Design and Optimization of TMOReRAM Based Synaptic Devices J.F. Kang1#, B. Gao1, P. Huang1, Z. Chen1, L.F. Liu1, X.Y. Liu1, S.M. Yu2, H.-S. P. Wong3 #E-mail: 1Institute 2School
3Department
[email protected]
of Microelectronics, Peking University of CIDSE, Arizona State University
of Electrical Engineering, Stanford University
2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
1
Outline Introduction Physical
Mechanism Defect Engineering Approach Optimization of Synapse Summary
2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
2
Outline Introduction Physical
Mechanism Defect Engineering Approach Optimization of Synapse Summary
2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
3
Introduction Resistive Switching (RS)
• Many materials have been used to demonstrate the reversible bi-stable resistance states (LRS and HRS), which can be switched by voltage, named as resistive
switching (RS)
RRAM • These RS materials can be used to construct a device, with a typical sandwiched structure, termed as RRAM (Resistive-switching Random Access Memory).
2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
4
Introduction
Two Switching Modes [#] Unipolar
Bipolar
depend on amplitude of applied voltage but not on polarity
depend on the polarity of the applied voltage
[#] H.-S. P. Wong et al., Proc. IEEE, 100, p.1951, 2012. 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
5
Introduction
Excellent performances have been demonstrated in transition metal oxide (TMO)-ReRAM [1-6]. Scalability: <10nm devices demonstrated [1-2]
Compatibility with CMOS using fab-friendly
materials [1-4] • HfO2, TaOx, WOx, Ti, Ta, TiN, NiSi
Switching speed: <1ns [6] Switching voltage: <1.5V Endurance: >1010 cycles [5]
Read disturb:
2015 SPICE Workshop
K-S Li et al, VLSI-T2014, C-W. Hsu et al, IEDM2013 W. Chien et al, IEDM2010. X.A. Tran et al, IEDM2011
[5] H.Y. Lee et al, IEDM2010.
Retention: >10 yrs [6]
>1010
[1] [2] [3] [4]
times [3]
[6] Y.S.Chen et al, IEDM2009
June 29-July 3 2015
Mainz, Germany
6
Introduction
Capability to High Density Integration [1,2]
Vertical MOSFET
32/16 Gb Test Chips have been demonstrated [3,4] [1] [2] [3] [4]
H.-Y. Chen, et al., IEDM2012, p.497 (Stanford & PKU); ITRS 2013, http://www.itrs.net, PIDS Chapter T-Y Liu et al, ISSCC2013, p493 (Sandisk & Toshiba) R. Fackenthail, et al, ISSCC2014, p338 (Micron & Sony)
2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
7
Introduction
New Function Application Concept of RRAM based memristor [1] Memristive switches: both store logic values and perform logic operations [2]
[1] D. B. Strukov et al, Nature 2008, 453, p.80 [2] J. Borghetti, Nature 2010, 464, p.873-876 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
8
Introduction RRAM based synapses for neuromorphic computing systems [#]
[#] S.M. Yu et al, IEDM2012, p.239 (Stanford and PKU)
Most demonstrated in the bipolar switching mode 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
9
Introduction
For applications Understand the physical mechnisms of RS Seek technical solutions to construct RRAM devices to achieve targeted performances [1]
In this talk, we will also address Low energy and robust synapse performances of TMO-RRAM [2, 3] Potential for application in a neuromorphic visual system [2] …..
