toward neuromorphic computing

Upload: dolev-krongauz

Post on 04-Apr-2018

218 views

Category:

Documents


0 download

TRANSCRIPT

  • 7/31/2019 Toward Neuromorphic Computing

    1/8

    Nanotube devices based crossbar architecture: toward neuromorphic computing

    This article has been downloaded from IOPscience. Please scroll down to see the full text article.

    2010 Nanotechnology 21 175202

    (http://iopscience.iop.org/0957-4484/21/17/175202)

    Download details:

    IP Address: 89.87.247.231

    The article was downloaded on 11/04/2010 at 21:22

    Please note that terms and conditions apply.

    The Table of Contents and more related content is available

    ome Search Collections Journals About Contact us My IOPscience

    http://www.iop.org/Terms_&_Conditionshttp://iopscience.iop.org/0957-4484/21/17http://iopscience.iop.org/0957-4484/21/17/175202/relatedhttp://iopscience.iop.org/http://iopscience.iop.org/searchhttp://iopscience.iop.org/collectionshttp://iopscience.iop.org/journalshttp://iopscience.iop.org/page/aboutioppublishinghttp://iopscience.iop.org/contacthttp://iopscience.iop.org/myiopsciencehttp://iopscience.iop.org/myiopsciencehttp://iopscience.iop.org/contacthttp://iopscience.iop.org/page/aboutioppublishinghttp://iopscience.iop.org/journalshttp://iopscience.iop.org/collectionshttp://iopscience.iop.org/searchhttp://iopscience.iop.org/http://iopscience.iop.org/0957-4484/21/17/175202/relatedhttp://iopscience.iop.org/0957-4484/21/17http://www.iop.org/Terms_&_Conditions
  • 7/31/2019 Toward Neuromorphic Computing

    2/8

    IOP PUBLISHING NANOTECHNOLOGY

    Nanotechnology 21 (2010) 175202 (7pp) doi:10.1088/0957-4484/21/17/175202

    Nanotube devices based crossbararchitecture: toward neuromorphiccomputing

    W S Zhao1, G Agnus2, V Derycke2, A Filoramo2, J-P Bourgoin2 andC Gamrat1

    1 CEA, LIST, Embedded Computing Laboratory, Building 528, 91191, Gif-sur-Yvette, France2 CEA, IRAMIS, Service de Physique de lEtat Condense (CNRS URA 2464),

    Laboratoire dElectronique Moleculaire, 91191 Gif-sur-Yvette, France

    E-mail: [email protected] and [email protected]

    Received 10 December 2009, in final form 29 January 2010

    Published 6 April 2010

    Online at stacks.iop.org/Nano/21/175202

    Abstract

    Nanoscale devices such as carbon nanotube and nanowires based transistors, memristors and

    molecular devices are expected to play an important role in the development of new computing

    architectures. While their size represents a decisive advantage in terms of integration density, it

    also raises the critical question of how to efficiently address large numbers of densely integrated

    nanodevices without the need for complex multi-layer interconnection topologies similar to

    those used in CMOS technology. Two-terminal programmable devices in crossbar geometry

    seem particularly attractive, but suffer from severe addressing difficulties due to cross-talk,

    which implies complex programming procedures. Three-terminal devices can be easilyaddressed individually, but with limited gain in terms of interconnect integration. We show how

    optically gated carbon nanotube devices enable efficient individual addressing when arranged in

    a crossbar geometry with shared gate electrodes. This topology is particularly well suited for

    parallel programming or learning in the context of neuromorphic computing architectures.

    (Some figures in this article are in colour only in the electronic version)

    1. Introduction

    For at least three decades, complementary metal-oxide

    semiconductor (CMOS) technology has been the dominant

    technology for integrated circuits (ICs) and computing

    systems. Nevertheless, miniaturization of CMOS ICs will

    probably be limited close to the 11 nm node because of

    physical barriers such as quantum effects3. Below this size,

    Boolean computing will face very severe difficulties, due in

    particular to high leakage currents, which will dramatically

    impact the power consumption performances of CMOS

    circuits [1]. In this context, devices based on nano-objects

    such as carbon nanotubes [2], nanowires [3, 4], graphene [5]

    or molecules [6] are viewed as possible means to go beyond

    Moores law. In order to make the implementation of

    nanodevices with CMOS circuits tolerant to the intrinsic

    3 International Technology Roadmaps for the Semiconductors (ITRS), 2008Update Review.

    high variability and defect density originating from their

    small size and from fabrication issues (potentially involving

    self-assembly processes) [7], intense research activity is

    presently developing to evaluate the use of adaptive neuralnetwork principles [8, 9] (figure 1) in which the synapses

    are built at the nanoscale [1013] while the neurons are

    made with conventional CMOS circuits. Such a strategy,

    while elegant, raises the critical issue of how to address

    individual nanodevices within large assemblies so as to allow

    programming or learning. Two-terminal nanodevices in

    crossbar arrays combined with CMOS based control circuits

    are considered as one of the most promising computing

    architectures [1320] but the simplicity of the topology

    introduces severe difficulties concerning individual addressing,

    for example cross-talk effects.

