Enhanced read resolution in reconfigurable memristive synapses for Spiking Neural Networks

The synapse is a key element circuit in any memristor-based neuromorphic computing system. A memristor is a two-terminal analog memory device. Memristive synapses suffer from various challenges including high voltage, SET or RESET failure, and READ margin issues that can degrade the distinguishability of stored weights. Enhancing READ resolution is very important to improving the reliability of memristive synapses. Usually, the READ resolution is very small for a memristive synapse with a 4-bit data precision. This work considers a step-by-step analysis to enhance the READ current resolution or the read current difference between two resistance levels for a current-controlled memristor-based synapse. An empirical model is used to characterize the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hbox {HfO}}_{2}$$\end{document}HfO2 based memristive device. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\textrm{st}$$\end{document}1st and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2\textrm{nd}$$\end{document}2nd stage device of our proposed synapse design can be scaled to enhance the READ current margin up to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}∼ 4.3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}× and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}∼ 21%, respectively. Moreover, READ current resolution can be enhanced with run-time adaptation techniques such as READ voltage scaling and body biasing. The READ voltage scaling and body biasing can improve the READ current resolution by about 46% and 15%, respectively. TENNLab’s neuromorphic computing framework is leveraged to evaluate the effect of READ current resolution on classification, control, and reservoir computing applications. Higher READ current resolution shows better accuracy than lower resolution even when facing different levels of read noise.


I. INTRODUCTION
T HE development of Artificial Neural Networks (ANNs) and Deep Neural Networks (DNNs) inspired by the remarkable information processing abilities of mammalian brains, while also achieving low power consumption and minimal latency.Due to their exceptional classification accuracy, DNNs are attracting considerable interest as the preferred classifier in numerous machine learning and computer vision applications [1].However, DNNs are typically executed on the von Neumann machines and are hence limited by the separation of memory and processing units, also known as the von Neumann bottleneck [2].Besides that, the extensive computational requirements, high power consumption, and memory bandwidth associated with DNNs make them unsuitable for mobile applications, where limitations in area and power are significant constraints [3].
In order to address the limitations in power and memory capacities of traditional computing architectures mentioned above, coupled with inspiration drawn from the efficiency of the biological nervous system, a novel concept of neuromorphic architectures has emerged.These architectures typically employ Spiking Neural Networks (SNNs), which aim to mimic the intricacies of the biological nervous system with greater fidelity by employing binary pulses as a means of communication.These architectures represent a distinct paradigm from the conventional von Neumann architectures in terms of the colocation of memory and processing unit, demonstrating promising results in specific applications [4]- [6].Neuromorphic architectures not only offer better energy efficiency, but also promises parallel signal processing, fault tolerance, and reconfigurability.Furthermore, they can be implemented using diverse silicon-based technologies, large-scale architectures, and computational models of neural components [7]- [10].
Neuromorphic systems also referred to as Neuroprocessors leverage the colocation of memory and processing units, where neurons act as the computational units, interconnected by synaptic memory elements.Synapses contain the weighted connections between neurons and can be implemented using digital [11] or analog [12], [13] circuits.Nevertheless, incorporating a multitude of synapses presents several obstacles, including the efficient handling of storage space needed for weighted connections and the accommodation of diverse synaptic learning techniques that demand adaptable weight storage [9], [14].Memristors are potential candidates to address these issues.Memristors have proven to be more compact and power efficient for synaptic implementation compared to SRAM and capacitor-based implementations of the same resolutions [9], [15].Memristive synapses have also been proved to have extended memory retention time [15].First postulated by Leon Chua, memristors are described as the fourth fundamental passive circuit element [16].A memristor is a two-terminal device with analog memory properties that originate from its ability to switch resistances.When a voltage beyond a certain threshold is applied to its terminals, the resistance is modified.A memristor state is also non-volatile making it a promising candidate for weight storage.Due to their compatibility with CMOS technology and non-volatile properties, memristors are well-suited for analog computation [17].
In this work, we use a current-controlled synapse designed using a TiN TE/HfO 2 /TiN BE memristor.To use this synapse, several operations need to be performed on the memristor, such as FORMING, RESET, SET, and READ.This memristor can vary its resistance in a few kΩ to over 150kΩ [12], [13].However, for our design, we exclusively use the low resistance states (LRS) of the memristors due to the improved reliability in this operating region, and to avoid a high degree of variability encountered near the high resistance states (HRS).The memristor can be programmed into different low-resistance states by precisely controlling the compliance current during SET operation that overcomes the issues of variability and limited resolution [18].
However, although the synapse is designed while taking the reliability concerns into consideration, the synapse is vulnerable to limited resolution due to the use of a narrow resistance range in the low-resistance regions [19].This occurs when the current generated by the different resistance states is undifferentiable.This limitation can cause a reduction in the learning performance of SNNs [9], [19].Overlapping synaptic currents also make the synapse to be susceptible to noise and process variation [9].Another disadvantage of the small difference between the synaptic states is that it complicates the analog-to-digital converter (ADC) design significantly [9].In order for the ADC to recognize different resistance states for digital conversion, the current output difference between the resistance states needs to be high enough for a compact power efficient design [9].This work aims at improving the current resolution of the memristive synaptic circuit using several techniques.
The key contributions of this paper are as follows.1) READ current resolution of a current compliance memristive synapse is enhanced .
2) READ current resolution is enhanced with proper device scaling.
3) READ current resolution is made re-configurable at run time with READ voltage scaling.4) READ current resolution is adaptable at run time with body biasing.
5) The TENNLab neuromorphic framework is utilized to observe the effect of READ current resolution on SNNs.Higher READ current resolution illustrates better accuracy with a lower possibility of a read error.The rest of the paper is organized as follows.Section II briefly describes a Verilog-A model, our proposed synapse with a hafnium oxide-based memristor device, and its READ operation.Section III shows the proper device sizing to enhance the READ current resolution.Section IV illustrates two techniques to make the READ current resolution re-configurable at run time.Section V will evaluate the design performance based on different test cases.The next section (VI) exhibits the effect of the READ current resolution or weight resolution on Spiking Neural Networks (SNNs).A detailed comparison with prior works is analyzed in Section VII.Finally, the paper is concluded in Section VIII with prospective future work.