[1] B. Gao, et al. IEEE T-ED, 60(4), pp 1379, 2013; [2] S. Yu, et al. IEDM 2012, p.239; [3] B. Gao, et al, ACS Nano, 8, p. 6998, 2014
2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
10
Outline Introduction Physical
Mechanism Defect Engineering Approach Optimization of Synapse Summary
2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
11
Physical Mechanism
For the resistive switching (RS) behavior of TMO-RRAM Filament effect has been widely accepted RS is due to the formation and rupture of conducting filaments
R. Waser, Nature. Mat. 2007
However, the physical natures of filaments and the crucial effects to dominate the formation and rupture of filaments are still argued • Conducting filament (CF) type: Vo or metallic ions? • Dominant effect for SET/RESET: G-R or S/D? Thermal or E-field? • Mechanisms of unipolar and bipolar: Same or not? 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
12
Physical Mechanism
A Unified Physical Mechanism [1,2] To clarify fundamental properties of resistive switching behaviors in TMO-ReRAM [1] N. Xu et al, VLSI-T 2008, p.100 [2] B. Gao et al, IEDM2011, p.417
The mechanism is based on filament effect on RS [3] [3] R. Waser, Nature. Mat. 2007
2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
13
Unified Physical Mechanism The unified physical mechanism is proposed to clarify these argued issues: Microscopic physical properties correlated with resistive switching in TMO-based RRAM (including unipolar and bipolar) To explain various resistive switching
characteristics observed in TMO-RRAM To predict performances of TMO-RRAM
2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
14
Unified Physical Mechanism Schematic microscopic properties of RS in TMO-RRAM (B. Gao et al, IEDM2011, p.417) Oxygen
Oxygen vacancy
1. Filament: A percolation path consisting of VO defects 2. Formation and rupture of filaments are correlated with generation and recombination of VO 3. Forming/SET: Generation of new Vo defects and O2- ions induced by Efield and thermal effects in rupture region Vo defects may be in different states: Filled state (VO) with 2 electrons in Vo Unfilled state (Vo2+) w/o electron in Vo
2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
15
Unified Physical Mechanism Schematic microscopic properties of RS in TMO-RRAM 4. RESET: Recombination among charged V02+ and O25. Two essential conditions for RESET 1) Occurrence of V02+ states induced by a critical E-field 2) Presence of moveable O2Formation of the state V02+ in the filament at a critical E-field significant capture section stable recombination state (LO) 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
16
Unified Physical Mechanism 6. Conduction Properties: due to electron transport along Vo filaments Hopping conduction
• Semiconductor-like: Vo are separated from each other • Metallic-like: Vo are closed each other in the clustered
Metallic conduction
2015 SPICE Workshop
• First principle calculations support this opinion
June 29-July 3 2015
Mainz, Germany
17
Outline Introduction Physical
Mechanism Defect Engineering Approach Optimization of Synapse Summary
2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
18
The resistive switching characteristics are correlated with geometry of Vo filament generation, recombination, and distributions of Vo
It is crucial to control Vo distributions and filament geometry to achieve targeted performances
2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
19
Defect Engineering Approach According to crystal defect theory, the generation and recombination probability of Vo is governed by 𝜸𝑬𝒍𝒐𝒄 − 𝜺𝒂 Eloc: Local electric field 𝒑 = 𝒆𝒙𝒑( ) 𝒌𝑻 A Defect Engineering Approach is proposed [*]
a: Formation energy of Vo
[*] B. Gao et al, IEEE Tran. ED,.Vol 60, p.1379, 2013 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
20
Defect Engineering Approach
A Defect Engineering Approach is proposed A. Material-Oriented Cell Design
B. Innovation Operation Scheme 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
21
Defect Engineering Approach
A. Material-Oriented Cell Design Calculated formation energy a of Vo [1, 2] Undoped (eV)
Ti (eV)
Al (eV)
La (eV)
Ga (eV)
HfO2
6.53/6.40a
6.48
4.09
3.42
ZrO2
6.37/6.09b
6.11
3.66
3.74
3.77
a) A. S. Foster et al. PRB 65, 174117(2002) ; b) A. S. Foster et al. PRB 64, 224108(2001) ; c) T. R. Paudel et al. PRB 77, 205202(2008)
Trivalent La or Al doping could effectively reduce a [1] H.W. Zhang et al, APL 96, 2010 2015 SPICE Workshop
[2] B. Gao et al, VLSI2009
June 29-July 3 2015
Mainz, Germany
22
Defect Engineering Approach
A. Material-Oriented Cell Design
In the resistive switching (RS) layers of Al- or Ladoped HfO2 or ZrO2 [1-2] Vo are preferentially generated near the trivalent Al or La sites Filaments are preferentially formed along the dopant sites Better controllability of resistive switching could be achieved by using proper doping approaches
[1] H.W. Zhang et al, APL 96, 2010 2015 SPICE Workshop
[2] B. Gao et al, VLSI2009
June 29-July 3 2015
Mainz, Germany
23
Defect Engineering Approach
A. Material-Oriented Cell Design: Doping Effect
Vo distributions and CFs are full-randomly 2015 SPICE Workshop
Vo and CFs are formed near the dopant sites June 29-July 3 2015
Mainz, Germany
24
Defect Engineering Approach
99.5 98 90 Undoped 70 50 Doped 30 10 LRS
2 0.5
10
2
3
HRS 4
10 10 Resistance ()
10
5
Cumulative Probability (%)
Cumulative Probability (%)
Improved Uniformity by proper doping 99.5 98
Undoped Doped
90 70 50 30 10 2 0.5
Set
Reset -2
-1
0
1
2
Voltage (V)
Expected uniformity improvement is identified by experiments 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
25
Defect Engineering Approach Gradual transitions both in SET and RESET -3
Sharp
10 10 10
Current (A)
Current (A)
10
-4
Set Process
-5
Gradual
Undoped Doped
-6
0.0
0.4
0.8
1.2
1.6
10
-3
10
-4
10
-5
10
-6
10
-7
10
-8
-3.0V -2.5V -2.0V -1.5V
-3.0 -2.5 -2.0 -1.5 -1.0 -0.5 0.0
Voltage (V)
Voltage (V)