    We recently demonstrated that optically gated carbon

    nanotube field effect transistors (OG-CNTFETs) [2123] canbe operated as two-terminal programmable resistors and

    0957-4484/10/175202+07$30.00 2010 IOP Publishing Ltd Printed in the UK & the USA1

    http://dx.doi.org/10.1088/0957-4484/21/17/175202mailto:[email protected]:[email protected]://stacks.iop.org/Nano/21/175202http://stacks.iop.org/Nano/21/175202mailto:[email protected]:[email protected]://dx.doi.org/10.1088/0957-4484/21/17/175202
  • 7/31/2019 Toward Neuromorphic Computing

    3/8

    Nanotechnology 21 (2010) 175202 W S Zhao et al

    Figure 1. Generic diagram of a single-layer neural networkcomposed ofi inputs, j neurons and (i j ) synapses. A singleneuron and its associated synapses form a perceptron. Incomingsignals (Xi ) are multiplied by the synaptic weights ( Wij ) beforereaching the neuron that integrates the signals and triggers a responseusing, for example, a threshold function.

    that, in such a configuration, they have all the required

    characteristics of artificial synapses [23]. In particular, their

    resistivity can be programmed over a wide range and then

    stored in a non-volatile way so that it can play the role of a

    synaptic weight (Wi j in figure 1) in neural network topologies.

    The programming relies on a single illumination step of

    the whole circuit followed by local electrical stimulations

    of the input electrode (source or drain) of the different

    devices. Most importantly, we showed that in simple

    circuits composed of multiple OG-CNTFETs, each device

    can be independently programmed at a given resistivity value

    independently of its initial (post-fabrication) characteristics sothat the programming step fully compensates for the process

    variability [23]. This early work was focused at the device level

    and no circuit implementation was studied.

    In this paper, we present a crossbar topology for arrays of

    nanotube based synapses, in which gate electrodes are shared

    by all the synapses connected to the same neuron. Appropriate

    polarization of these gate electrodes allows the protection of

    already-trained synapses from further modification and fulfils

    a key requirement to achieve efficient addressing. Using

    simulations, we establish the potential of such a topology in

    terms of the programming/learning efficiency of large arrays.

    2. Optically gated carbon nanotube field effecttransistor (OG-CNTFET)

    2.1. OG-CNTFETs as memory transistors

    An OG-CNTFET consists of a p-type carbon nanotube field

    effect transistor (CNTFET) in back-gate geometry and coated

    with a thin film of photoconductive polymer, for example

    P3OT (poly(3-octylthiophene-2,5-diyl)) [21]. The FET

    channel can be variously composed of a single semiconducting

    single wall carbon nanotube (SWNT) (as schematically

    represented in figure 2(a)); multiple parallel semiconducting

    SWNTs (figure 2(b)) or an array of SWNTs (figure 2(c))potentially composed of both metallic and semiconducting

    SWNTs (but adjusted so as to provide an FET with a good

    off-state). At constant sourcedrain bias VDS and positive gate

    bias VGS, such a FET is in its off-state but can be turned-on using light at an appropriate wavelength and intensity (see

    figures 2(a) and (b) in [21] for details on wavelength and lightintensity dependence). Indeed, an illumination pulse generates

    electronhole pairs in the polymer, a fraction of which getsdissociated. While photo-induced holes are depleted from thedevice in this bias configuration, electrons get trapped (in thegate dielectric, close to the nanotubeSiO2 interface [22]). The

    trapped electrons apply a negative potential to the nanotubechannel, which becomes more conductive. After the light is

    turned off, a significant part of these electrons remain trapped

    so that the change forms a non-volatile memory state (weverified that the light-induced trapped charges show no signof decay after >40 h). Using either negative electrical pulses

    on the gate electrode [21, 22] or positive electrical pulses on

    the source electrode (figure 2(d)) [23], these electrons can bede-trapped in a controlled manner. By adjusting the number

    of trapped charges, the resistivity of the device channel can beprecisely programmed, as shown for example in figure 2(e),

    which displays eight I(V) curves of the same device at

    different stages of programming. The experimental detailsare described in [21] for single nanotube devices and in [23]

    for resistance programming using devices based on nanotubenetworks. The robustness of the write-read-erase cycles isillustrated in [24].