II. CURRENT-CONTROLLED MEMRISTIVE SYNAPSE A. HfO 2 Based Device Modeling
A Verilog-A model is utilized to simulate the HfO 2 based memristive devices [20].In this model, mathematical equations are derived based on the memristance state and the required time to switch the states between HRS to LRS or LRS to HRS.The I-V characteristics of this device are taken under consideration to derive the empirical model of the memristor.The threshold voltage and switching time are two sets of important measured parameters for this model.There are some fitting constants in the model, which are utilized to fit the device's I-V characteristics to observe the simulation behavior as closely as measured data.The sigmoid window function is also considered to be incorporated with different patterns of switching time between HRS to LRS and LRS to HRS.This Verilog-A model is also capable of detecting the RESET failure if the RESET voltage is crossed its functional window.All mathematical equations, I-V curve, measured and, simulation details are available in our prior work [20].

B. Proposed Synapse
Memristors are widely used to construct brain-inspired synapses, which is the key element for neuromorphic computing.There are different synapse flavors based on the memristor's recipe (combination of materials).Various materials are utilized to build the memristors such as HfO 2 , Ta 2 O 5 , NbO 2 , and so on.Our proposed architecture is designed using a TiN TE/HfO 2 / TiN BE memristor.Fig. 1 shows the proposed synapse with READ devices.However, the first step of this synapse is the onetime FORM operation.Thick oxide transistors are used for this design to take care of high voltage around 3.3 V for forming.A thick oxide transistor is also useful to reduce flicker noise.Unformed memristors are usually exhibit resistance in the range of ∼8 MΩ to ∼10 MΩ.After forming, the memristor's resistance level will be in few kΩ.Hence, the synapse needs to RESET to a higher resistance state (HRS), which is typically hundreds of kΩ.Finally, our device is ready to SET /write /program to a specific low resistance state (LRS) from a HRS.Due to less variability, the programming region is selected in the LRS region.The targeted LRS for this design is from 5 kΩ to 20 kΩ.As a result, we can utilize 4-bit precision with 1 kΩ resistance resolution.This design needs a set of RESET and SET to program in a new LRS value.After a successful SET operation, the synapse is ready for a READ operation.
Fig. 1 shows the proposed current-controlled synapse with READ circuitry.V READ and V READB signals are utilized to access the memristor (R) during a READ operation.V DD READ and V READ are 1.2 V and 0.6 V respectively during a READ operation.Due to the READ signal assertion, there will be a small 1 st stage current through the memristor.This current will create a voltage to operate transistor M N 2 .Finally, a READ Current is sensed from the drain of the M N 2 .The body of this M OSF ET is utilized as a separated signal to control the threshold voltage of M N 2 .READ is a very sensitive operation for memristor-based synapses.Especially, the READ margin between two resistance levels (e.g. 5 kΩ and 6 kΩ ) needs to be good enough to read the data or weight properly.Most of the time the difference between the two resistance levels is few nA for low power design, which is very hard to sense properly.A research paper shows the READ current between 5 kΩ and 6 kΩ is 20 nA [21].A few techniques can be utilized to overcome this low READ margin /resolution.All the techniques are explained below with proper analysis.

