ArticlePDF Available

Abstract and Figures

In spite of the increasing use of machine learning techniques, in-memory computing and hardware have increased the interest to accelerate neural network operation. Henceforth, novel embedded nonvolatile memories (eNVMs) for highly scaled technology nodes, like ferroelectric field effect transistors (FeFETs), are heavily studied and very promising. Furthermore, inference and on-chip learning can be fostered by further eNVM technology options, such as multibit operation and linear switching. In this article, we present the advantages of hafnium oxide-based FeFETs for such purposes due to their basic three-terminal structure, which allows to selectively activate or deactivate selected devices as well as tune linearity and dynamic range for certain applications. Furthermore, we discuss the impact of the material properties of the ferroelectric layer, the interface layer thickness, and scaling on the device performance. Here, we demonstrate good device properties even for highly scaled devices (100 nm x 100 nm).
Content may be subject to copyright.
Ferroelectric field effect transistors as a synapse
for neuromorphic application
M. Lederer, T. K¨
ampfe Member, IEEE, T. Ali, F. M¨
uller, R. Olivo, R. Hoffmann, N. Laleni and K. Seidel
AbstractIn spite of the increasing use of machine
learning techniques, in-memory computing and hardware
has raised interest to accelerate neural network operation.
Hereforth, novel embedded non-volatile memories (eNVM)
for highly scaled technology nodes, like ferroelectric field
effect transistors (FeFETs), are heavily studied and very
promising. Furthermore, inference and on-chip learning
can be fostered by further eNVM technology options, such
as multi-bit operation and linear switching. In this article
we present the advantages of hafnium oxide based FeFETs
for such purposes due to its basic three terminal structure,
which allows to selectively activate or deactivate selected
devices as well as tune linearity and dynamic range for
certain applications. Furthermore, we discuss the impact
of the material properties of the ferroelectric layer, the
interface layer thickness and scaling on the device perfor-
mance. Here, we demonstrate good device properties even
for highly scaled devices (100 nm x 100 nm).
Index Termsferroelectric, FeFET, hafnium oxide, non-
volatile memory, neuromorphic hardware, synapse
DUE to the von-Neumann bottleneck for memory-
intensive algorithms such as approximate computing the
demand of non-volatile memories has rapidly increased [1].
New architectures like near- or in-memory computing have
gained much attention and multiple new non-volatile memory
devices, like resistive random access memory (RRAM), phase-
change random access memory (PCRAM) or devices based on
ferroelectric HfO2, e.g. ferroelectric tunnel junctions (FTJs)
or ferroelectric field effect transistors (FeFETs), have been
suggested as suitable devices for storing the weight infor-
mation or acting as neuron [1]–[6]. For the implementation
of these devices in large, reliable and power efficient neural
networks, they must satisfy multiple requirements [2], [7].
Besides large on/off conductance ratio and symmetric weight
update, recent studies [7] have highlighted the importance of
device variability, endurance and retention.
This paragraph of the first footnote will contain the date on which
you submitted your paper for review. This research was funded by
the ECSEL Joint Undertaking project TEMPO in collaboration with the
European Union’s Horizon 2020 Framework Program for Research and
Innovation (H2020/2014-2020) and National Authorities, under Grant
No. 826655, and by the German Bundesministerium f¨
ur Wirtschaft
(BMWi), by the State of Saxony in the frame of the Important Project
of Common European Interest (IPCEI).
M. Lederer, T. K ¨
ampfe, T. Ali, F. M¨
uller, R. Olivo, R. Hoffmann,
N. Laleni and K. Seidel are with the Fraunhofer IPMS, Center Na-
noelectronic Technologies (CNT), Dresden 01109, Germany (e-mail:
As mentioned before, devices based on ferroelectric HfO2
have been suggested as possible candidates [2], [4], [5], [8]–
[11]. Due to its compatibility to complementary-metal-oxide-
semiconductor (CMOS) processes, high coercive field and
persistent ferroelectricity for ultra-thin films, this material can
easily be integrated in current technology nodes [12]. Non-
volatile memory devices, especially ferroelectric field effect
transistors (FeFETs), have been already demonstrated in 28 nm
and 22 nm high-k-metal-gate (HKMG) CMOS technology
nodes [13], [14]. Hereby, Si- and Zr-doped HfO2(HSO/HZO)
have been widely used [15], [16]. Furthermore, a multitude
of other dopants, like La, Y, or Al, can be used to stabilize
the ferroelectric phase [12], as well as stress configurations,
which promote ferroelectricity even in undoped films [17].
In this article, the main differences between HfO2FeFETs
and common resistive synapses are discussed. Furthermore, the
impact of gate voltage during read-out as well as the signal
sequence on the weight update of the device and influences of
the device integration in regards of the ferroelectric layer are
elaborated. Finally, figures of merit with regard of endurance,
retention, and scaling are discussed. Here, good linear behavior
and dynamic range (DR) were demonstrated for a 100 nm x
100 nm device. Thus, demonstrating the advantage of HfO2
based FeFETs for hardware accelerated neural networks and
their readiness for application in current technology.
The ferroelectric layers of the FeFETs were produced using
atomic layer deposition (ALD) utilizing chlorine based pre-
cursors with a thickness of 10 nm on SiO2or SiON interface.
For the investigation on layer thickness influences, 5 nm films
were prepared additionally. In case of Si- and Zr-doped HfO2
(HSO and HZO), a cycling ratio of 16:1 and 1:1 was used,
respectively. A 10 nm TiN film deposited by physical vapor
deposition (PVD) was used as capping layer. More details on
the full integration flow can be found elsewhere [18]. The
highly scaled embedded non-volatile (eNVM) devices were
produced by using a non-invasive eNVM process [13]. For
the polarization hysteresis of the ferroelectric layer, a 10 nm
bottom TiN electrode was deposited via ALD, followed by
the aforementioned deposition of HSO or HZO. The layer
was then capped by a 10 nm PVD TiN top electrode and
crystallized by rapid thermal annealing at 800C. Capacitor
structures were formed by sputtering Ti/Pt using a shadow
mask and subsequent wet etch.