    2.2. OG-CNTFET programming using the gate protection

    effect

    Our initial prototype circuits [23] used OG-CNTFETs as

    two-terminal devices: we fix a constant gate potential fora series of devices sharing the same gate electrode andprogram them using the source electrode. We first illuminatethe circuit globally to set all the devices in their state of

    minimal resistivity and then program each nanotube resistor

    independently by applying electrical pulses on the input(source) electrodes. However, while very tempting from an

    integration point of view, this strategy is quite limited interms of programming/learning efficiency. Indeed, in a simplecrossbar, programming pulses would act on a full series of

    devices and not on individual ones. To overcome this stronglimitation, we use a specific property of OG-CNTFETs that is

    illustrated in figure 3. During programming, electrons are de-

    trapped from the gate dielectric only when the electric field isoriented in the appropriate direction, so that only negative VGSpulses or positive VDS pulses are efficient. As a consequence,

    if a strong positive VGS bias is applied, it shields (or protect)the device from being programmed by positive VDS pulses.

    This is shown in figure 3, where the current through an OG-CNTFET submitted to programming pulses is displayed in

    two configurations of gate bias: figure 3(b) corresponds to

    the conventional programming configuration: VGS = 0 Vand positive VDS pulses modify the final resistivity value.Figure 3(c) corresponds to VGS = 6 V. The same programming

    pulses are applied but the final current is unaffected. Note that

    in this latter configuration the current level is low during theprogramming cycle because VGS > 0 sets the FET in it off-

    state.

    2

  • 7/31/2019 Toward Neuromorphic Computing

    4/8

    Nanotechnology 21 (2010) 175202 W S Zhao et al

    Figure 2. (a) Schematic representation of a single nanotube based OG-CNTFET, (b) OG-CNTFET based on multiple semiconductingnanotubes in parallel, (c) OG-CNTFET based on a network of nanotubes, (d) output current of a nanotube network based OG-CNTFET atconstant VGS = +2 V. Light pulses (at = 457 nm and P = 40 W cm

    2) increase the conductance of the channel while electrical pulsesdecrease the conductance (data extracted from [23]), (e) example of different programmed states of the same nanotube network basedOG-CNTFET. The eight values of resistivity ranges from 5 M to 1.8 G.

    3. OG-CNTFET based crossbar architecture for

    neuromorphic computing

    3.1. Architecture

    Based on this gate protection effect we developed an original

    neuromorphic computing architecture. Instead of a single

    and fully global gate electrode, we propose to implement one

    gate electrode per row of devices in a crossbar geometry as

    represented in figure 4. In this configuration, all OG-CNTFETs

    within the same row share both the same gate electrode and

    the same source electrode. They correspond to synapses

    connected to the same neuron which provides the function

    Yi . The drain terminals (pre-synapse terminals) of all OG-

    CNTFETs within the same column are also connected together

    and carry the inputs signals (Xi ). The neurons needed here

    are simple CMOS based comparators. They receive, from the

    post-synapse terminals, a current (Isi ) which is directly the sum

    of the contributions from the different synapses and compare

    it to reference values to be targeted (Irefi ). The neurons can

    also apply a feedback signal (Vgi ) to the gate electrode of

    their own row of synapses. The operation of this architecture

    includes three phases: initializing, learning and computing. At

    first, light initializes globally all the OG-CNTFETs to their

    low resistance state. Then electrical pulses are applied to all

    the inputs (Xi ) to update the resistance of the OG-CNTFETs

    (or the weight of the synapses). At each neuron, the totalcurrent from a row (Isi ), measured between the programming

    pulses for given values of the inputs, progressively decreases

    toward Irefi , the known values to be learnt. Once this valueis reached in a given row, the neuron of this row sends a

    feedback signal to its corresponding shared gate electrode,

    which consists of a positive Vgi value. This activates the

    protection mechanism and prevents any further modification

    of all the resistivity values within this row. The incoming input

    signals keep modifying devices in the other rows until all the

    functions have been learnt. Once all the synaptic weights have

    been set to appropriate values, the network can be used for

    ultra-rapid parallel computing of multiple functions. In such

    a configuration of fixed weights, the circuit takes advantage

    of the exceptional transport properties of carbon nanotubes. It

    is indeed envisioned that carbon nanotube electronic deviceswould be much faster than devices based on conventional

    semiconductors [25, 26]. Recent experimental results tend

    to support these projections (see for example [27, 28] and

    references therein).

    The principal advantage of the gate protection effect is to

    allow the crossbar architecture to learn several functions in a

    massively parallel way. Indeed, input signals can be sent to the

    full array of nanodevices at the same time. Most importantly, as

    each row stops being modified when it has reached the targeted

    value, the total duration to learn a large number of functions is

    simply the duration of the learning phase of a single function:

    the one requiring the largest number of programming pulses.

    This implies a dramatic increase in the learning speed whencompared with conventional serial learning procedures.

    3

  • 7/31/2019 Toward Neuromorphic Computing

    5/8

    Nanotechnology 21 (2010) 175202 W S Zhao et al

    Figure 3. Evolution of source drain current flowing through an OG-CNTFET as a function of time in two different experimental conditions.(b) Corresponds to the case where VGS is maintained constant at 0 V. After a short illumination step that sets the device in a low resistancestate, positive input bias pulses (a) are applied to modify the device resistivity. (c) Corresponds to the same experiment but conducted atVGS = 6 V during the programming input pulses. This bias prevents any non-volatile change in resistivity. The programming currents duringthe VDS pulses (not visible in log scale) are in the 0.841.06 A range in (b) and 0.35 A in (c).