III. READ CURRENT RESOLUTION ENHANCEMENT WITH PROPER DEVICE SCALING
At first, READ device sizing is considered to observe the effect on READ current resolution.A 65 nm CMOS process is utilized to construct and conduct Cadence Spectre simulations.A Verilog-A model is utilized to characterize the HfO 2 based memristor device [20].Fig. 2(a) shows effect on READ current resolution with the sizing of M P 1 and M N 1 .Here, the width and length of M N 2 are set at 0.5 µm.In addition, the width of the M P 1 is varied from 0.5 µm to 4 µm and the width of the M N 1 is varied from 1 µm to 4 µm.In addition, the length of M P 1 and M N 1 (both are thick oxide transistor) are set at 0.5 µm.When the width of both M N 1 and M P 1 is minimal, the READ current resolution (one memristive level to another, e.g. 5 kΩ to 6 kΩ) is at least 19 nA.Due to an optimized read procedure, the READ current shows very stable resolutions compared to prior work [21].In this work, a regular pfet is utilized to control the read voltage at the drain of M N 2. Whereas a diode-connected pfet was connected in prior work.Hence, the width of the M P 1 set at 0.5 µm and the width of the M N 1 varies from 2 µm to 4 µm.The READ current resolution is 64 nA, when the width of the M N 1 and M P 1 are 4 µm and 0.5 µm respectively.The larger size of M N 1 and smaller size of M P 1 allow suitable gate voltage for the M N 2 to provide a high-resolution READ current.Twelve different sizing combinations are observed for READ current resolution.When the width of the M P 1 and M N 1 are 1 µm and 4 µm, the READ current resolution is about 81 nA for 4-bit data precision.According to the last test case, if the width of M P 1 increases significantly and its size becomes the same as M N 1 , then the READ current resolution does not show significant benefit on sizing.A better sizing combination is observed, when the M N 1 and M P 1 are not the same and M P 1 is smaller than M N 1 .According to our sizing analysis, the READ resolution provides best performance when the width of M N 1 and M P 1 are 4 µm and 1 µm respectively.Next, the effect of length and width of the M N 2 is observed, with the width of the M N 1 and M P 1 transistors are set at 4 µm and 1 µm respectively.
The length and width of M N 2 are varied to observe the effect on READ current resolution.Fig. 3 (a) illustrates the READ current resolutions, when the width of the device is varied from 0.5 µm to 4 µm and the length is fixed at 0.5 µm.When the length and width of the device are set at a minimal size, the READ current resolution is about 81 nA.The READ current resolution is drastically decreased as the width of the M N 2 is scaled up.At the same time, the READ current should be increased with M N 2 up-sizing.As a result, the first test case shows better READ current resolutions with lower power consumption.
Next, the width of the M N 2 is set at 0.5 µm and the length is varied from 0.5 µm to 4 µm.The READ current resolution is decreased with the increment of length.At this point, the overall READ current will be decreased by up-sizing the length of M N 2 , with higher latency.Finally, both length and width are increased simultaneously.Fig. 3 (c) shows the READ current resolutions when both length and width are increased.READ current resolutions are increased as the length and width are up-scaled at the same time.About 21% improvement in READ current resolutions can be achieved by up-sizing the length and width of M N 2 simultaneously.Fig. 3 (d) shows the READ current level with different sizing combinations.READ current resolution can be increased with the overhead of area and READ current.Fig. 3 (d) shows, at 5 kΩ memristive weight, the READ current is 21.8 µA with minimal length and width of the M N 2 .On the other hand, when both length and width of M N 2 are increased to 4 µm, the READ current is increased by 22.48%.At the same time, the READ current resolution is increased by 21%.There is a clear trade-off between READ current resolution and READ current overhead.In addition, the overall design area is also influenced by a larger length of M N 2.
TABLE I shows the optimized sizing configuration to enhance the READ current resolution.The width of M P 1 and M N 1 are set at 1 µm and 4 µm respectively.Both M OSF ET 's length are fixed at 0.5 µm.In addition, the length and width of M N 2 is considered as 4 µm.The READ current resolution is about 98 nA with this optimized sizing.In the next section, the READ current resolution will be adapted dynamically with V READ and MN2 Bias signals.

