerstrom, et al., “Coupled-oscillator associative memory array operation for pattern recognition,” IEEE Journal on Exploratory Solid-State Computational Devices and Circuits, vol. 1, pp. 85–93, 2015. N. Shukla, A. Parihar, E. Freeman, H. Paik, G. Stone, et al., “Synchronized charge oscillations in correlated electron systems,” Scientific Reports, vol. 4, p. 4964, 2014. A. Parihar, N. Shukla, S. Datta, and A. Raychowdhury, “Exploiting synchronization properties of correlated electron devices in a nonboolean computing fabric for template matching,” IEEE Journal on Emerging and Selected Topics in Circuits and Systems, vol. 4, pp. 450–459, 2014 N. Shukla, W. -T Tsai, M. Jerry, M. Barth, V. Narayanan, S. Datta, “Ultra Low Power Coupled Oscillator Array for Computer Vision Application,” in 2016 Symposium on VSLI Technology, 2016. S. Li, X. Liu, S.K. Nandi, D.K. Venkatachalam and R.G. Elliman, “Coupling dynamics of Nb/Nb2O5 relaxation oscillators,” Nanotechnology, vol. 28, no. 12, 125201, 2017. L. Gao, P.-Y. Chen and S. Yu, “Exploiting NbOx metal-insulatortransition device as oscillation neuron for neuro-inspired computing,” IEEE Electron Devices Technology and Manufacturing (EDTM), pp. 152-153, 2017. T. C. Jackson, A. A. Sharma, J. A. Bain, J. A. Weldon, and L. Pileggi, “Oscillatory neural networks based on TMO nano-oscillators and multi-level RRAM cells,” IEEE journal on Emerging and Selected Topics in Circuits and Systems, vol. 5, pp. 230–241, 2015. P.-Y. Chen, J.-S. Seo, Y. Cao and S. Yu, “Compact oscillation neuron exploiting metal-insulator-transition for neuromorphic computing,” IEEE/ACM International Conference on ComputerAided Design (ICCAD), pp. 1-6, 2016. A. B. Posadas, A. O’Hara, S. Rangan, R. A. Bartynski and A. Demkov, “Band gap of epitaxial in-plane-dimerized single-phase NbO2 films,” Applied Physics Letters, vol. 104, 2014. A. A. Sharma, Y. kesim, M. Shulaker, C. Kuo, C. Augustine, H. S. -P. Wong, S. Mitra, M. Skowronski, J. A. Bain and J. A. Weldon, “Low-Power, “High-performance S-NDR Oscillators for Stereo (3D) Vision using Directly-Coupled Oscillator Networks,” in 2016 Symposium on VSLI Technology, 2016. P. Maffezzoni, L. Daniel, N. Shukla, S. Datta, and A. Raychowdhury, “Modeling and simulation of vanadium dioxide relaxation oscillators,” IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 62, pp. 2207–2215, 2015. E. Cha, J. Park, J. Woo, D. Lee, A. Prakash, and H. Hwang, “Comprehensive scaling study of NbO2 insulator-metal-transition selector for cross point array application,” Applied Physics Letters, vol. 108, p. 153502, 2016. A. Prakash, D. Deleruyelle, J. Song, M. Bocquet, and H. Hwang, “Resistance controllability and variability improvement in a TaOxbased resistive memory for multilevel storage application,” Applied Physics Letters, vol. 106, p. 233104, 2015.

2168-6734 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.