For electrical analysis was conducted utilizing an automated
waferprober and precision semiconductor analyzer. Drain volt-
Fig. 1. Schematic layout of a two terminal (a) and three terminal (b) device using a ferroelectric layer. Different remanent polarization states of
a ferroelectric film can be addressed by returning at a certain voltage in a triangular waveform (c). FeFETs (three terminal device) can be poled
almost continues by pulsing between the two extreme states (d). A possible crossbar layout using FeFETs is presented in (e).
age was set to 100 mV for all measurements. Pulse sequences
with amplitudes ranging from 2.5 V to 4.5 V and widths from
50 ns to 300 ns (10 ns steps) were tested. The signal inves-
tigation was carried out on devices with a width and length
of 25 µm each. Electrical characterization for investigating
dopant, pulse sequences with varying width (50 ns to 300 ns)
were used with an amplitude of 3.5 V and -4 V for potentiation
and depression, respectively. For investigating the thickness
influence, the amplitudes were increased to 4 V and -4.5V,
respectively. Devices under test had a width of 20 µm and
length of 15 µm.
Retention measurements for the intermediate states were
performed by setting the state using a pulse of varying
amplitude and measuring transfer characteristics after certain
amounts of time. For cycle-to-cycle variability, a pulse train
with varying amplitude (from 2.5 V to 4 V for program and
from -3 V to -4.5 V for erase with a width of 100 ns) was used
repeating up to 30 times on a FeFET with width and length
dimensions of 10 µm each. For investigating the device-to-
device variability, 20 devices with a channel width and length
of 5 µm each were measured using a pulse sequence with
varying amplitude (from -3 V to -4.55 V with a width of
100 ns).
Ferroelectric HfO2can be used in two different design
schemes for embedded non-volatile memories. Similar to
resistive memories, it can be used in a two terminal device
configuration (see Figure 1a). This design concept is usu-
ally applied by connecting the ferroelectric capacitor to the
drain of a select transistor in conventional ferroelectric RAM
(FeRAM). This concept results in a destructive read-out of
the polarization [19]. Applying this two terminal device in
a resistive crossbar does not promise good properties, as the
tunnel electroresistance (TER) shift due to the local electric
field resulting from the polarization state as in ferroelectric
tunnel junctions (FTJs) is very low and leakage current is
likely to deteriorate the material.
An alternative structure is the FeFET, which is a three
terminal device. Here, the ferroelectric material is integrated
in the gate stack of a field effect transistor (see Figure 1b). An
equivalent device can be formed, by connecting a ferroelectric
capacitor to the gate contact of a transistor, as used for
back-end of line (BEoL) integration [20]. A three terminal
device offers many advantages over two terminal devices.
Firstly, three terminal configuration enables a decoupling of
writing and reading paths. Secondly, the gate contact allows
the deactivation of selected devices as well as the application
of certain inhibit voltages for selective programming in an
array configuration. Thirdly, due to a capacitive coupling of the
polarization state and the surface potential of the channel, high
on/off ratios can be achieved. These and other functionalities
resulting from the third terminal, will be discussed later on.
Furthermore, due to the FeFETs similarity to high-k metal
gate transistors, these devices can be easily co-integrated into
highly scaled technology nodes [21].
Stable intermediate states, which are necessary for
e.g. multi-bit precision based hardware accelerated neural-
networks, can be formed by different polarization states of the
ferroelectric material [23]. While the unit cell allows only a
very limited and finite number of orientations of the polariza-
tion axis with a fixed remanent polarization, domain structure
as well as differences in the crystallographic orientation of
grains enable an almost continuous number of these states.
Figure 1c shows the different remanent polarization states,
which are achieved by applying different voltage amplitudes
using a triangular wave form to a ferroelectric capacitor.
As illustrated in Figure 1d, the polarization states in a
Fig. 2. Alteration in the current distribution of the intermediate states in dependence of the gate voltage for FeFETs with a 10 nm HSO layer
and SiO2interface layer. (a) shows the transfer characteristics of the intermediate states. By extracting the currents for certain gate voltages and
normalizing them to the local current maximum, the resulting current state distributions for depression (b) can be visualized, analogously potentiation
can be analyzed. As shown in (b), the three types of sequences result in curves being very linear in case of seq. 2 (at -3.5 V and 50 ns to 250 ns
with 10 ns increment) and seq. 3 (at 100 ns and -3 V to -4.5 V with 0.05 V increment), whereas a quite non-linear behavior is found seq. 1 (at
-4.5 V and 100 ns). For quantitative analysis, a non-linearity coefficient [2], [22] was introduced. Respective curves for certain non-linearity values
are shown in (c). The opposing trend of linearity in dependence of the gate voltage (VG) can be seen clearly in for potentiation and depression.
Furthermore, dynamic range shows a gate voltage dependence (d), here shown for sequence 2 with pulse amplitudes of 3.5 V/-4 V. Reducing the
pulse width (e) or amplitude (f) can improve the linearity.
FeFET can be addressed by applying multiple pulses, e.g. of
the same shape. By changing the polarity of the pulse, the
polarization can be inverted. Nevertheless, the device can also
be fully programmed or erased by applying a single pulse with
a sufficient amplitude.