IV. RECONFIGURABLE READ CURRENT RESOLUTION
READ current resolution can be adapted at run time.Various applications can perform better at enhanced READ current or weight resolution.To enhance the application's performance, a reconfigurable or run-time adaptation of READ current resolution is proposed with different circuit techniques.Specifically, READ voltage (V READ) scaling is a useful technique to influence the READ current resolution at run time.

A. READ Voltage Scaling
Fig. 1 shows V READ is the gate voltage of M N 1 .Initially, the applied amplitude of this signal was 0.6 V. Fig. 4 (a) shows the READ current resolution at 0.6 V is 98 nA for 4-bit data.Here, the optimized sizing is utilized from TABLE I.The READ current resolution is about 143 nA at 0.63 V as V READ.About 46% of READ current resolution can be enhanced   with V READ scaling at run time.If the V READ is increased more than 0.63 V for this particular sizing or configuration, then the resolution is decreased.A stronger turn-on of M N 1 influences the final READ current negatively.As a result, the resolution is decreased with excessive V READ.Fig. 4 (b) shows READ current with different gate voltages of M N 1 .When the READ voltage is 0.6 V, the READ current is about 26.8 µA at 5 kΩ.On the other hand, if the READ voltage is scaled up to 0.63 V, the READ current will be decreased to about 19.1 µA.About 29% READ current can be optimized when the READ voltage is scaled from 0.6 V to 0.63 V. READ current shows a significantly lower value at 0.64 V.However, at 0.64 V the READ current resolution is reduced significantly compared to the value at 0.63 V.In addition, at 0.64 V the std.dev. of READ current is ∼ 7% higher than std.dev. at 0.63 V. Thus, the READ current at 0.63 V is more reliable than at 0.64 V.According to our design optimization, the READ current resolution is higher at 0.63 V among all the READ voltages.In addition, the READ current resolution can be varied from 98 nA to 143 nA with V READ signal scaling.In the next sub-section, another device technique is utilized to enhance the READ current resolution at run time.

B. READ Device M N 2 Biasing
READ device M N 2 biasing is another technique to manage synaptic READ current at run time.Here the READ voltage V READ is set at 0.63 V with optimized sizing.Fig. 5 (a) shows the READ current resolution at different bias voltages at the body of M N 2 .As we know body biasing can change the threshold voltage of a M OSF ET to control the current flowing through it.Here a positive body voltage is applied on M N 2 to improve the READ current resolution.When the body voltage is 0 V, the READ current resolution of our proposed synapse is about 143 nA.If the body voltage is scaled up to 0.3 V, then the READ current resolution is further up to 15%.In addition, if the body biasing voltage is increased more, then a reverse phenomenon is observed with READ current resolution.Due to the channel effect of the NMOS, the reverse phenomenon is observed with higher body bias voltage.At 0.4 V the READ current resolution is decreased to 88 nA.As a result 0.3 V is an optimized body bias voltage for this design scenario.Fig. 5 (b) exhibits the READ current of our synapse at different body biasing scenarios.Overall READ current is increased with the increment of body biasing voltage.About 47% READ current is increased with body biasing to enhance the READ current resolution at run time.Finally, it can be said that the READ current resolution of our proposed synapse can be adapted at run time with different body biasing voltages.

