Neuromorphic Computing

VLSI 2016: Four-Layer 3D Vertical RRAM Integrated with FinFET as a Versatile Computing
Unit for Brain-Inspired Cognitive Information Processing – HS Philip Wong (H. Li et al.,
Stanford University)
 For the first time, a four-layer HfOx-based 3D vertical RRAM, the “tallest” one ever reported, is developed and integrated
with FinFET selector.
 Uniform memory performance across four layers is obtained (±0.8V switching, 10 6 endurance, 104s@125°C).
 SPICE simulations show that high drive current of pillar select transistors is required for high-rise 3D RRAM arrays.
 The four-layer 3D RRAM is a versatile computing unit for (a) brain-inspired computing and (b) in-memory computing.
 The 3D architecture with dense and balanced neuron-synapse connections provides 55% EDP savings and 74% VDD
reduction (enhanced robustness) compared with conventional 2D architecture.

mmhlabs.org 11/03/2021 2
Sci. Adv. 2016: Organic core-sheath nanowire artificial synapses with femtojoule energy
consumption (W. Xu et al., Pohang University of Science and Technology)
 Demonstration of organic nanowire (ONW)
synaptic transistors (STs) that emulate the
important working principles of a biological
synapse.

 The ONWs emulate the morphology of

nerve fibers. With a core-sheath–
structured ONW active channel and a well-
confined 300-nm channel length obtained
using ONW lithography, ~1.23 fJ per
synaptic event for individual ONW was
attained, which rivals that of biological
synapses.

mmhlabs.org 11/03/2021 3
VLSI 2016: Novel RRAM-enabled 1T1R synapse capable of low-power STDP via burst-mode
communication and real-time unsupervised machine learning (S. Ambrogio et al.,
Politecnico di Milano University)
 Demonstration of a new synapse based on one-transistor/one-resistor
(1T1R) structure performing spike-timing dependent plasticity (STDP)
with lower area than 2T1R.
 STDP functionality and robust learning are shown by both deterministic
and stochastic synapses.
 Demonstration of 105 reduction of power consumption thanks to burst
operation and 86% efficiency of MNIST handwritten-digit classification.

mmhlabs.org 11/03/2021 4
Nature Comm 2016: Neuromorphic device architectures with global connectivity through
electrolyte gating (P. Gkoupidenis et al., Ecole Nationale Superieure des Mines)

 Demonstration of the global control of an

array of organic devices based on
poly(3,4ethylenedioxythiophene):poly(styrene
sulf) that are immersed in an electrolyte, a
behaviour that resembles homeoplasticity
phenomena of the neural environment.

 Authors this effect to produce behaviour that

is reminiscent of the coupling between local
activity and global oscillations in the
biological neural networks.

 Authors show that the electrolyte establishes

complex connections between individual
devices, and leverage these connections to
implement coincidence detection.

mmhlabs.org 11/03/2021 5
Nature Elect 2016: Learning through ferroelectric domain dynamics in solid-state synapses(S.
Boyn et al., Universite Paris-Saclay)

 Authors established that STDP can be

harnessed from intrinsically
inhomogeneous polarization switching in
ferroelectric memristors.

 Combining time-dependent transport

measurements, ferroelectric domain
imaging, and effective Hamiltonian-
based atomistic MD simulations, authors
show that the ferroelectric switching
underlying resistive changes in these
devices can be described by a well-
established nucleation-limited model.

 Using this physical model, reliable

prediction of the conductance evolution
of ferroelectric synapses with varying
neural inputs can be achieved

mmhlabs.org 11/03/2021 6
Nature Mat 2017: Memristors with diffusive dynamics as synaptic emulators for neuromorphic
computing (Z. Wang et al., University of Massachusetts)