For integrating this device in a crossbar structure, an ad-
ditional line connected to the gate terminal of the devices is
necessary. A possible crossbar schematic is shown in Figure
1e. Consequently, the write signal, which is addressed to the
gate terminal, and the input signal, addressed to the source
terminal, during read operation are physically separated and
A. Impact of gate voltage on pulse-switching
Furthermore, the intermediate states in a FeFET can be
addressed by applying rectangular pulse sequences. When
measuring transfer characteristics after each pulse, an almost
continuous shift in the threshold voltage over a broad range
can be observed (see Figure 2a). From this, a high dynamic
range greater than four orders of magnitude, limited only to the
transistor characteristics at suitable VG, as well as the ability
to switch off devices completely for low enough gate voltages
can be deduced.
Due to the third terminal, linearity and dynamic range are
influenced both by the gate voltage (VG) during readout and
by the signal sequence shape applied during write operation.
Previously, three different pulse schemes (see Figure 2b) were
discussed to program the states in a multi-level memory cell
[2]. The easiest one (sequence 1) is the repetition of the same
pulse multiple times. Alternatively, the pulse width (sequence
2) or the pulse amplitude (sequence 3) can be increased with
the pulse count.
The latter is in terms of theory straightforward for fer-
roelectrics, as only domains with an effective coercive field
smaller or equal to the voltage drop across the ferroelectric can
switch, like discussed for Figure 1c. The reason sequence 2
can be used for multi-state switching is attributed to nucleation
limited switching (NLS) present in hafnium oxide based films
On the contrary, sequence 1 is not expected to result in
a very linear behavior. As the pulse time and amplitude
remain constant, neither NLS nor classical switching propose
the addressability of multiple states. Origins for addressing
different states with this sequence could be explained by
changes in the voltage drop across the ferroelectric due to the
changes in the polarization and therefore local electrical field.
On the other hand, a neuron like integrate and fire scheme
was reported in literature for seq. 1 and has been referred to
as accumulated switching [25].
As shown for depression, the three sequences (with com-
parable signal parameters) reveal a clear trend to more linear
curves for seq. 1 to 3 at a gate voltage of 0.8 V (see Figure
2b). In order to quantitatively compare the measured behavior,
a fitting model proposed by Yu et al. based on the equations 1
and 2, where P is the pulse number and A, B the fit parameters,
was used to extract a non-linearity coefficient α[2], [22].
Additionally, as it can be already deduced from Figure 2a,
changes in the current distribution of the intermediate states
will differ in dependence of the gate voltage due to the
nonlinear behavior of the transfer characteristic.
ID=B(1 eP
(A+ 0.162) (2)
As shown in Figure 2c for seq. 1, the fitting of each
distribution results in clear dependence of linearity on the gate
voltage. Here, potentiation and depression show very contrary
trends, which is a result of the asymmetric switching behavior
of FeFETs, regarding programming and erasing, combined
with the shape of transfer characteristics. While potentiation
shows an increase of non-linearity with higher gate voltages,
depression shows strong reduction. In both cases, values tend
to saturate for higher gate voltages. Similar trends are observed
for the other sequences.
Combined with the average conductance change and on/off
ratio (see Figure 2d), it allows to optimize the operating
conditions of the crossbar for a specific neural-network ap-
plication. Either it can be optimized for a maximum of on/off
ratio by choosing an appropriate gate voltage, or for a more
symmetric behavior. Furthermore, a change in the threshold
voltage targeting allows to further tune the device.
As already depicted in Figure 2e, this behavior can be
further tuned by adapting the pulse’s amplitude and width.
In contrast to seq. 1, seq. 2 and seq. 3 can only modulate am-
plitude or width, respectively. As shown in Figure 2e, shorter
pulses result in an increase of linearity for potentiation and
depression, with changes being more pronounced in the latter.
Analogously, a lower amplitude shifts also both directions to a
lower non-linearity coefficient (see Figure 2f). This trend has
been found consistent for all three sequences. Thus allowing to
achieve also low non-linearity in seq. 1 based pulse schemes.
Consequently, improving the signal for high writing speed
as well as for low power consumption falls in line with the
optimization for a more linear behavior. Furthermore, a high
dynamic range larger than three magnitudes can be achieved
by adapting gate voltage, substrate doping and writing signal.
Power consumption during write operation is estimated by
simulation to be in a range of 10 fJ to 6 fJ for pulse
amplitudes in the range of 4 V to 2 V, respectively. This is
significantly lower compared to RRAM and PCRAM, with a
power consumption of around 0.1 pJ and 10 pJ, respectively
[26]. As the requirements for on-chip training and interference
are quite different, the implications for those two are discussed
separately in the following paragraphs.
As the linearity of writing plays a reduced role in pure
interference mode, the focus lies here on dynamic range.
Therefore, devices should be optimized in the gate voltage for
the maximum in on/off ratio. Additionally the write signals can
be improved for increased on/off ratio, even if this results in
a slight decreases in linearity. Furthermore, the devices can
be operated in unipolar mode, if this improves the figures
of merit. Due to the absence of on-chip training in pure
interference mode, the requirements on variability of the
devices also increase.
On the contrary, the requirements on linearity and symmetry
become much more stringent, if on-chip training is included.
Therefore, the long term potentiation (LTP) and depression
(LTD) signals should be optimized for symmetry, while still
having a linear behavior. Consequently, bipolar mode should
be used. Dynamic range is still not negligible, as interference
will be performed afterwards. The requirements on variability
become more loose, as the training will adapt the weights
depending on the device properties. Regarding the writing
sequence, seq. 1 would be favorable as it simplifies the signal
circuits. Seq. 2 should also be possible to integrate in an easy
manner, thus enabling stronger improvements on the Figure of
merits. On the contrary, seq. 3 is expected to be quite difficult
to integrate due to changing voltages.