V. READ CURRENT RESOLUTION AND DESIGN PERFORMANCE EVALUATION
TABLE II shows an evaluation of device sizing, READ current resolution, and READ power.There are five test cases considered for the evaluation.The first test case is constructed with base device sizing.Here, the V READ and body bias (MN2 Bias) are 0.6 V and 0 V respectively for first three test cases.The first test case shows 19 nA READ current resolution with 18.87 µW as a max READ power.Here, both stages' power (1 st and 2 nd ) are considered for the max READ power.The second test case is the 1 st stage device sizing.Here the READ current resolution is 81 nA with 4-bit precision which is about 4.3x higher than the base sizing test case.Only 5% of power overhead is observed compared to the base test case.The next test case is focused on 2 nd stage device sizing with 1 st stage sizing.Here, the READ current resolution is 98 nA, which is 5.16x more improved than the base test case (Base sizing).Here, the READ power overhead is only 0.16% compared to the base test case.The fourth test case is to increase the V READ to 0.63 V from 0.6 V.In this scenario, the READ current resolution is 143 nA, which is 7.53% enhanced than the base test case.In addition, the max READ power shows 1.43% overhead compared to the base test case.Finally, the body biasing is applied to the body of M N 2 .After applying 0.3 V, the READ current resolution is about 165 nA, which is 8.86x improved compared to the base test case.Here the max READ power overhead is about 3.60% compared to the base test case.The READ voltage on the source of the M N 2 is automatically adjusted based on the gate voltage of the M N 2 .Due to that, the READ power is quite stable in this design.Here the Monte Carlo simulation is observed with 1000 samples in a Cadence Spectre environment to analyze the READ current variation.The READ current shows about 0.65x variations with 8.68x resolution improvement.

VI. IMPACT OF READ CURRENT RESOLUTION ON APPLICATIONS
To evaluate the impact of READ current resolution on applications, we specifically wanted to evaluate how what impact READ errors would have on performance.Our hypothesis is that when the current resolution is higher, there will be fewer or no read errors, while when the current resolution is lower, read errors are significantly more likely.To perform this evaluation, we leveraged the TENNLab neuromorphic computing framework [22], which allows for evaluation of neuromorphic processors using different applications and algorithms.Within the framework, we used the RISP neuromorphic simulator [23], with integrate and fire neurons and synapses with 4-bit weight resolution.
To specifically study the impact of READ current resolution, we evaluated how spiking neural networks with different likelihoods of read failures for each synapse.In particular, for a particular network evaluation, we defined a likelihood of read failures for each synapse read, wherein the weight read would be one level off (either the level above or level below), which would be more likely to happen for low current resolution.We trained the networks with these read failures using an evolutionary optimization training approach for spiking neural networks and neuromorphic system called EONS [24].We trained for the iris dataset, the wine dataset, and the breast cancer dataset, three commonly used toy datasets in machine learning that are available in the UCI machine learning repository [25].We trained 100 networks for each of six different likelihoods of read errors for the synaptic weight values: 0, 10, 20, 30, 40, and 50 percent for each dataset.Figure 6 shows the results for these simple datasets on both training and testing performance.This figure shows that, in general, the best  overall testing performance was achieved by no read errors at all; however, on average, some noise on the read errors does not necessarily hurt performance significantly, either in training or testing.For the results in Figure 6, it is worth noting that the networks were trained and tested using read errors.Figure 7 shows the results for when networks are trained without read errors and then tested with varying likelihoods of read errors per synapse, which would likely be the case for networks trained in simulation and then deployed to hardware.In this case, we can see that read errors cause a decrease in testing performance for each dataset.Finally, because this work enables higher-precision weights to be used on synapses with more reliability, we investigated the impact of precision on performance.Figure 8 shows the results for different levels of bit-precision (2, 3, and 4) on the synaptic weights.Here, we can see that 2-bit weight synapses perform significantly worse across all three datasets than 3-and 4-bit precision synapses, as expected.