 Demonstration of a new class of

memristors as synaptic emulators
that function primarily on the
basis of diffusion (rather than
drift) dynamics.
 The microscopic nature of both
the threshold switching and
relaxation of the diffusive
memristor is revealed for the first
time by in situ HRTEM and
explained by nanoparticle
dynamics simulation.
 The Ag dynamics of the diffusive
memristors functionally resemble
the synaptic Ca2+ behaviour in
chemical synapses and lead to a
direct and natural emulation of
multiple synaptic functions for
both short term and long-term
plasticity

mmhlabs.org 11/03/2021 7
Nature Mat 2017: A non-volatile organic electrochemical device as a low-voltage artificial
synapse for neuromorphic computing (Y.V.D Burgt et al., Stanford University)

 Demonstration of demonstrate a new organic electronic device made with inexpensive and commercially
available plastic materials that behaves as an artificial synapse.
 Artificial synapse exhibits a large number of non-volatile and reproducible states (>500) and operates at very
low voltages.
 Artificial synapse switches with low energy density and we project that just ∼35 aJ is sufficient to switch a sub-
micron device, a number smaller than that of biological synapses

mmhlabs.org 11/03/2021 8
Nature Comm. 2017: Face classification using electronic synapses – HS Philip Wong (P. Yao et
al., Tsinghua University)

 A neuromorphic network is developed using bidirectional analogue 1024-cell-1T1R RRAM array.

 The optimized RRAM metal oxide stack (TiN/TaOx/HfAlyOx/TiN) exhibits gradual and continuous weight change.
 An integrated neuromorphic network hardware system is built and trained online for grey-scale face classification.
 Two schemes have been tested (with and without write verify), both achieved high recognition rate, but with a trade-off:
 “With write-verify” scheme shows a better approach providing 4.61x faster converging speed, 1.05x higher
recognition accuracy and 4.41 lower energy consumption
 “Without write-verify” scheme simplifies the operation to a great degree.
 The resistive switching memory cell can be scaled down to 10 nm, which provides around 10 11 synapses per cm2.

Typical I–V curve of a single 1T1R cell for a quasi-DC sweep,

the gate voltage is 1.8 and 8 V during SET and RESET
process, respectively. Inset is a transmission electron
microscope (TEM) image of the RRAM device.

mmhlabs.org 11/03/2021 9
 VLSI 2017: Novel ferroelectric FET based synapse for neuromorphic systems (H.
Mulaosmanovic et al., NaMLab gGmbH)

 First demonstration of a synapse based on a single ferroelectric FET (FeFET) integrated in a 28nm HKMG technology,
having hafnium oxide as the ferroelectric and a resistive element in series.
 The gradual and non-volatile ferroelectric switching is exploited to mimic the synaptic weight.
 Both the spike-timing dependent plasticity (STDP) and the signal transmission are demonstrated and the effect of the
spike properties and circuit design on STDP is discussed .

mmhlabs.org 11/03/2021 10
Nature 2017: Neuromorphic computing with nanoscale spintronic oscillators (J. Torrejon et al.,
Université Paris-Saclay)

 Experimental demonstration that a nanoscale spintronic oscillator (a magnetic tunnel junction) can be used to
achieve spoken-digit recognition with an accuracy similar to that of state-of-the-art neural networks.
 Determination of the regime of magnetization dynamics that leads to the greatest performance.

mmhlabs.org 11/03/2021 11
Sci. Adv. 2017: On-chip photonic synapse (Z. Cheng et al., University of Oxford)

 Demonstration of a hardware synapse, implemented entirely in the optical domain via a photonic integrated-circuit
approach.
 Using purely optical means brings the benefits of ultrafast operation speed, virtually unlimited bandwidth, and no
electrical interconnect power losses.
 The synapse uses phase-change materials combined with integrated silicon nitride waveguides.
 Crucially, the synaptic weight can be randomly set simply by varying the number of optical pulses sent down the
waveguide

mmhlabs.org 11/03/2021 12
Nature Comm 2017: Flexible three-dimensional artificial synapse networks with correlated
learning and trainable memory capability (C. Wu et al., Hanyang University)
a b c
 Demonstration of flexible three-dimensional
artificial chemical synapse networks, in
which two-terminal memristive devices,
namely, electronic synapses (e-synapses),
are connected by vertically stacking
crossbar electrodes.