To judge whether the multi-bit states of FeFET technology
are complying with the requirements of implementation into
resistive crossbars for neural netowrk applications, additional
Figures of merit in context of reliability have to be taken into
account. As seen in Figure 3a, retention measurements of the
intermediate states were conducted up to 5×103s. We observe,
that after a short initial relaxation, the intermediate states
retain stable. This initial relaxation originates from electron
detrapping, which is especially pronounced and takes slightly
longer for the high current state due to more electron trapping
during the larger programming pulse.
Furthermore, we observed very good cycle-to-cycle variabil-
ity (see Figure 3b). Additionally, results suggest a good device-
to-device variability, which is mainly influenced by saturation
current variability and therefore not inherent to the switching
behavior of the ferroelectric layer, as can be seen in Figure 3c.
The resuts in 3b furthermore suggest a high symmetry between
the potentiation and depression branch while preserving a
high linearity, which eases technical realizations. A ratio of
minimum and maximum resistance of >103can be achieved,
as well as linearly switched 32 states (5-bit). The device-to-
device variation was investigated and only a small variation
was observed, the cycle-to-cycle-variation is limited as was
shown in 3b. The endurance is furthermore expected to be
improved upon sub-loop operation and hence even could
comply with on-chip learning needs which are expected to
be in the order of 109cycles.
B. Cell design considerations
The weight update behavior is further influenced by the
device integration of the FeFET gate stack, consisting of the
gate material, interface layer, ferroelectric and top electrode.
Influences of different materials, thicknesses, and process
related parameters on the device properties in respect to
binary operation have already been addressed elsewhere [18],
[27]–[30]. Here, influences of the material properties, layer
thickness as well as impacts resulting from scaling on the
analog switching properties of the device will be discussed in
the following paragraphs.
First, the impact of different dopants on the device were
investigated using HSO and HZO. Dynamic hysteresis mea-
surements of these two materials already show strong dif-
ference (see Figure 4a). While the HSO shows a higher
Fig. 3. Reliability of FeFETs in terms of synaptic devices. Retention measurements of intermediate states (a) in HSO SiO2FeFETs indicate very
stable states up to 105s. Furthermore, very linear and stable operation for repeated cycles is found (b) for this material stack, thus having a very
low cycle-to-cycle variability. Comparison of 20 devices (HZO FeFETs with SiON interface) further suggests low device-to-device variability (c).
Fig. 4. Differences between HZO and HSO based devices with SiON interface and impact of scaling. Polarization hysteresis (a) shows lower
coercive field but higher remanent polarization for HSO. 10 nm FeFETs based on HZO show lower nonlinearity coefficient (b) compared to HSO
based devices with a film thickness of 10 nm. However, the dynamic range (c) of HSO based devices is higher. Impact of thickness scaling of HZO
SiON FeFETs on the nonlinearity coefficient is shown in (d). Intermediate states (e) are still observable in highly scaled HSO devices (100nm x
100nm) using varying amplitudes from -2 V to -4 V (0.05 V increment) at 200 ns. Furthermore, good linearity is achievable (f).
remanent polarization (PR), the HZO film shows a higher
coercive field (Ec) with a broader distribution. Consequently,
differences in the non-linearity coefficient as well as on/off
ratio are observable (see Figure 4b and 4c, respectively).
The lower non-linearity is in very good agreement with the
expectations, as the ratio of signal amplitude to coercive field
(Vmax/EcdF E ) is much lower. Furthermore, the broader
Ecdistribution supports a more pronounced subloop behavior.
The resulting drawback on the other hand is the reduced on/off
ratio, resulting analogously from the Vmax/EcdF E ratio.
This can be of course counteracted by increasing the applied
voltage. This would also result in utilizing a larger portion
of the memory window, which is expected to be larger for
HZO devices [31]. Consequently, HZO devices would enable
a larger dynamic range and better non-linearity along the
drawback of increased voltage amplitudes during LTP/LTD.
The reduced dynamic range of HSO FeFETs with SiON
interface (Fig. 4c) compared to HSO devices with SiO2
interface (Fig. 2d) is originates from the difference in the
voltage divider resulting from the interface and ferroelectric
capacitance, thus changing the electric field drop across the
ferroelectric layer for identical pulse conditions.
Changes to a larger thickness of the ferroelectric layer result
in a stretching of the transfer characteristics due to the change
in electrical field across the ferroelectric and semiconductor.
Consequently, a stretching of the characteristic curves for non-
linearity (see Figure 4d) and dynamic range.
Furthermore, a strong impact with scaling to very small de-
vice areas is expected, as the total amount of grains inside the
layer decreases drastically [32]. Nevertheless, a high number
of intermediate states can be observed in devices as small as
100x100 nm2(see Figure 4e). Furthermore, very linear and
analog like switching can be observed while still exhibiting
a high dynamic range, as seen in Figure 4f. Additionally,
zero width domain walls have been suggested to exist in
ferroelectric hafnium oxide due to flat phonon bands and
thus enabling domain sizes as small as single unit cells [33].
This stresses the possibility of analog operating FeFETs even
in strongly scaled devices and further optimizations in the
material processing for small grain sizes or orientation control
should improve this behavior.
In summary, we demonstrated that hafnium oxide based
FeFETs offer a versatile device, which can easily be optimized
for specific applications like hardware accelerated neural-
networks by simply changing signals or the gate voltage.
Furthermore, by adapting the pulse scheme, not only seq.
3 (voltage variation) but also seq. 2 (pulse width variation)
and seq. 1 (identical pulses) can achieve a very linear current
distribution of the addressed states. Due to the third terminal it
is furthermore possible to deactivate selected devices, allowing
very dense crossbars, as only 1T cells are required. The power
consumption of these devices are also very low due to the
large on/off ratio and low read currents. In point of design
considerations, material improvements to larger memory win-
dows or a less steep slope at Ec, resembling a broader domain
orientation distribution, were demonstrated as being favorable.
