A Review of Resistor Switching Devices for Memory and Neuromorphic Computing Applications
Hamsavahini R1*, Saurabh Nagar2, Kaustab Ghosh3
1 Department of Electronics and Communication Engineering, BMSIT&M, Bangalore, India.
2 SENSE, VIT University, Vellore, Tamilnadu, India.
3 SENSE, VIT University, Chennai, Tamilnadu, India.
*Corresponding Author
Hamsavahini R,
Department of Electronics and Communication Engineering, BMSIT&M, Bangalore, India.
Tel: 0741633550
E-mail: hamsavahini@bmsit.in
Received: September 08, 2020; Accepted: September 26, 2020; Published: September 29, 2020
Citation: Hamsavahini R, Saurabh Nagar, Kaustab Ghosh. MA Review of Resistor Switching Devices for Memory and Neuromorphic Computing Applications. Int J Nano Stud Technol. 2020;9(1):135-139. DOI : dx.doi.org/10.19070/2167-8685-2000025
Copyright: Hamsavahini R© 2020. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution and reproduction in any medium, provided the original author and source are credited.
Abstract
High-density memories are required for various applications in the area of signal and image processing, Digital system
design, and neuralnetworks. Researchers are looking for an intermediate solution to fill the gap between DRAM and Flash
NAND in the memory hierarchy. The development of Resistive Switching Technologies (RRAM) proposes a potential
solution to this demand for fast, low cost, high density, and non-volatile memory. This paper discusses the schematic of
Resistive Switching (RS) devices, The Principle and operation of memristor crossbar array as memory, potential applications
of memristors as memory, Logic design applications, and neuromorphic computations.
2.Memristor Theory
3.Resistive Switching Memory Device Applications
4.Conclusion
5.References
Key Words
Melamine; Alloxan; Supramolecular System; Microwire; Crystal; Complex.
Introduction
Over the past decades, successive downscaling of the
complementary metal-oxide-semiconductor (CMOS) transistors
consistent with Moore’s law has contributed to the exponential
progress in computing and knowledge technology. The amount
of transistors per microprocessor chip has doubled about every
two years with increased clock speed or rate of execution of
instructions. The transistor size is approaching the elemental
physical limit of around 2–3 nm beyond which quantum tunneling
and quantum uncertainties will make transistors unreliable to
work in conventional circuits. Additionally, the performance gap
between processing and memory units has increased dramatically,
and therefore the data movement between these units within
the conventional Von-Neumann architecture starts to be the
foremost dominant factor for energy consumption and system
throughput [4, 5]. This problem, widely referred to as the Von-Neumann bottleneck, is going to be further exacerbated within
the data-intensive applications like machine learning tasks. to deal
with these issues, new materials, devices, and new computing
architectures are now extensively investigated to enrich and
possibly replace conventional CMOS devices and circuits. One
such device may be a non-volatile Resistive Switching (RS) device.
rather than using electrons and holes to store data in conventional
memory units, RS devices store data by reconfiguring the interior
ion (oxygen ion or metal cation) distribution in nanoscale solidstate
films.
A typical RS device stores data in the form of different values of
resistance, and has a simple two-terminal resistor-like structure
with a functional thin film sandwiched by two electrodes, as
schematically shown in Figure. 1. RS devices first received broad
interest as they effectively reduce the performance gap between
existing memory (i.e. static random-access memory (SRAM) and
dynamic random-access memory (DRAM)) and non-volatile
data storage solutions, and can potentially replace the above
memoriesdue to the fast switching speed, low power operation,
scalability and good reliability offered by RS devices [1].