 The e-synapses resemble the key features

of biological synapses: unilateral connection,
long-term potentiation/depression, a spike-
timingdependent plasticity learning rule,
paired-pulse facilitation, and ultralow-power
consumption.

 The three-dimensional artificial synapse

networks enable a direct emulation of
correlated learning and trainable memory
capability with strong tolerances to input
faults and variations

mmhlabs.org 11/03/2021 13
Sci. Adv. 2017: Memristive stochastic plasticity enables mimicking of neural synchrony:
Memristive circuit emulates an optical illusion (M. Ignatov et al., Christian-Albrechts-Universität
zu Kiel)
 Demonstration that memristive devices
allow a new degree of freedom for this
concept: a local memory that supports
transient connectivity patterns in oscillator
networks.

 On the basis of the probability distribution

of the resistance switching process of Ag-
doped titanium dioxide memristive
devices, a local plasticity model is
proposed, which causes an autonomous
phase and frequency locking in an
oscillator network.

 Authors show that the implemented

network allows achievement of the
transition from asynchronous to multiple
synchronous states

mmhlabs.org 11/03/2021 14
IEDM 2017: 2D Molybdenum Disulfide (MoS2) Transistors Driving RRAMs with 1T1R Configuration
– HS Philip Wong (R. Yang et al., Stanford University)

 Demonstration of the first 1-transistor-1-

resistor (1T1R) memory cell using the
atomically thin MoS2 FET and resistive
random access memory (RRAM).
 This 1T1R demonstration realizes a key
milestone for tight integration of memory with
logic in a monolithic 3D integrated chip.
 The monolayer MoS2 is grown by chemical
vapor deposition (CVD), suitable for wafer-
scale fabrication.
 The MoS2 FETs have ON-state current of
190 µA/µm at VD = 2.5 V, showing strong
driving capability for RRAM.
 Metal-oxide RRAMs are fabricated at low
process temperature, compatible with MoS 2
FET fabrication.
 1T1R measurements show higher
resistances, and less resistance and voltage
variation compared with measurements
using only the RRAM. .

mmhlabs.org 11/03/2021 15
 IEDM 2017: Investigation of Statistical Retention of Filamentary Analog RRAM for
Neuromophic Computing (M. Zhao et al., Tsinghua University)

 Analog retention behavior of filamentary RRAM array is investigated for the first time
 The conductance distributions of all the conductance levels follow normal distribution with the standard
deviations increasing linearly with the square root of time
 Physical mechanism for analog retention is elucidated based on defect theory and Brownian Motion;
 A compact model is proposed to predict the retention behavior of analog filamentary RRAM and to evaluate the
performance of neuromorphic system.

mmhlabs.org 11/03/2021 16
IEDM 2017: Modeling Disorder Effect of the Oxygen Vacancy Distribution in
Filamentary Analog RRAM for Neuromorphic Computing (B. Gao et al., Tsinghua University)

 Analog RRAM has been recently developed where multiple weak conductive filaments (CF) are created instead of an
abrupt one. This device can switch in nanoseconds with excellent retention and edurance. However, device physics
were still not understood well.
 In this work, the authors perform Kinetic Monte Carlo simulations to show that oxygen vacancies distribution dictates the
analog behavior, current noise and temperature coefficient.
 Authors propose an order parameter “Ov” to quantify the analog behavior
 Ov parameter should be around 0.25 to transition the abrupt CF into the analog switching mechanism

mmhlabs.org 11/03/2021 17
 IEDM 2017: Gate-tunable memristors from monolayer MoS2 (V. K. Sangwan et al.,
Northwestern University)

 The fabrication of gate-tunable memristors based on monolayer MoS2 grown by chemical vapor deposition (CVD).
 Channel consisting of polycrystalline MoS2 films with grain sizes of 3-5 µm.
 Device show switching ratios up to ~500, with the resistance in individual states being continuously gate-tunable by over
three orders of magnitude.
 Device physics: resistive switching results from dynamically varying threshold voltage and Schottky barrier heights,
whose underlying physical mechanism is vacancy migration and/or charge trapping.
 Top-gated devices achieve reversible tuning of threshold voltage.