Furthermore, smaller grains are preferable for further device
scaling. Consequently, microstructure engineering of grain size
and crystallographic orientation will be of major importance
for ferroelectric synaptic devices. In addition, changes in the
layer thickness of the ferroelectric film allow to improve the
device for specific applications.
[1] G. W. Burr, R. M. Shelby, A. Sebastian, S. Kim, S. Kim, S. Sidler,
K. Virwani, M. Ishii, P. Narayanan, A. Fumarola, L. L. Sanches,
I. Boybat, M. Le Gallo, K. Moon, J. Woo, H. Hwang, and Y. Leblebici,
“Neuromorphic computing using non-volatile memory,Adv. Phys. X,
vol. 2, no. 1, pp. 89–124, 2017, doi:10.1080/23746149.2016.1259585.
[2] M. Jerry, P.-Y. Chen, J. Zhang, P. Sharma, K. Ni, S. Yu, and S. Datta,
“Ferroelectric fet analog synapse for acceleration of deep neural net-
work training,” in IEEE International Electron Devices Meeting, ser.
IEDM. San Francisco, CA, USA: IEEE, 2017, pp. 6.2.1–6.2.4,
[3] G. W. Burr, P. Narayanan, R. M. Shelby, S. Sidler, I. Boybat, C. Di
Nolfo, and Y. Leblebici, “Large-scale neural networks implemented with
non-volatile memory as the synaptic weight element: Comparative per-
formance analysis (accuracy, speed, and power),” in IEEE International
Electron Devices Meeting, ser. IEDM. Washington, DC: IEEE, 2015,
pp. 4.4.1–4.4.4, doi:10.1109/IEDM.2015.7409625.
[4] K. Ota, J. Deguchi, S. Fujii, M. Saitoh, M. Yamaguchi, R. Berdan,
T. Marukame, Y. Nishi, K. Matsuo, K. Takahashi, Y. Kamiya,
and S. Miyano, “Performance maximization of in-memory reinforce-
ment learning with variability-controlled hf 1-x zr x o 2 ferro-
electric tunnel junctions,” in IEEE International Electron Devices
Meeting. San Francisco, CA, USA: IEEE, 2019, pp. 6.2.1–6.2.4,
[5] K. Ni, X. Yin, A. F. Laguna, S. Joshi, S. D¨
unkel, M. Trentzsch,
J. M¨
ueller, S. Beyer, M. Niemier, X. S. Hu, and S. Datta, “Ferroelectric
ternary content-addressable memory for one-shot learning,” Nat Elec-
tron, vol. 2, no. 11, pp. 521–529, 2019, doi:10.1038/s41928-019-0321-3.
[6] R. Berdan, T. Marukame, K. Ota, M. Yamaguchi, M. Saitoh, S. Fujii,
J. Deguchi, and Y. Nishi, “Low-power linear computation using non-
linear ferroelectric tunnel junction memristors,” Nat Electron, vol. 1, p.
216, 2020, doi:10.1038/s41928-020-0405-0.
[7] E. A. Cartier, W. Kim, N. Gong, T. Gokmen, M. M. Frank, D. M.
Bishop, Y. Kim, S. Kim, T. Ando, E. Y. Wu, P. Adusumilli, J. Rozen,
P. M. Solomon, W. Haensch, M. J. BrightSky, A. Sebastian, G. W.
Burr, and V. Narayanan, “Reliability challenges with materials for
analog computing,” in IEEE International Reliability Physics Sym-
posium, ser. IRPS. Monterey, CA, USA: IEEE, 2019, pp. 1–10,
[8] R. Yang, “In-memory computing with ferroelectrics,Nat Electron,
vol. 19, p. 41, 2020, doi:10.1038/s41928-020-0411-2.
[9] S. Oh, H. Hwang, and I. K. Yoo, “Ferroelectric materials for neu-
romorphic computing,” APL Mater., vol. 7, no. 9, p. 091109, 2019,
[10] B. Obradovic, T. Rakshit, R. Hatcher, J. Kittl, R. Sengupta, J. G.
Hong, and M. S. Rodder, “A multi-bit neuromorphic weight cell using
ferroelectric fets, suitable for soc integration,” IEEE J. Electron Devices
Soc., vol. 6, pp. 438–448, 2018, doi:10.1109/JEDS.2018.2817628.
[11] H. Mulaosmanovic, J. Ocker, S. Muller, M. Noack, J. Muller, P. Po-
lakowski, T. Mikolajick, and S. Slesazeck, “Novel ferroelectric fet
based synapse for neuromorphic systems,” in Symposium on VLSI
Technology, ser. VLSI. Kyoto, Japan: IEEE, 2017, pp. T176–T177,
[12] J. M¨
uller, P. Polakowski, S. M ¨
uller, and T. Mikolajick, “Ferroelectric
hafnium oxide based materials and devices: Assessment of current status
and future prospects,” ECS J. Solid State Sci. Technol., vol. 4, no. 5, pp.
N30–N35, 2015, doi:10.1149/2.0081505jss.