VII. COMPARISON WITH PRIOR WORK
READ current resolution is a big challenge for the memristive-based synapse or memory design.Usually, the minimum current resolution is a few nA.Due to that, it is challenging for the circuit designer to sense the current level properly with ADC or CMOS neurons.In our neural network analysis, we consider a CMOS neuron to observe the charge accumulation and fire [26].In this work, a memristor-based synapse is optimized to enhance the READ current resolutions.At first device sizing is considered to optimize the resolution.Hence, READ voltage and body bias are considered to enhance the resolution at run  time.Here, the optimal size is adopted from TABLE I.In addition, the READ voltage and body bias voltages are selected as 0.63 V and 0.3 V respectively.According to TABLE III, the minimum READ current resolution is 165 nA with 4-bit data precision.As we know, the larger resistance level shows a lower current resolution.In our case, the current level difference between 19 kΩ and 20 kΩ is considered to determine the minimum current resolution.At room temperature the value is at least 165 nA.The maximum READ current ∼ 28 nA is captured at 5 kΩ.Our synapse is programmed at LRS to avoid inherent process variations.
Another research article shows the minimum READ current resolution of a 3T1R synapse is 20 nA for 4-bit data precision.Our proposed design shows 8.25x resolution improvement compared to their work [21].Energy-efficient and high-performance synapse is presented in another research paper, where the minimum READ current resolution is 20 nA [19].This design is also based on HfO 2 based memristor and 65 nm CMOS process.The programming region is between 3 kΩ to 18 kΩ, which provides low inherent process variation with 4-bit data density.Our proposed design also illustrated 8.25x current resolution improvement compared to their work [19].A TiO 2 based memristive memory is presented in [27], which is programmed from 1 kΩ to 1 GΩ.Due to the utilization of HRS, the inherent process variation is higher for this design compare to our proposed design.This design also covers four programming states.As a result, our proposed design shows a 4x more dense data storage capacity compared to their design [27].The max READ current is about 3.21x higher than our proposed design.The minimum READ current resolution is about 90 nA, which is 1.83x lower than our proposed design.
Another research group is presented a TiO 2 based multi-level resistive memory, which is programmed from 10 kΩ to 100 MΩ [28].Due to programming in HRS, the inherent process variation is higher than our design.Their design is only programmed in six different states, which exhibits 2.67x lower memory density compared to our proposed design.The maximum READ current of their design is 100 µA, which is 3.57x higher than our proposed design.Moreover, the minimum READ current resolution of their design is ∼ 17 nA, whereas our proposed design shows ∼ 9.7x enhanced resolution compared to their design [28].A HfO 2 based multi-level cell is presented with 2-bit memory density, which is 4x lower memory density than our proposed synapse [29].Their device is programmed from 10 kΩ to 1 MΩ, which shows higher inherent process variation compared to our design.Their device draws lower maximum READ current than our design.However, the minimum READ current resolution is 2.04x lower than our proposed design.Another multi-level resistive memory is presented using a-ZnO, which is programmed between 24 MΩ to 176 MΩ [30].Their design shows higher process variation and 3.2x lower memory density than our proposed design.Their maximum READ current is lower than our design.However, the minimum READ current resolution is extremely low compared to our design.A HfO 2 based multi-level RRAM is presented in a research article, where authors programmed their device between 0.8 kΩ to 100 MΩ with 8 programming states [31].Our proposed design shows lower process variation and 2x higher memory density than their design.The maximum READ current of their design is extremely higher than our proposed design.Our proposed design also shows a 3.67x improved minimum READ current resolution than their design [31].Another research group presented their device with TiO 2 , which is programmed between 1 MΩ to 10 GΩ with six programming states [32].This design shows a 2.67x lower memory density than our proposed design.Due to their HRS programming region, the process variation is higher than our design.The maximum READ current of their design is lower than our proposed design.However, their minimum READ current resolution is extremely lower than our proposed design.Moreover, our design shows power savings compare to prior works [27], [28], [31].In addition, our proposed design shows more linear READ current compared to other designs [27]- [32].Due to a compact programming range, our design shows better linearity compared to prior works.A wide programming range causes non-linear behavior of READ current.
After comparing our proposed design with prior works from different research groups, it can be said that our proposed design shows lower inherent process variation with higher memory density.In addition, our proposed design shows an enhanced READ current resolution compared to others' designs.