    3.2. Architecture simulation based on an electrical model of

    OG-CNTFET

    Based on experimental measurements, we developed a

    SPICE-like electrical model of OG-CNTFET [29] in the

    Cadence Spectre simulator, which is the main platform

    tool for integrated circuits development [30]. This model

    provides the interface to simulate OG-CNTFET based synaptic

    networks with CMOS based neurons and allows the testing

    of neuromorphic circuits. Figure 5 shows the simulation of a

    neural network composed of four neurons, with each neuron

    associated to 16 synapses (i.e. a total of 64 synapses while

    only 16 are represented in figure 4 for clarity). As mentioned

    above, there are three operating phases for this architecture:

    the first one is the initializing phase which consists in resetting

    all the synapses to their minimal weight. Three light pulses

    are used in the simulation to realize this task (in analogy with

    figure 2(d)) but a single longer pulse would be equivalently

    efficient. The second one is the learning phase during which

    the weights of the synapses are updated to target reference

    currents at different neurons. The last one is the computing

    phase.

    This simulation shows an example in which all the 16

    Xi values are set to 0.4 V between pulses and the targeted

    values for this specific configuration of the inputs are chosen

    as: Iref0 to Iref3 = 10 nA, 30 nA, 20 nA and 40 nA respectively

    (the reference current levels can be chosen at arbitrary valuesin the programming range which, in the present state of

    Figure 4. Crossbar architecture comprising four neurons (CMOSbased) and 16 synapses (OG-CNTFETs). Each neuron is associatedwith four synapses within a row which share both the same electricalgate (Vgi ) and the same source terminal. The drain terminals (verticalpre-synapse terminals) connect synapses from different neurons andserve to apply both the programming pulses during the learningphase and the input signals during the computing phase. All thesynapses in the array also share a common optical gate terminal: theP3OT film. The output of the functions (Yi ) and the feedback signalsto the gate (Vgi ) are produced by simple CMOS based neurons.

    4

  • 7/31/2019 Toward Neuromorphic Computing

    6/8

    Nanotechnology 21 (2010) 175202 W S Zhao et al

    Figure 5. Electrical simulation of OG-CNTFET based neuromorphic circuits comprising four neurons and 64 synapses (16 per neuron). Thereference currents to target in the example are respectively 10 nA, 30 nA, 20 nA and 40 nA. (a) Voltage pulses applied at the inputs (X0X3);

    in order to simplify the simulation, they are at the same magnitude and frequency for each input, which is not mandatory. ((b)(e)) Thecurrents Is at the post-synapse terminals, (f) voltage at the gate terminal of function Y2. It is initialized to 0 and is set to a high level by thefeedback from the comparator which then protects the relevant synapses from being updated.

    development of the technology, spans three [23] to four [21]

    orders of magnitude). This example highlights some important

    characteristics of the architecture, such as the parallel function

    learning (four functions Yi are learnt at the same time) and the

    total learning duration corresponding to the time necessary to

    learn the function Y0, which is the slowest one (i.e. the one

    requiring the largest number of programming pulses) in this

    example. The other functions Y1 to Y3, being ready before Y0,do not contribute to the total programming time.

    This model also includes a random parameter (the minimal

    resistivity of the device after illumination) which allows one to

    simulate the effect of variability (mismatch) in OG-CNTFETs.

    This parameter was chosen so as to reflect the variability in

    performances obtained from experimental measurements [23].

    The good operation of this simulation thus also demonstrates

    the high reliability of this neuromorphic architecture in the

    presence of device-to-device variability.

    It is important to note that in the developed model, the

    parameters were adjusted to correspond to experimental data

    acquired on non-scaled-down devices, such as those used

    in figure 3. The next generations of such circuits will be

    programmed with sub-s pulses using OG-CNTFETs with a

    scale-down gate oxide thickness (experimental evidence will

    be published separately). We should also note that the final

    currents at the end of the learning step are close to the reference

    currents, but with a moderate accuracy (94%). This is due

    to the insufficient number of synapses per neuron (16) in this

    example. When we associate each neuron with 100 synapses,

    the accuracy becomes >98%. An arbitrary high precision

    can be reached by appropriately scaling up the number of

    synapses per neuron. As the synapses will be built at the

    nanoscale, while the neuron will be CMOS based, this type of

    architecture would intrinsically be in a configuration of a largesynapses/neurons ratio compatible with high precision. With

    large numbers of synapses, parallel learning procedures, such

    as the one developed here, becomes a mandatory requirement.