[13] J. M¨
uller, E. Yurchuk, T. Schlosser, J. Paul, R. Hoffmann, S. Muller,
D. Martin, S. Slesazeck, P. Polakowski, J. Sundqvist, M. Czer-
nohorsky, K. Seidel, P. Kucher, R. Boschke, M. Trentzsch, K. Gebauer,
U. Schroder, and T. Mikolajick, “Ferroelectricity in HfO2 enables
nonvolatile data storage in 28 nm HKMG,” in Symposium on VLSI
Technology, ser. VLSI. Honolulu, HI, USA: IEEE, 2012, pp. 25–26,
[14] S. D¨
unkel, M. Trentzsch, R. Richter, P. Moll, C. Fuchs, O. Gehring,
M. Majer, S. Wittek, B. Muller, T. Melde, H. Mulaosmanovic, S. Sle-
sazeck, S. Muller, J. Ocker, M. Noack, D.-A. Lohr, P. Polakowski,
J. Muller, T. Mikolajick, J. Hontschel, B. Rice, J. Pellerin, and S. Beyer,
“A FeFET based super-low-power ultra-fast embedded NVM technology
for 22nm FDSOI and beyond,” in IEEE International Electron Devices
Meeting, ser. IEDM. San Francisco, CA, USA: IEEE, 2017, pp. 19.7.1–
19.7.4, doi:10.1109/IEDM.2017.8268425.
[15] T. S. B¨
oscke, J. M¨
uller, D. Br¨
auhaus, U. Schr¨
oder, and U. B¨
“Ferroelectricity in hafnium oxide thin films,” Appl. Phys. Lett., vol. 99,
no. 10, p. 102903, 2011, doi:10.1063/1.3634052.
[16] J. M¨
uller, T. S. B¨
oscke, U. Schr¨
oder, S. M¨
uller, D. Br¨
U. B¨
ottger, L. Frey, and T. Mikolajick, “Ferroelectricity in simple binary
ZrO2 and HfO2,” Nano Lett, vol. 12, no. 8, pp. 4318–4323, 2012,
[17] P. Polakowski and J. M¨
uller, “Ferroelectricity in undoped hafnium
oxide,” Appl. Phys. Lett., vol. 106, no. 23, p. 232905, 2015,
[18] T. Ali, P. Polakowski, S. Riedel, T. B¨
uttner, T. K¨
ampfe, M. Rudolph,
B. P¨
atzold, K. Seidel, D. L¨
ohr, R. Hoffmann, M. Czernohorsky,
K. K¨
uhnel, P. Steinke, J. Calvo, K. Zimmermann, and J. M¨
uller, “High
endurance ferroelectric hafnium oxide-based fefet memory without
retention penalty,IEEE Trans. Electron Devices, vol. 65, no. 9, pp.
3769–3774, 2018, doi:10.1109/TED.2018.2856818.
[19] J. M¨
uller, T. S. B¨
oscke, S. M¨
uller, E. Yurchuk, P. Polakowski, J. Paul,
D. Martin, T. Schenk, K. Khullar, A. Kersch, W. Weinreich, S. Riedel,
K. Seidel, A. Kumar, T. M. Arruda, S. V. Kalinin, T. Schlosser,
R. Boschke, R. van Bentum, U. Schr¨
oder, and T. Mikolajick, “Ferroelec-
tric hafnium oxide: A CMOS-compatible and highly scalable approach
to future ferroelectric memories,” in IEEE International Electron De-
vices Meeting, ser. IEDM. Washington, DC, USA: IEEE, 2013, pp.
10.8.1–10.8.4, doi:10.1109/IEDM.2013.6724605.
[20] D. Lehninger, R. Olivo, T. Ali, M. W. Lederer, T. K¨
ampfe, C. Mart,
K. Biedermann, K. K¨
uhnel, L. Roy, M. Kalkani, and K. Seidel, “Back–
end–of–line compatible low temperature furnace anneal for ferroelectric
hafnium zirconium oxide formation,” Phys. Status Solidi A, vol. 217,
no. 8, p. 1900840, 2020, doi:10.1002/pssa.201900840.
[21] S. Beyer, S. Dunkel, M. Trentzsch, J. Muller, A. Hellmich, D. Utess,
J. Paul, D. Kleimaier, J. Pellerin, S. Muller, J. Ocker, A. Benoist,
H. Zhou, M. Mennenga, M. Schuster, F. Tassan, M. Noack, A. Pourkera-
mati, F. Muller, M. W. Lederer, T. Ali, R. Hoffmann, T. Kampfe, K. Sei-
del, H. Mulaosmanovic, E. T. Breyer, T. Mikolajick, and S. Slesazeck,
“Fefet: A versatile cmos compatible device with game-changing poten-
tial,” in IEEE International Memory Workshop, ser. IMW. Dresden,
Germany: IEEE, 2020, pp. 1–4, doi:10.1109/IMW48823.2020.9108150.
[22] S. Yu, P.-Y. Chen, Y. Cao, L. Xia, Y. Wang, and H. Wu, “Scaling-
up resistive synaptic arrays for neuro-inspired architecture: Chal-
lenges and prospect,” in IEEE International Electron Devices Meet-
ing, ser. IEDM. Washington, DC: IEEE, 2015, pp. 17.3.1–17.3.4,
[23] K. Lee, H.-J. Lee, T. Y. Lee, H. H. Lim, M. S. Song, H. K. Yoo,
D. I. Suh, J. G. Lee, Z. Zhu, A. Yoon, M. R. MacDonald, X. Lei,
K. Park, J. Park, J. H. Lee, and S. C. Chae, “Stable subloop behavior
in ferroelectric si-doped hfo2,” ACS Appl. Mater. Interfaces., 2019,
[24] H. Mulaosmanovic, J. Ocker, S. M¨
uller, U. Schroeder, J. M¨
P. Polakowski, S. Flachowsky, R. van Bentum, T. Mikolajick, and
S. Slesazeck, “Switching kinetics in nanoscale hafnium oxide based
ferroelectric field-effect transistors,ACS Appl. Mater. Interfaces., vol. 9,
no. 4, pp. 3792–3798, 2017, doi:10.1021/acsami.6b13866.