mmhlabs.org 11/03/2021 18
 IEDM 2017: Room Temperature 2D Memristive Transistor with Optical Short Term
Plasticity (X. Xie et al., UC Santa Barbara)

 Demonstration of the first room

temperature light sensitive memristive
transistor that can serve as artificial retina
device and memristive receiver for optical-
electrical neuromorphic interface.
 1T-phase quantum dot superlattice is
created on 2H-phase monolayer single
crystal MoS2 back gated FET by focused
electron beam irradiation.
 Quantum for is shown to work as charge
traps that induce memristive resistance.
 Memristive resistance can be controlled
by applying gate bias and shows short
term plasticity to light stimulation.

mmhlabs.org 11/03/2021 19
 IEDM 2017: Time Dependent Variability in RRAM-based Analog Neuromorphic
System for Pattern Recognition (J. Kang et al., Peking University)

 Time dependent variability (TDV) is investigated in RRAMs and its interaction with RRAM based pattern recognition.
 Authors show that even when the circuits are well trained, TDV effect can introduce drops in recognition accuracy
during the operating conditions
 TDV impact increases when higher resistances are used for the circuit implementation.
 TDV impact can not be suppressed by scaling up with more synapses or increasing the response time
 Training algorithm needs to be chosen carefully to reduce TDV impact

mmhlabs.org 11/03/2021 20
IEDM 2017: Fast, Energy-Efficient, Robust, and Reproducible Mixed-Signal Neuromorphic
Classifier Based on Embedded NOR Flash Memory Technology (X. Guo et al., UC Santa Barbara)

  Designed, fabricated, and tested a prototype mixed-signal, 28×28-binary-input, 10-ouput, 3-ayer neuromorphic
network based on embedded nonvolatile floating-gate cell arrays redesigned from a commercial 180-nm NOR flash
memory.
 Each array performs a very fast and energy-efficient analog vector-by-matrix multiplication, which is the bottleneck for
signal propagation in neuromorphic networks.
 All functional components of the prototype circuit, including 2 synaptic arrays with 101,780 floating-gate synaptic
cells, 74 analog neurons, and the peripheral circuitry for weight adjustment and I/O operations, have a total area
below 1 mm2.
 Its testing on the MNIST benchmark set has shown a classification fidelity of 94.65%, close to the 96.2% obtained in
simulation. The classification of one pattern takes <1 s time and ~20 nJ energy – both numbers >10 3× better than
those of the 28-nm IBM TrueNorth digital chip for the same task at a similar fidelity.

mmhlabs.org 11/03/2021 21
 IEDM 2017: Device and circuit optimization of RRAM for Neuromorphic computing –
HS Philip Wong (H. Wu et al., Tsinghua University)

 A human face recognition task was demonstrated on a 1k-bit 1T1R array using an online training perceptron
network.
 The RRAM device structure and materials stack were optimized to achieve reliable bidirectional analog switching
behavior.
 A binarized-hidden-layer (BHL) circuit architecture is proposed to minimize the needs of A/D and D/A converters
between RRAM crossbars.
 Several RRAM non-ideal characteristics were carefully evaluated for handwritten digits’ recognition task with
proposed BHL architecture and modified neural network algorithm.

mmhlabs.org 11/03/2021 22
Sci. Adv. 2018: Ultralow power artificial synapses using nanotextured magnetic Josephson
junctions (M.L. Schneider et al., National Institute of Standards Technology)

 Demonstration of a new form of artificial synapse based on

dynamically reconfigurable superconducting Josephson
junctions with magnetic nanoclusters in the barrier.

 The spiking energy per pulse varies with the magnetic

configuration, but in our demonstration devices, the spiking
energy is always less than 1 aJ. This compares very favorably
with the roughly 10 fJ per synaptic event in the human brain.