    4. Discussion and conclusions

    4.1. Comparison with previous nanodevice based crossbar

    architectures

    Without the gate protection effect provided by the OG-CNTFET, this architecture would be similar to some nano-

    neuromorphic computing architectures which have been

    studied recently [17, 31, 32]. They are all based on crossbar

    structures and tunable two-terminals nanodevices as synapses,

    but use different methods to program the functions. Some of

    them update the relevant synapses and program the functions

    in series, as there is no efficient solution to protect the

    nanodevices associated with the programmed functions in the

    array [33]. This leads to very long learning times for large

    circuits. We think that a parallel learning capability is one of

    the essential requirements to build neuromorphic machines for

    practical applications, and promises higher performance thanCMOS based computing systems [34, 35]. For the crossbar

    architecture based on memristors, a parallel learning process

    could potentially be carried out, as their conductance would

    not be updated if the potential difference between their two

    terminals is lower than a given threshold voltage [17, 18].

    Therefore, one of the most efficient solutions would be to

    use the low threshold voltage (e.g. 0.7 V) of memristors to

    control the parallel learning [18, 35] or programming [36].

    However, this method leads to important cross-talk effects,

    as the threshold voltages of memristors could be different

    from each other and difficult to precisely control due to

    nanofabrication issues. For large crossbar architectures, these

    difficulties would be severe due to voltage drops along theconnecting wires. Another problem for these solutions is that

    5

  • 7/31/2019 Toward Neuromorphic Computing

    7/8

    Nanotechnology 21 (2010) 175202 W S Zhao et al

    each neuron requires a complex CMOS control circuit. This

    would enlarge greatly the final die area [18, 37].

    4.2. Expected performances (die area, reliability, learning

    speed and power consumption)

    Architecture design is one of the most crucial steps to buildhigh performance neural network circuits. It determines

    the power consumption, the die area and both the learning

    and the computing speed. In order to reduce the final die

    area and simplify the circuit initialization, the whole array

    of OG-CNTFETs shares the same global optical stimulus.

    In [23], we used a light spot of 200 m2. Such surface

    could accommodate a large number of scaled-down devices

    (see section 4.3 below). The proposed crossbar geometry is

    based only on shared electrodes: for the input, the output

    and the feedback signals. It thus minimizes the nanodevice

    interconnections and allows the development of a process

    with only two levels of metallization. Unlike for memristors

    in crossbar architectures, cross-talk problems do not occur

    between adjacent OG-CNTFETs during and after the function

    learning. Indeed, the protection is not based on a precise

    voltage value that could vary from device to device. The post-

    synapse terminal is always grounded and the FET channel

    is fully closed by the high positive bias applied on the gate

    electrodes of the protected devices. This allows a reliable

    parallel learning process even in very large arrays. As the gate

    electrodes are shared per neuron, the use of the third terminal

    will not substantially lower the device density. Furthermore

    they can be implemented vertically under the two-terminal

    programmable resistors (we showed recently that silicon wires

    can be used as shared gate electrodes for OG-CNTFETs, butthe description of this process flow and the associated electrical

    results are beyond the scope of this paper). Thereby this

    architecture also promises a low die area. Another important

    advantage of this architecture is the low standby power

    consumption during the learning phase. As the application of a

    positive Vg bias reduces the current down to nearly zero when

    the protection is activated (figure 3(c)), the power consumption

    saving for such neuromorphic circuits is optimized by the

    shutdown of the learning process per row as soon as the target

    level is reached. In addition, as the multi-level states of OG-

    CNTFETs are non-volatile, the circuit would also work with

    low standby power consumption during the computing phase.

    4.3. Scaling, speed and implementation perspectives

    The data presented in this paper were obtained on OG-

    CNTFETs based on nanotube networks (figure 2(c)). The main

    purpose was to illustrate both the programmability (figure 1(e))

    and the gate protection mechanism (figure 3), which are the

    basis of the studied architecture. For the simulations to be as

    realistic as possible, the data were also used to calibrate the

    models. OG-CNTFETs can be built using individual nanotubes

    and short channel lengths (for example, we used 100 nm

    in [21]). Several groups also demonstrated conventional

    nanotube transistors with channel length down to 10 nm [38]and with good performance down to 40 nm [39]. Recently,

    we used silicon wires as shared back-gate electrodes for OG-

    CNTFETs [40]. In this geometry, scaled-down devices could

    be integrated down to a conservative density of 1010 cm2.

    As a result, it is expected that the CMOS part used for the

    neurons would be the limiting factor in terms of die area. It

    is thus important to keep the functionality of this silicon based

    part very simple, which is the case in the studied architecture(neurons are CMOS comparators).

    The magnitude of the manipulated signals is also

    important for future implementations. With individual

    nanotubes, we achieved ON-currents above 1 A (at VDS =

    400 mV) and a programmability range of four orders of

    magnitude [21]. Optimized single nanotube transistors can

    drive currents as high as 20 A with Ion/Ioff ratio above 106.