[25] H. Mulaosmanovic, E. Chicca, M. Bertele, T. Mikolajick, and S. Sle-
sazeck, “Mimicking biological neurons with a nanoscale ferroelec-
tric transistor,Nanoscale, vol. 10, no. 46, pp. 21 755–21 763, 2018,
[26] F. Zahoor, T. Z. Azni Zulkifli, and F. A. Khanday, “Resistive random
access memory (rram): an overview of materials, switching mechanism,
performance, multilevel cell (mlc) storage, modeling, and applications,
Nanoscale Res Lett, vol. 15, no. 1, p. 90, 2020, doi:10.1186/s11671-
[27] T.-P. Ma and J.-P. Han, “Why is nonvolatile ferroelectric memory field-
effect transistor still elusive?” IEEE Electron Device Lett., vol. 23, no. 7,
pp. 386–388, 2002, doi:10.1109/LED.2002.1015207.
[28] N. Gong and T.-P. Ma, “Why is fe–hfo 2 more suitable than pzt or
sbt for scaled nonvolatile 1-t memory cell? a retention perspective,”
IEEE Electron Device Lett., vol. 37, no. 9, pp. 1123–1126, 2016,
[29] J. M¨
uller, P. Polakowski, S. M ¨
uller, H. Mulaosmanovic, J. Ocker,
T. Mikolajick, S. Slesazeck, S. Flachowsky, and M. Trentzsch,
“High endurance strategies for hafnium oxide based ferroelectric
field effect transistor,” in Non-Volatile Memory Technology Sympo-
sium, ser. NVMTS. Pittsburgh, PA, USA: IEEE, 2016, pp. 1–7,
[30] T. Ali, P. Polakowski, T. B¨
uttner, T. Kampfe, M. Rudolph, B. Pat-
zold, R. Hoffmann, M. Czernohorsky, K. Kuhnel, P. Steinke, L. M.
Eng, and K. Seidel, “Principles and challenges for binary oxide
based ferroelectric memory fefet,” in IEEE International Memory
Workshop, ser. IMW. Monterey, CA, USA: IEEE, 2019, pp. 1–4,
[31] M. W. Lederer, T. K¨
ampfe, N. Vogel, D. Utess, B. Volkmann, T. Ali,
R. Olivo, J. M¨
uller, S. Beyer, M. Trentzsch, K. Seidel, and L. M. Eng,
“Structural and electrical comparison of Si and Zr doped hafnium oxide
thin films and integrated FeFETs utilizing transmission kikuchi diffrac-
tion,” Nanomaterials, vol. 10, no. 2, 2020, doi:10.3390/nano10020384.
[32] H. Mulaosmanovic, S. Slesazeck, J. Ocker, M. Pesic, S. M¨
uller, S. Fla-
chowsky, J. M ¨
uller, P. Polakowski, J. Paul, S. Jansen, S. Kolodinski,
C. Richter, S. Piontek, T. Schenk, A. Kersch, C. K¨
unneth, R. van
Bentum, U. Schr¨
oder, and T. Mikolajick, “Evidence of single do-
main switching in hafnium oxide based fefets: Enabler for multi-
level fefet memory cells,” in IEEE International Electron Devices
Meeting, ser. IEDM. Washington, DC: IEEE, 2015, pp. 26.8.1–26.8.3,
[33] H.-J. Lee, M. Lee, K. Lee, J. Jo, H. Yang, Y. Kim, S. C. Chae,
U. Waghmare, and J. H. Lee, “Scale-free ferroelectricity induced by
flat phonon bands in hfo2,” Science, vol. 11, no. 369, pp. 1343–1347,
2020, doi:10.1126/science.aba0067.
... At the same time, the maximum applied amplitude is not sufficient to switch all domains, thus reducing the accessed effective memory window and therefore the dynamic range. 273 Another influence on the switching process during LTP and LTD is the thickness of the hafnium oxide layer. As shown in Fig. 6.23d, thinner layers result in similar non-linearity coefficients. ...
... Hence, thinner layers might be of interest for low voltage applications. 273 Nevertheless, the scaling of thickness is limited due to the in section 6.1.1 discussed increase in depolarization field. Finally, by tuning the switching properties based on the previous discussions, linear switching behavior with a high number of intermediate states can be achieved even in highly scaled devices, e.g. for 90x90 nm 2 as depicted in Fig. 6.23e and 6.23f. ...
... Retention measurements of the intermediate states in FeFETs clearly predict long term stability with an expected retention of up to 10 years (see Fig. 6.24a). 273 Moreover, it has been demonstrated that the polarization state retention is not affected significantly by operating at increased temperature (up to 120°C). Due to the pyroelectric properties of the material, however, the polarization value decreases with increasing temperature, affecting the threshold voltage of the device, as well as the temperature dependent carrier concentration in the semiconductor. ...
Full-text available
The discovery of ferroelectricity in hafnium oxide spurred a growing research field due to hafnium oxides compatibility with processes in microelectronics as well as its unique properties. Notably, its application in non-volatile memories, neuromorphic devices as well as piezo- and pyroelectric sensors is investigated. However, the behavior of ferroelectric hafnium oxide is not understood into depth compared to common perovskite structure ferroelectrics. Due the the metastable nature of the ferroelectric phase, process conditions have a strong influence during and after its deposition. In this work, the physical properties of hafnium oxide, process influences on the microstructure as well as reliability aspects in non-volatile and neuromorphic devices are investigated. With respect to the physical properties, strong evidence is provided that the antiferroelectric-like behavior in hafnium oxide based thin films is governed by ferroelastic 90° domain wall movement. Furthermore, the discovery of an electric field-induced crystallization process in this material system is reported. For the analysis of the microstructure, the novel method of transmission Kikuchi diffraction is introduced, allowing an investigation of the local crystallographic phase, orientation and grain structure. Here, strong crystallographic textures are observed in dependence of the substrate, doping concentration and annealing temperature. Based on these results, the observed reliability behavior in the electronic devices is explainable and engineering of the present defect landscape enables further optimization. Finally, the behavior in neuromorphic devices is explored as well as process and design guidelines for the desired behavior are provided.