Passive circuit theory can be described asa set of relationships
between electromagnetic quantities:
1) Voltage vis the change inmagnetic flux Φ for time t
2) Current iis the change in electric charge q for time t
3) Resistor Ris a linear relationship between voltage and current (dv = Rdi)
4) Capacitor Cis the linear relationship between voltage and electric charge (dq = Cdv)
5) Inductor Lis the linear relationship between magnetic flux Φ and current i (dΦ = Ldi)
Out of the six possible relationships, the sole two electromagnetic
quantities that there are not any pairings are magnetic flux and
charge. But, in 1971 Leon Chua [6] postulated that mathematically,
a fourth fundamental passive circuit element could exist, called a
memristor, that binds the charge q to the linkage flux Φ as shown
in Figure. 2 during which the electrical symbol of a memristor is
additionally indicated:
dΦ = Mdq ----- (1)
The memristoris defined in terms of a non-linear functional relationship between the flux linkage Φ(t) and the amount of electric charge that has flowed through the device q(t):
f(Φ(t),q(t)) = 0 ---- (2)
where Φ(t) and q(t)are time-domain integrals of memristor electric voltage v and electric current i, respectively:
and
The variable flux linkage Φ(t) is borrowed from the circuit characteristic of an inductor, and it does not represent a magnetic field here.In Eq. (2) the derivative of one respect to the other depends on the value of one or the other. The charge-dependent rate of change of flux with chargecan be used to describe the memristor function as:
Introducing Eq, (3) in Eq, (5), we have:
where RM(q) is the small-signal memristancedefined at the operating point. In this way, we have acharge-controlled or current-controlled memristor. Similarly, simplifying Eq, (4) in (5), we have:
In this way, we will obtain a flux-controlled or voltage-controlled memristor. The current-controlled memristor is often modelled as a classical resistor whose resistance is controlled by the timedomain integral of the present flowing through the memristor. Analogously, the voltage-controlled memristor behaves as a conductor whose conductance depends on the time domain integral of terminal voltage.
In both cases, need an electronically controlled resistor,conductor and an integrator. Table.1 covers all meaningful ratios of differentials of i, q, Φ, and v. Figure. 3 shows the current-voltage characteristics for memristor, with its pinched hysteresis loop [2].
With reference to Eq. (6), memristor function can also be written in the following way;
v(t)= M(q(t)) i(t) ----- (8)
The power consumption characteristic equation of a memristor is given by,
P(t) = V(t) I(t) = I2(t)M(q(t)) ----- (9)
The memristor stores the logical value in terms of its resistance,
in contrast to standard memories, which use a charge to represent
data. This resistance is controlled by applying a voltage across the
memristor. Memristors are often fabricated between two metals,
which act as the top and the bottom electrodes of a switching
dielectric material. Hence, memristors are often fabricated within
the metal layers as a part of a typical CMOS Back End of Line
(BEOL) process. Memristive memory generally utilizes a crossbar
structure, which enables a particularly dense memory array with a
memory cell area of 4F2, where F is the technology feature size.
Figure.4 shows one such design of a memristive memory crossbar
array. Voltage drivers, row/column decoders, and sense amplifiers
are used as a component of the peripheral circuit to support write
and read operations, similar to other memory technologies.
To perform a write operation, a write voltage Vwrite, above the threshold voltage (Von and Voff, which switches the memristor to LRS and HRS, respectively), is applied across the target memristor through the wordlines and bitlines. For a memristor with asymmetric switching characteristics (i.e.,Von = Voff ), two different write voltages are applied for writing logic 1 (i.e., VSET ) and 0 (i.e., VRESET ). Read operations are performed by applying a voltage Vread , with a magnitude less than the threshold voltage for switching, and measuring current passing through the device employing a sense amplifier [3].
Resistive Switching Memory Device Applications
RS devices are considered as one of the simplest applicants for
future non-volatile memory applications [7, 8]. Unlike chargebased
memories, like dynamic random access memory (DRAM)
and NAND non-volatile storage, which affect performance
degradation as the scaling limit is approached, non-charge based
memories including RS devices offer solutions to increase Moore’s
law. especially, ionic RS devices offer outstanding performance
specs including scalability, high switching speed, long retention
time, high endurance, large on/off ratio, and low power operation.
Memory applications uses the resistance states of RS devices
represent a touch (‘0’ or ‘1’) or multi bits (e.g. 2bits: ‘00’, ‘01’, ‘10’,
or ‘11’), which may be read by sensing current through the RS
device during the read operation.