    In a configuration such as the one illustrated in figure 2(b)

    based on nanotubes in parallel, higher current levels can be

    achieved without significantly increasing the device width.

    Recent progress in the CVD growth of aligned SWNTs show

    the potential of this configuration [41, 42].

    Concerning speed, our recent experimental results show

    that by scaling down the gate dielectric thickness to 2 nm,

    programming pulses shorter than 1 s can be used [40].

    These results are in preparation for publication. After the

    programming phase, the circuit would benefit from the intrinsic

    high speed of carbon nanotube electronics. In 2009, we

    showed, for example, that CNTFETs with cut-off frequencies

    as high as 80 GHz can be built [27].

    The physical implementation of an optical gate may be

    difficult to achieve, even though there is a very intense activity

    in the field of mixed electrical/optical electronics (note in this

    context that nanotube transistors can be used as both light

    detectors [43] and light sources [44]). We estimate that thestudied architecture could be implemented in a fully electrical

    way. Indeed, its principle relies on devices having two control

    signals. The only requirement for these two control signals

    is that one must be common to all the devices while the

    second one must be shared by devices within the same row.

    In the present implementation, light constitutes an efficient

    way to share a global stimulus. But this global signal could

    be a global back-gate electrode. In that case, the second one

    would be a shared top-gate electrode. With the recent progress

    in carbon nanotube based memory devices [45, 46], such an

    implementation can be reasonably envisaged.

    4.4. Conclusions

    In conclusion, we studied an original crossbar architecture to

    integrate carbon nanotube based programmable devices within

    a neuromorphic architecture. An experimental demonstration

    of a mechanism that allows protecting already programmed

    devices from further modification was provided. This

    mechanism allows very efficient parallel learning of multiple

    functions within large arrays. This hybrid nano/CMOS

    architecture promises high reliability, high density and fast

    learning speed, which are the crucial requirements to building

    neuromorphic computing machines for practical applications

    and to providing performances exceeding those of presentcomputing systems. Using a precise functional model

    6

  • 7/31/2019 Toward Neuromorphic Computing

    8/8

    Nanotechnology 21 (2010) 175202 W S Zhao et al

    of OG-CNTFETs, we simulated this crossbar architecture

    and highlighted the principal characteristics of its expected

    performances.

    Acknowledgments

    The author thanks B Linares-Barranco and J-O Klein forfruitful discussions, P Chenevier and S Campidelli for help

    with nanotube purification, S Lenfant and D Vuillaume for

    important technological help and scientific input. This work

    was partially funded by the European Union through the FP7

    Project Nabab (Contract FP7-216777), by the French National

    Research Agency through the Panini Project (ANR-07-ARFU-

    008) and by the Region Ile-de-France for the nanofabrication

    facility at CEA-SPEC.

    References

    [1] Kim N S, Austin T, Blaauw D, Mudge T, Krisztian F, Hu J S,Irwin M J, Kandemir M and Narayanan V 2003 IEEE Trans.Comput. 16 68

    [2] Avouris P, Chen Z and Perebeinos V 2007 Nat. Nanotechnol.2 605

    [3] Li Y, Qian F, Xiang J and Lieber C M 2006 Mater. Today 9 18[4] Thelander C et al 2006 Mater. Today 9 28[5] Geim A K and Novoselov K S 2007 Nat. Nanotechnol. 6 183[6] Joachim C, Gimzewski J K and Aviram A 2000 Nature

    408 541[7] Lu W and Lieber C M 2007 Nat. Mater. 6 841[8] Gurney K 1997 An Introduction to Neural Networks (London:

    Routledge) ISBN 1-85728-673-1[9] Rosenblatt F 1958 Psychol. Rev. 65 386

    [10] Strukov D B, Snider G S, Stewart D R and Williams R S 2008Nature 453 80

    [11] Lai Q, Li Z, Zhang L, Li X, Stickle W F, Zhu Z, Gu Z,Kamins T I, Williams R S and Chen Y 2008 Nano Lett.8 876

    [12] Alibart F, Pleutin S, Guerin D, Novembre C, Lenfant S,Lmimouni K, Gamrat C and Vuillaume D 2010 Adv. Funct.

    Mater. 20 330[13] Strukov D B and Likharev K K 2005 Nanotechnology 16 888[14] Stan M R, Franzon P D, Goldstein S C, Lach J C and

    Ziegler M M 2003 Proc. IEEE 91 1940[15] DeHon A 2003 IEEE Trans. Nanotechnol. 2 23[16] Snider G S and Williams R S 2007 Nanotechnology 18 035204[17] Snider G S 2008 Molecular-junction-nanowire-crossbar-based

    neural networkUS Patent Specification 7359888

    [18] Snider G S 2008 Proc. IEEE/ACM Int. Symp. on NanoscaleArchitectures pp 8592

    [19] Jo S H, Kim K H and Lu W 2009 Nano Lett. 9 870[20] Xia Q et al 2009 Nano Lett. 9 3640

    [21] Borghetti J, Derycke V, Lenfant S, Chenevier P, Filoramo A,Goffman M, Vuillaume D and Bourgoin J-P 2006 Adv.