 Each artificial synapse is composed of a Si barrier containing

Mn nanoclusters with superconducting Nb electrodes.

 The critical current of each synapse junction, which is

analogous to the synaptic weight, can be tuned using input
voltage spikes that change the spin alignment of Mn
nanoclusters. We demonstrate synaptic weight training with
electrical pulses as small as 3 aJ.

 The Josephson plasma frequencies of the devices, which

determine the dynamical time scales, all exceed 100 GHz.

mmhlabs.org 11/03/2021 23
Nature Materials 2018: SiGe epitaxial memory for neuromorphic computing with reproducible
high performance based on engineered dislocations (S. Choi et al., MIT)

 Demonstration of analog resistive

switching devices that possess desired
characteristics for neuromorphic
computing networks with minimal
performance variations using a single-
crystalline SiGe layer epitaxially grown on
Si as a switching medium.
 Such epitaxial random access memories
utilize threading dislocations in SiGe to
confine metal filaments in a defined, one-
dimensional channel.
 This confinement results in drastically
enhanced switching uniformity and long
retention/high endurance with a high
analog on/off ratio.
 Simulations using the MNIST handwritten
recognition data set prove that epitaxial
random access memories can operate
with an online learning accuracy of 95.1%.

mmhlabs.org 11/03/2021 24
Science 2018: A bioinspired flexible organic artificial afferent nerve – Zhenan Bao (Y. Kim et al.,
Stanford University)

 Authors used flexible organic electronics to

mimic the functions of a sensory nerve.

 Artificial afferent nerve collects pressure

information (1 to 80 kilopascals) from clusters
of pressure sensors, converts the pressure
information into action potentials (0 to 100
hertz) by using ring oscillators, and integrates
the action potentials from multiple ring
oscillators with a synaptic transistor.

 Biomimetic hierarchical structures can detect

movement of an object, combine
simultaneous pressure inputs, and distinguish
braille characters.

 Authors connected the artificial afferent nerve

to motor nerves to construct a hybrid
bioelectronic reflex arc to actuate muscles.

mmhlabs.org 11/03/2021 25
Nature Comm 2018: Neuromorphic computing with multi-memristive synapses (I. Boybat et al.,
IBM)

 Demonstration of a novel synaptic

architecture comprising multiple memristive
devices with non-ideal characteristics to
efficiently implement learning in neural
networks.
 This architecture is shown to overcome
several significant challenges that are
characteristic to nanoscale memristive
devices proposed for synaptic
implementation, such as the asymmetric
conductance response, limitations in
resolution and dynamic range, as well as
device-level variability.
 The architecture is applicable to a wide
range of neural networks and memristive
technologies and is crossbar-compatible.
 The high potential of the concept is
demonstrated experimentally in a large-
scale SNN performing unsupervised
learning.

mmhlabs.org 11/03/2021 26
 VLSI 2018: Neuromorphic Technology Based on Charge Storage Memory Devices (S.T. Lee
et al., Seoul University)

 Four synaptic devices are introduced for spiking neural networks (SNNs) and deep neural networks (DNNs).
 Unsupervised learning is successfully demonstrated by applying the STDP learning rule reflecting the LTP/LTD characteristics of
the fabricated TFT-type NOR flash memory cells.
 Gated Schottky diode (GSD) and vertical NAND flash cell are proposed as synaptic device for DNNs.
 Using matched simulation, higher learning accuracy with GSD and NAND synaptic devices are obtained compared to that with a
memristor-based synapse.
 Measured synaptic properties of the vertical NAND cells are reported for the first time.

mmhlabs.org 11/03/2021 27
 VLSI 2018: A Methodology to Improve Linearity of Analog RRAM for Neuromorphic
Computing (W. Wu et al., Tsinghua University)

 Work presents a novel methodology to improve the conductance

tuning linearity of the filamentary RRAM.
 An electro-thermal modulation layer is designed and introduced to
control the distribution of electric field and temperature in the
filament region.
 For the first time, a HfOx based RRAM is demonstrated with linear
analog SET, linear analog RESET, 50ns speed, 10× analog tuning
window, 100kΩ on-state resistance, and high temperature
retention for multilevel states.

mmhlabs.org 11/03/2021 28
Nature Elect. 2018: Electronic synapses made of layered two dimensional materials– HS Philip
Wong (Y. Shi et al., Soochow University)

 Authors show that multilayer hexagonal boron nitride (h-BN) can be used as
a resistive switching medium to fabricate high-performance electronic
synapses.