Synaptic functions: In addition to memory applications, RS
devices are considered promising candidates for bio-inspired
computing and in-memory computing. When applied in
computing systems, the RS devices are often utilized in a crossbar
form to perform Vector matrix operation (VMM). During this
approach, the values within the matrix are stored as the analog
conductance values of the RS devices within the crossbar array.
The input vector is applied as voltage pulses with different pulse
amplitudes or different pulse widths to the rows of the crossbar.
The currents or charges collected at the columns of the crossbar
represent the resulting VMM outputs. As a result, the computeintensive
VMM operations are often obtained during a single
step, greatly improving the energy efficiency and throughput
beyond the restrictions of standard computing.In the nervous
system, a neuron can communicate with other neurons by
passing electrical or chemical signals through synapses [9]. Each
neuron is often connected with thousands of other neurons
with different connection strength, i.e. synaptic weight, which
determines how efficient the input spikes from one neuron (the pre-synaptic neuron) are often delivered to the receiving neuron
(the postsynaptic neuron). Additionally, the synaptic weight are
generally updated by spikes from the pre- and post-neurons,
allowing the system to achieve learning and form memory.
These structures and functions are implemented with RS device
networks, thus making it possible to realize highly-efficient bioinspired
computing hardware.
VMM with crossbar array: In deep learning algorithms, the
VMM operation (Multiply-Accumulate (MAC) operation) is
the main computing operation for training and inference but
is extremely resource expensive for conventional computing
systems to implement. To accelerate VMM efficiently, the graphics
processing unit has been extensively used to improve parallelism
by using 1,000 s of computing cores with a high-throughput
connection to the memory. Algorithm studies to more efficiently
map the neural networks (NNs) onto the hardware have also
been conducted [10]. New hardware “accelerators”, like the
tensor processing unit (TPU), was also designed to enhance the
efficiency of matrix operations and have benefited the success
through optimizations of the digital circuit and architecture
design for these relatively narrow types of operations [11]. Unlike
conventional hardware systems, RS crossbar array structures can
naturally perform VMM during a single read step [12]. The Neural
Network structure is readily mapped to the crossbar arrays, where
the RS devices located at each cross point can store the weight
matrix values also as producing an output depending on the input
and therefore the weight. The inputs and outputs of the network
(or a layer within the network) are connected to rows and columns
of the crossbar array, respectively as shown in Figure 6. During
inference, read voltage pulses are fed to the rows of the crossbar
particular to the input signals, the VMM outputs are collected as
the current through the crossbar array at the columns as shown
in Figure 6. The VMM operation is completed concurrently with
none of the data movement between processing and memory
unitsirrespective of the matrix size, thus offering very high
parallelism that results in superior computing throughput and
very high energy efficiency [13, 14].
RS devices have also been used for logic operations in electronic
circuits, which offers a high compute density and non-volatility. A
fundamental Boolean logic operation, material implication (IMP),
isimplemented as a logic gate two RS devices. Logic values can
be represented by the resistance of the RS devices (i.e., “0” for
the High resistance State-HRS and “1” for the Low resistance
state-LRSand the IMP operation can be achieved based on the
voltage divider effect. With well-designed voltage pulses and
series resistance values, the resistance of the output RS device (q)
is determined by the input logic state (p), producing the desired
truth table for the IMP operation. With iterative IMP operations,
all other Boolean logic operations can then be achieved [15, 16].
Beyond logic applications, RS devices are often used as nonvolatile
switches in field-programmable gate arrays (FPGA)
systems [17-19]. Switches using RS devices are reprogrammed to
reconfigure the connections and therefore the functionality of
the FPGA, resulting in improved density and power metrics. RS
devices can act as logic elements that summarise the data paths
connecting logic gates into digital circuits, enabling FPGA-like
functionality.