    Mater. 18 2535[22] Anghel C, Derycke V, Filoramo A, Lenfant S, Giffard B,

    Vuillaume D and Bourgoin J-P 2008 Nano Lett. 8 3619[23] Agnus G, Zhao W S, Derycke V, Filoramo A, Lhuillier Y,

    Lenfant S, Vuillaume D, Gamrat C and Bourgoin J-P 2009Adv. Mater. 21 702

    [24] Derycke V et al 2009 C. R. Physique 10 330[25] Burke P 2004 Solid State Electron. 48 1981[26] Guo J, Hasan S, Javey A, Bosman G and Lundstrom M 2005

    IEEE Trans. Nanotechnol. 4 715[27] Kocabas C et al 2009 Nano Lett. 9 1937[28] Nougaret L, Happy H, Dambrine G, Derycke V, Bourgoin J P,

    Green A A and Hersam M C 2009 Appl. Phys. Lett.94 243505

    [29] Zhao W S, Agnus G, Derycke V, Filoramo A, Gamrat C andBourgoin J-P 2009 Proc. IEEE Int. Conf. Nanotechnologypp 8736

    [30] Virtuoso Advanced Analysis Tools User Guide 2007 Cadence5.1.41 Spectre Guide http://www.cadence.com

    [31] Cavin R, Huntchby J A and Zhirnov V 2008 Computer41 33[32] Nugent A 2006 Utilized nanotechnology apparatus using a

    neural network, a solution and a connection Gap US PatentSpecification 7039619

    [33] He M, Klein J O and Belhaire E 2008 Electron. Lett. 44 575[34] Kramer A H and Sangiovanni-Vincentelli A 1989 Efficient

    parallel learning algorithms for neural networks Advances inNeural Information Processing Systems vol 1(San Francisco, CA: Morgan Kaufmann Publishers Inc.)ISBN:1-558-60015-9

    [35] Klein J O and Belhaire E 2009 French Patent0705532[36] Ziegler M M and Stan M R 2003 IEEE Trans. Nanotechnol.

    2 217[37] Pershin Y and Di Ventra M 2009 arXiv:0905.2935v2[38] Chiu P W, Chen K J and Lin C Y 2008 Appl. Phys. Lett.

    92 152111[39] Lin Y-M, Appenzeller J, Chen Z, Chen Z-G, Cheng H-M and

    Avouris P 2005 IEEE Electron Device Lett. 26 823[40] Agnus G, Filoramo A, Lenfant S, Vuillaume D, Bourgoin J-P

    and Derycke V 2010 High-speed programming of carbonnanotube based optoelectronic memory devices using siliconnanowires as gate electrode, unpublished

    [41] Kang S J, Kocabas C, Ozel T, Shim M, Pimparkar N,Alam M A, Rotkin S V and Rogers J A 2007 Nat.

    Nanotechnol. 2 230[42] Ding L, Tselev A, Wang J, Yuan D, Chu H, McNicholas T P,

    Li Y and Liu J 2009 Selective growth of well-alignedsemiconducting single-walled carbon nanotubes Nano Lett.9 800

    [43] Freitag M, Martin Y, Misewich J A, Martel R and Avouris P2003 Nano Lett. 3 1067

    [44] Misewich J A, Martel R, Avouris P, Tsang J C, Heinze S and

    Tersoff J 2003 Science 300 783[45] Rinki M, Johansson A, Paraoanu G S and Torma P 2009 Nano