 The devices can operate in a volatile or non-volatile regime, enabling the
emulation of a range of synaptic-like behaviour, including both short- and
longterm plasticity.

 The behaviour results from a resistive switching mechanism in the h-BN

stack, based on the generation of boron vacancies that can be filled by
metallic ions from the adjacent electrodes.

 The power consumption in standby and per transition can reach as low as
0.1 fW and 600 pW, respectively, and with switching times reaching less than
10 ns, demonstrating their potential for use in energy-efficient brain-like
computing.

 The volatile and non-volatile nature of the RS has been confirmed at the
nanoscale via CAFM, demonstrating excellent potential for scalability.

mmhlabs.org 11/03/2021 29
Nature Comm 2018: Artificial optic-neural synapse for colored and color-mixed pattern
recognition– HS Philip Wong (S. Seo et al., Sungkyunkwan University)

 Demonstration of an optic-neural
synaptic device by implementing
synaptic and optical-sensing
functions together on h-BN/WSe2
heterostructure.
 This device mimics the colored and
color-mixed pattern recognition
capabilities of the human vision
system when arranged in an optic-
neural network.
 The synaptic device demonstrates
a close to linear weight update
trajectory while providing a large
number of stable conduction states
with less than 1% variation per
state.
 The device operates with low
voltage spikes of 0.3 V and
consumes only 66 fJ per spike.

mmhlabs.org 11/03/2021 30
 IEDM 2018: ECRAM as Scalable Synaptic Cell for High-Speed, Low-Power
Neuromorphic Computing (J. Tang et al., IBM)

 Demonstration of a nonvolatile Electro-Chemical Random-Access Memory (ECRAM) based on lithium (Li) ion
intercalation in tungsten oxide (WO3) for high-speed, lowpower neuromorphic computing.
 Symmetric and linear update on the channel conductance is achieved using gate current pulses, where up to 1000
discrete states with large dynamic range and good retention are demonstrated.
 MNIST simulation based on the experimental data shows an accuracy of 96%.
 For the first time, high-speed programming with pulse width down to 5 ns and device operation at scales down to 300×300
nm2 are shown.
 It is also verified that the conductance change scales linearly with pulse width, amplitude and charge, projecting an
ultralow switching energy ~1 fJ for 100×100 nm2 devices.

mmhlabs.org 11/03/2021 31
IEDM 2018: 3D Monolithic Stacked 1T1R cells using Monolayer MoS2 FET and hBN RRAM
Fabricated at Low (150ιC) Temperature- HS Philip Wong (C.H. Wang et al., Stanford University)

 Demonstration of 3D monolithically integrated two level stacked 1-transistor/1-resistor (1T1R) memory cells, using monolayer
MoS2 transistors and few-layer hBN RRAMs, fabricated at temperatures below 150 ºC.
 The stacking process is scalable to an arbitrarily large number of layers and on any substrate material without foreseeable
physical limitations.
 The 1T1R cells can be switched with programming current <130 μA and voltage < 1 V, close to typical CMOS logic voltages.
 These cells are promising for in-memory and neuromorphic computing because (1) the hBN RRAM has gradual set and reset
switching due to multiple weak-filaments formed along local defects and (2) the MoS 2 transistor has low off-current due to the
large band gap of monolayer MoS2 (Eg > 2 eV).
 Linearity of RRAM resistance change is well-controlled by the gate voltage of the transistor.