Conclusions
In this paper, we have highlighted the history, operation of
RS devices as a memory for reading and write operation and
its potential applications. RS devices have made remarkable
progress over the last 15 years. RS devices have offered commercial productsfor memory applications and have also
been extensively used for neuromorphic computing applications,
providing significant benefits for real-time data processing with
highthroughput and low energy consumption.
References
- Lee SH, Zhu X, Lu WD. Nanoscale resistive switching devices for memory andcomputing applications. Nano Research. 2020; 1-16.
- Marani R, Gelao G, Perri AG. A review on memristor applications. arXiv preprintarXiv:1506.06899. 2015.
- Talati N, Ben-Hur R, Wald N, Haj-Ali, Reuben J, Kvatinsky S. mMPU-A Real Processing-in-Memory Architecture to Combat the von Neumann Bottleneck. In Applications of Emerging Memory Technology. Springer, Singapore. 2020; 191-213.
- Wulf WA, McKee SA. Hitting the memory wall: implications of the obvious. ACM SIGARCH computer architecture news. 1995; 23(1): 20-24.
- Horowitz M. 1.1 computing's energy problem (and what we can do about it).In 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC). 2014: 10-14.
- Chua L. Memristor-the missing circuit element. IEEE Transactions on circuit theory. 1971; 18(5): 507-519.
- Kim KH, Gaba S, Wheeler D, Cruz-Albrecht JM, Hussain T, Srinivasa N, et al. functional hybrid memristor crossbar-array/CMOS system for data storage and neuromorphic applications. Nano letters. 2012; 12(1): 389-395. PMID: 22141918.
- Kawahara A, Azuma R, Ikeda Y, Kawai K, Katoh Y, Hayakawa Y, et al. An 8 Mb multi-layered cross-point ReRAM macro with 443 MB/s write throughput. IEEE Journal of Solid-State Circuits. 2012; 48(1): 178-185.
- Markram H, Lübke J, Frotscher M, Sakmann B. Regulation of synaptic efficacy by coincidence of postsynaptic APs and EPSPs. Science. 1997; 275(5297): 213-215. PMID: 8985014.
- Chen YH, Krishna T, Emer JS, Sze V. Eyeriss: An energy-efficient reconfigurable accelerator for deep convolutional neural networks. IEEE journal of solid-state circuits. 2016; 52(1): 127-138.
- Jouppi NP, Young C, Patil N, Patterson D, Agrawal G, Bajwa R, et al. Indatacenter performance analysis of a tensor processing unit. In Proceedings of the 44th Annual International Symposium on Computer Architecture. 2017; 1-12.
- Di Ventra M, Pershin YV. The parallel approach. Nature Physics. 2013; 9(4): 200-202.
- Zidan MA, Strachan JP, Lu WD. The future of electronics based on memristive systems. Nature Electronics. 2018; 1(1): 22-29.
- Chen B, Cai F, Zhou J, Ma W, Sheridan P, Lu WD. December. Efficient in-memory computing architecture based on crossbar arrays. In 2015 IEEE International Electron Devices Meeting (IEDM). IEEE. 2015; 17-5.
- Borghetti J, Snider GS, Kuekes PJ, Yang JJ, Stewart DR, Williams RS. Memristive’switches enable ‘stateful’logic operations via material implication. Nature. 2010; 464(7290): 873-876.M
- Linn E, Rosezin R, Tappertzhofen S, Böttger U, Waser R. Beyond von Neumann logic operations in passive crossbar arrays alongside memory operations. Nanotechnology. 2012; 23(30): 305205.
- Xia Q, Robinett W, Cumbie MW, Banerjee N, Cardinali TJ, Yang JJ.. Memristor− CMOS hybrid integrated circuits for reconfigurable logic. Nano letters. 2009; 9(10): 3640-3645. PMID: 19722537.
- Strukov DB, Likharev KK. CMOL FPGA: a reconfigurable architecture for hybrid digital circuits with two-terminal nanodevices. Nanotechnology. 2005; 16(6): 888.
- Snider GS, Williams RS. Nano/CMOS architectures using a field-programmable nanowire interconnect. Nanotechnology. 2007; 18(3): 035204.