• Better controllability on RS processes achieved in doped HfOx devices • This is beneficial for RRAM as a synapse 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
26
Defect Engineering Approach
B. Innovation Operation Scheme • Vo density is dependent on local electric field and switching time • Operation schemes (switching time and local electric field) can be used to control Vo distributions Different operation schemes can be expected to achieve different response characteristics!! 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
27
Defect Engineering Approach
B. Innovation Operation Scheme
Non linear resistance change as a function of pulses is observed when short pulses are applied. 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
28
Defect Engineering Approach B. Innovation operation scheme
Nearly linear resistance change with pulses is realized when wider pulses are applied. 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
29
Outline Introduction Physical
Mechanism Defect Engineering Approach Optimization of Synapse Summary
2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
30
TMO-based Synaptic Devices
Neuromorphic Visual Systems A great amounts of synapses are needed A typical CMOS-based binary synapse consisted of a 8T-SRAM cell [*] New synapse is needed
[*] SRAM based CMOS hardware (IBM, CICC 2011) 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
31
TMO-based Synaptic Devices TMO-RRAM-based synapse is promising
Analogy between biological Analogy between biological and and artificial RRAM synapse. RRAM based neural networks. 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
32
Optimization of Synapse
Artificial Visual System-based on RRAM and Winner-Take-All algorithm is constructed Integrated-andfire neuron circuit
1st layer: 32× 32 neurons; 2nd layer: 4× 4 neurons between 1st layer and 2nd layer: 16348 RRAM synapses 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
33
Synaptic Device Behavior Multi-level resistance states and ultra-low spike energy <1pJ are demonstrated [#] [#] S. Yu, et al, IEDM2012, p.239
Dependence of energy/spike on initial R for training process 2015 SPICE Workshop
Measured and fitted training process with pulse amplitudes.
June 29-July 3 2015
Mainz, Germany
34
Model of TMO-based Synapse A model is developed for the training process of TMO-RRAM synapse
• Resistance variation effect during training process • Model parameters can extracted from measured data. 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
35
TMO-based Synaptic Devices Resistance Evolution under 400 RESET pulses • In low resistance regime, fluctuation is smaller but suffers from high spike energy • In high resistance regime, low spike energy but larger fluctuation presented Larger fluctuation or variation may cause degradation of recognition accuracy of the neuromorphic systems 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
36
Simulated System Performances Training Images and Initial Conductance Map
2D Gaussian bar: Random center and random orientation
Before the training randomized around 20kΩ 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
37
Simulated System Performances Resistance Diverges and Orientation Map Emerges During the training
After the training 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
38
Optimized Synaptic Devices Can we realize synaptic performances with both low spike energy and high recognition accuracy ? ?
Geometric mean of more than 2 devices in parallel can significantly suppress the impact of intrinsic fluctuation effect 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
39
Optimized Synaptic Devices Optimized architecture of a neuromorphic system using robust synapse is proposed • A 1D1R synaptic cell is introduced • 1D is applied to perform logarithm function on the device resistance Geometric mean calculation on resistance is replaced by the logarithm function. 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
40
Simulated System Accuracy • Single RRAM device • Geometric mean of two devices • Two parallel 1D-1R cells
• Significant improvement on recognition accuracy is achieved by the architecture of 2 parallel 1D1R. • Array integration approach is a great challenge 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
41
Optimized Synaptic Devices
3D vertical ReRAM array architecture as synapses Easily to achieve high density of integration Significantly to immunize resistance variation during training process of synapses
A synapse: devices in the same pillar electrode 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
42
3D Vertical RRAM Arrays
Measured training process of top and bottom ReRAM devices in the 3D vertical array 2 layered devices are fabricated Nearly constant device performance both in top and bottom layers is measured. Significantly improved accuracy achieved. 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
43
TMO-based Synaptic Devices
Measured training process for the 3D vertical synaptic devices Different initial R states can be achieved by different current compliances Initial R is set to ~1MΩ, maximum energy consumption per spike <1 pJ. 2015 SPICE Workshop
B. Gao et al, ACS Nano 8, 6998, 2014
June 29-July 3 2015
Mainz, Germany
44
Outline Introduction Physical
Mechanism Defect Engineering Approach Optimization of Synapse Summary
2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
45
Summary A unified physical mechanism is proposed to elucidate the resistive switching of TMORRAM A defect engineering approach is developed to design and optimize RRAM performances
Excellent controllability on RS behaviors is demonstrated in optimized RRAM devices based on the defect engineering approach.
2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
46
Summary
Multi-level resistance states are realized in the optimized RRAM
Robust synaptic behaviors with sub-pJ energy per spike are realized in the optimized RRAM
Optimized architectures of TMO-RRAM synapse are proposed to improve system performances. 2015 SPICE Workshop
June 29-July 3 2015
Mainz, Germany
47