mmhlabs.org 11/03/2021 32
 IEDM 2018: A Surface Potential- and Physics- Based Compact Model for 2D Polycrystalline
MoS2 FET with Resistive Switching Behavior in Neuromorphic Computing (L. Wang et al., NUS)
 A surface potential- and physics based compact model for two
dimensional (2D) polycrystalline- molybdenum disulfide (MoS2) FETs
with resistive switching (RS) behavior is developed and verified by
experimental data.
 This model is incorporated with the theories of thermal activation
transport, grain boundary (GB) barrier and space charge limited
current (SCLC).
 Based on the GB induced disorders, the grain size, low temperature
and high electrical field dependent characteristics are studied.
 The predicted transfer and output characteristics have excellent
quantitative agreement with experimental results.
 Considering the hopping process induced defect- (i.e., sulfur
vacancy) redistribution, the GB (e.g., intersecting or bisecting GB)
dependent resistive switching behavior is physically investigated.
 Model is implemented to simulate the synaptic activity such as short-
term/long-term plasticity, which indicates the possibility of using 2D-
FETs for neuromorphic computing applications.

mmhlabs.org 11/03/2021 33
 IEDM 2018: Hybrid Structure of Silicon Nanocrystals and 2D WSe2 for Broadband
Optoelectronic Synaptic Devices (Z. Ni et al., Zhejiang University)

 Authors take advantage of the synergy of the strong broadband optical absorption of boron (B)-doped silicon nanocrystals (Si
NCs) and the efficient charge transport of two-dimensional (2D) WSe 2 to make synaptic devices based on the hybrid structure of
Si NCs and 2D WSe2.
 The Si-NC/WSe2 synaptic devices can be optically stimulated in a broad spectral region from the ultraviolet (UV) to near-
infrared (NIR), exhibiting important synaptic functionalities.
 The energy consumption of the Si-NC/WSe2 synaptic devices may be as low as ~ 75 fJ.

mmhlabs.org 11/03/2021 34
 IEDM 2018: High Performance 2D Perovskite/Graphene Optical Synapses as Artificial Eyes
(H. Tian et al., Tsinghua University)

 A novel concept of all distributed architecture using optical synapse has been proposed.
 An ultrasensitive artificial optical synapse based on a graphene/2D perovskite heterostructure shows very high photo-
responsivity up to 730 A/W and high stability up to 74 days.
 Moreover, our optical synapses has unique reconfigurable light-evoked excitatory/inhibitory functions, which is the key to enable
image recognition.
 The demonstration of an optical synapse array for direct pattern recognition shows an accuracy as high as 80%.

mmhlabs.org 11/03/2021 35
Nature Mat 2019: Ionic modulation and ionic coupling effects in MoS2 devices for
neuromorphic computing (X. Zhu et al., The University of Michigan, Ann Arbor)

 Demonstration of reversible modulation of

MoS2 films that is consistent with local 2H–
1T′ phase transitions by controlling the
migration of Li+ ions with an electric field,
where an increase/decrease in the local
Li+ ion concentration leads to the transition
between the 2H (semiconductor) and 1T′
(metal) phases.

 The resulting devices show excellent

memristive behaviour and can be directly
coupled with each other through local ionic
exchange, naturally leading to synaptic
competition and synaptic cooperation
effects observed in biology.

mmhlabs.org 11/03/2021 36
Nature Elect 2019: Ternary content-addressable memory with MoS2 transistors for massively
parallel data search – HS Philip Wong (R. Yang et al., Stanford University)

 Demonstration of 2T2R transition

metal dichalcogenide TCAM (TMD-
TCAM) cells made from CVD
monolayer MoS2 FETs and HfOx-
RRAMs.

 The TMD-TCAM cells exhibit a high

R-ratio comparable to SRAM-based
TCAMs, due to the low off-state
current of 1L MoS2 FETs and the
robust current control in the 1T1R
driving scheme.

 Authors showed that, when these

TCAM cells are integrated into a
TCAM array, a very large RBSM
can be achieved for a long word
length, allowing searches of a large
amount of data in parallel.

mmhlabs.org 11/03/2021 37