    Lett. 9 643[46] Fu W, Xu Z, Bai X, Gu C and Wang E 2009 Nano Lett. 9 921

    7

    http://dx.doi.org/10.1038/nnano.2007.300http://dx.doi.org/10.1038/nnano.2007.300http://dx.doi.org/10.1016/S1369-7021(06)71650-9http://dx.doi.org/10.1016/S1369-7021(06)71650-9http://dx.doi.org/10.1016/S1369-7021(06)71651-0http://dx.doi.org/10.1016/S1369-7021(06)71651-0http://dx.doi.org/10.1038/35046000http://dx.doi.org/10.1038/35046000http://dx.doi.org/10.1038/nmat2028http://dx.doi.org/10.1038/nmat2028http://dx.doi.org/10.1037/h0042519http://dx.doi.org/10.1037/h0042519http://dx.doi.org/10.1038/nature06932http://dx.doi.org/10.1038/nature06932http://dx.doi.org/10.1021/nl073112yhttp://dx.doi.org/10.1021/nl073112yhttp://dx.doi.org/10.1002/adfm.200901335http://dx.doi.org/10.1002/adfm.200901335http://dx.doi.org/10.1088/0957-4484/16/6/045http://dx.doi.org/10.1088/0957-4484/16/6/045http://dx.doi.org/10.1109/JPROC.2003.818327http://dx.doi.org/10.1109/JPROC.2003.818327http://dx.doi.org/10.1109/TNANO.2003.808508http://dx.doi.org/10.1109/TNANO.2003.808508http://dx.doi.org/10.1088/0957-4484/18/3/035204http://dx.doi.org/10.1088/0957-4484/18/3/035204http://dx.doi.org/10.1021/nl8037689http://dx.doi.org/10.1021/nl8037689http://dx.doi.org/10.1021/nl901874jhttp://dx.doi.org/10.1021/nl901874jhttp://dx.doi.org/10.1002/adma.200601138http://dx.doi.org/10.1002/adma.200601138http://dx.doi.org/10.1021/nl801543khttp://dx.doi.org/10.1021/nl801543khttp://dx.doi.org/10.1002/adma.200802692http://dx.doi.org/10.1002/adma.200802692http://dx.doi.org/10.1016/j.crhy.2009.05.006http://dx.doi.org/10.1016/j.crhy.2009.05.006http://dx.doi.org/10.1016/j.sse.2004.05.044http://dx.doi.org/10.1016/j.sse.2004.05.044http://dx.doi.org/10.1109/TNANO.2005.858601http://dx.doi.org/10.1109/TNANO.2005.858601http://dx.doi.org/10.1021/nl9001074http://dx.doi.org/10.1021/nl9001074http://dx.doi.org/10.1063/1.3155212http://dx.doi.org/10.1063/1.3155212http://www.cadence.com/http://dx.doi.org/10.1109/MC.2008.155http://dx.doi.org/10.1109/MC.2008.155http://dx.doi.org/10.1049/el:20080442http://dx.doi.org/10.1049/el:20080442http://dx.doi.org/10.1109/TNANO.2003.820804http://dx.doi.org/10.1109/TNANO.2003.820804http://arxiv.org/abs/0905.2935v2http://arxiv.org/abs/0905.2935v2http://dx.doi.org/10.1063/1.2912128http://dx.doi.org/10.1063/1.2912128http://dx.doi.org/10.1109/LED.2005.857704http://dx.doi.org/10.1109/LED.2005.857704http://dx.doi.org/10.1038/nnano.2007.77http://dx.doi.org/10.1038/nnano.2007.77http://dx.doi.org/10.1021/nl803496shttp://dx.doi.org/10.1021/nl803496shttp://dx.doi.org/10.1021/nl034313ehttp://dx.doi.org/10.1021/nl034313ehttp://dx.doi.org/10.1126/science.1081294http://dx.doi.org/10.1126/science.1081294http://dx.doi.org/10.1021/nl8029916http://dx.doi.org/10.1021/nl8029916http://dx.doi.org/10.1021/nl801656whttp://dx.doi.org/10.1021/nl801656whttp://dx.doi.org/10.1021/nl801656whttp://dx.doi.org/10.1021/nl8029916http://dx.doi.org/10.1126/science.1081294http://dx.doi.org/10.1021/nl034313ehttp://dx.doi.org/10.1021/nl803496shttp://dx.doi.org/10.1038/nnano.2007.77http://dx.doi.org/10.1109/LED.2005.857704http://dx.doi.org/10.1063/1.2912128http://arxiv.org/abs/0905.2935v2http://dx.doi.org/10.1109/TNANO.2003.820804http://dx.doi.org/10.1049/el:20080442http://dx.doi.org/10.1109/MC.2008.155http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://www.cadence.com/http://dx.doi.org/10.1063/1.3155212http://dx.doi.org/10.1021/nl9001074http://dx.doi.org/10.1109/TNANO.2005.858601http://dx.doi.org/10.1016/j.sse.2004.05.044http://dx.doi.org/10.1016/j.crhy.2009.05.006http://dx.doi.org/10.1002/adma.200802692http://dx.doi.org/10.1021/nl801543khttp://dx.doi.org/10.1002/adma.200601138http://dx.doi.org/10.1021/nl901874jhttp://dx.doi.org/10.1021/nl8037689http://dx.doi.org/10.1088/0957-4484/18/3/035204http://dx.doi.org/10.1109/TNANO.2003.808508http://dx.doi.org/10.1109/JPROC.2003.818327http://dx.doi.org/10.1088/0957-4484/16/6/045http://dx.doi.org/10.1002/adfm.200901335http://dx.doi.org/10.1021/nl073112yhttp://dx.doi.org/10.1038/nature06932http://dx.doi.org/10.1037/h0042519http://dx.doi.org/10.1038/nmat2028http://dx.doi.org/10.1038/35046000http://dx.doi.org/10.1016/S1369-7021(06)71651-0http://dx.doi.org/10.1016/S1369-7021(06)71650-9http://dx.doi.org/10.1038/nnano.2007.300