VIII. CONCLUSIONS AND FUTURE WORK
In this paper, a HfO 2 based current-controlled memristive synapse is optimized for READ operation.At first, the READ devices are optimized to enhance the READ current resolution.About 4.3x and 21% READ current resolution is enhanced with 1 st and 2 nd stage device sizing respectively.READ voltage scaling and body biasing are applied to enhance the READ current resolution at run time.About 46% and 15% READ current resolution is improved with READ voltage scaling and body biasing.A neuromorphic framework shows that a higher READ current resolution exhibits better accuracy compared to a lower resolution on a simple classification application.Lower resolutions are more likely to be affected by reading failures with higher noise.As a result, a higher READ current resolution makes the neuromorphic system more reliable.

Fig. 1 .
Fig. 1.Memristor with READ circuitry is illustrated.READ operation requires 1-PMOS and 2-NMOS.READ operation folded into two parts. 1 st stage current is generated with M P 1 and M N 1 .This current initiates a voltage to operate M N 2 in the linear region.Finally, the READ current will be sensed from the drain of M N 2 .The body of the M N 2 is considered as a dedicated signal to control the threshold of this device.

Fig. 2 .
Fig. 2. READ simulation results are illustrated based on the sizing of M P 1 and M N 1 .The length and width of M N 2 are fixed at 0.5 µm.(a) M P 1 is varied from 0.5 µm to 4 µm.In addition,M N 1 is varied from 1 µm to 4 µm.Larger M N 1 shows a higher impact on the READ current resolution.(b) shows the READ current scale with different size of M P 1 , when the width of the M N 1 is fixed at 4 µm.

Fig. 3 .
Fig. 3. Cadence simulation results for READ current resolutions with different width and length of M N 2 device.(a) shows the READ current resolution when the length of the M N 2 is fixed at 0.5 µm and the width is varied from 0.5 µm to 4 µm.The current resolution is drastically decreased with the increment of M N 2 's width.(b) shows the READ current resolution when the width of the M N 2 is fixed at 0.5 µm and the length is varied from 0.5 µm to 4 µm.The READ current resolution is also decreased with the increment of length of M N 2 .Finally (c) shows the READ current resolution when the length and width of M N 2 change simultaneously.READ current resolution is increased when the length and width of the M N 2 are increased at the same time.(d) shows the READ current level with different M N 2 sizing.About 22.48% READ current overhead is observed to improve 21% READ current resolution.

Fig. 4 .
Fig. 4. READ voltage (V READ) has a significant effect on READ current resolution.(a) shows the READ current resolution at different V READ voltage.READ current resolution is increased with the increment of the gate voltage of M N 1 .After a certain level of gate voltage increment, the READ resolution starts decreasing.(b) exhibits the READ current level at different READ voltages.As we increase the READ voltage at the gate of M N 1 , the READ current level starts decreasing.Due to a weak turn-on of M N 2 , the READ current level is decreased.Here is an interesting thing to notice, as we increase the gate voltage of M N 1 the overall READ current level is decreased.As a result, the READ current resolution is increased with overall READ current optimization.

Fig. 5 .
Fig. 5. READ device M N 2 plays an important role to make READ current resolution adaptable at run time.The body of M N 2 is scaled to enhance the resolution with READ power overhead.(a) shows the body biasing effect on READ current resolution.About 15% resolution can be enhanced with body biasing.(b) illustrates the READ current level at different biasing voltages.The READ current is increased with the increment of body biasing of M N 2 .

Fig. 6 .
Fig.6.Training and testing accuracy for iris, wine, and breast cancer datasets.In these results, the networks were trained and tested with read errors.

Fig. 7 .
Fig. 7. Testing accuracies for networks trained without read errors and tested with read errors.

Fig. 8 .
Fig. 8. Results for reduced precision on the synaptic weights.

TABLE I TRANSISTOR
SCALING TO ENHANCE READ CURRENT RESOLUTION

TABLE II READ
CURRENT RESOLUTION ENHANCEMENT AND DESIGN EVALUATION