... With the discovery of ferroelectricity in hafnium oxide thin films [1], which has been previously employed as high-k dielectric in high-k-metal-gate (HKMG) complementarymetal-oxide-semiconductor (CMOS) processes, researchers showed strong interest for the integration of this material into novel devices. Applications range from non-volatile memories [2,3] and neuromorphic devices [4,5] to piezoelectric micro-electro-mechanicalsystems (MEMS) [6][7][8], energy harvesting devices [9] and pyroelectric sensors [10,11]. The great interest in this material stems not only from its CMOS compatibility but also red from its unique properties: a high coercive field, combined with high remanent polarization and comparably low relative permittivity [12]. ...
Full-text available
Devices based on ferroelectric hafnium oxide are of major interest for sensor and memory applications. In particular, Si-doped hafnium oxide layers are investigated for the application in the front-end-of-line due to their resilience to high thermal treatments. Due to its very confined doping concentration range, Si:HfO2 layers based on thermal atomic layer deposition often exhibited a crossflow pattern across 300 mm wafer. Here, plasma enhanced atomic layer deposition is explored as an alternative method for producing Si-doped HfO2 layers, and their ferroelectric and pyroelectric properties are compared.
... Lederer et al. introduced the advantages of FeFETs based on hafnium oxide for such applications because they have a basic three-terminal structure that can selectively activate or deactivate selected devices and adjust linearity and dynamic range for certain applications. In addition, the effects of the material properties of the ferroelectric layer, the thickness of the interface layer, and the scaling on the device performance are discussed [12]. Chien et al. recommend embedding 256 kb resistive random access (ReRAM) in the microcontroller unit as a data buffer for communication with independent flash memory. ...
Full-text available
The research direction of the new generation of embedded memory can be summarized into two types of embedded nonvolatile memory and embedded volatile memory; the research on online testing of embedded memory started in the past ten years, and there are few research results. This article analyzes the feasibility of the noncontact IC chip in the embedded ferroelectric memory of the sports game auxiliary timing device and is aimed at obtaining an optimized embedded ferroelectric memory by analyzing the relevant data to achieve the update and update of the sports game timing device system. Early sports event timing methods generally use manual timing (stopwatch) or camera shooting timing; this method is inefficient, poor real-time, huge workload, and prone to errors. This research mainly focuses on the analysis and discussion of the material structure and performance of the embedded ferroelectric memory and the process of noncontact IC chip. This article uses custom welding circuit technology to prepare the best ferroelectric filter in the test part and verifies the influence of temperature on the material; in order to understand the properties of ferroelectric materials at the electronic and atomic level, a first-order statistical method is obtained. The numerical calculation results of the experiment verify that the evaluation value of the serial port synchronization module as a whole exceeds the pulse synchronization; the network synchronization as a whole exceeds the code synchronization, and the result of the network time service module is the opposite, but as a whole, each module of the noncontact IC chip has strong performance adaptability; in the application of auxiliary timing, the maintainability of noncontact IC chip is quite outstanding, and the maximum value is 7.97; a large number of complex simulation system tasks can be completed by simple and direct tasks.
... Further, ferroelectric field effect transistors (FeFETs) stand out for new era of FET devices in view of low power loss, non-volatile read ability as a memory device, and quick operation speed [19][20][21]. The utilization of ferroelectric material in the gatestack of MOSFET gives better resistance to the device execution since FE material shows the conduct of negative capacitance which can be seen from their charge energy bend [22][23][24]. ...
Full-text available
In this article, we investigated the analog performance and RF(Radio Frequency) performance of ferroelectric layer improved Field Effect Transistor device that is metal ferroelectric metal insulator metal oxide transistor (MFMIMOS) with spacer and without spacer. A proposed device MFMIMOS (spacer) with spacer enhances the ON-current (Ion) by 25% as compared to a without spacer device MFMIMOS, leakage current (Ioff) reduces by almost 37%, switching ratio (Ion/Ioff) enhanced by 99%, threshold voltage (Vth) increased by 0.29%, subthreshold swing (SS) reduced by almost 5.6% and drain induced barrier lowering (DIBL) lowed by 17.6% over MFMIMOS. We also examined the analog parameters for improving the performance of the proposed device MFMIMOS with a spacer at temperature (T) 300K such as a transconductance generation factor (TGF), transconductance (gm), intrinsic gain (Av), early voltage (Vea), intrinsic delay (Ti), and some RF parameters gain transconductance frequency product (GTFP), cut off frequency (Ft), and gain frequency product(GFP). In the variation of temperature on MFMIMOS(spacer) we observed improved assessment at room temperature rather than other temperatures T = 400K and T = 500K such as threshold voltage higher by 2.4 times over temperature T = 500K, switching ratio (Ion/Ioff), transconductance (gm), transconductance generation factor (TGF). All the simulated result is taken by Visual TCAD simulator with high compatibility of MFMIMOS (spacer) with spacer. Thus, the proposed device MFMIMOS (spacer) with spacer shows significantly improved analog performance than its conventional counterpart and improved performance at room temperature (T = 300K) rather than other temperatures T = 400K and T = 500K.
... Moreover, the potentiation and depression states should be symmetrical to obtain symmetric switching behavior (symmetric potentiation and depression responses) for ideal cases [28]. Desirable DR values of ideal neuromorphic devices are expected to be as large as possible [29]. The DR is correlated with the neural network accuracy and DR should have a minimum value of 10 to achieve network accuracies higher than 80% [30]. ...
Nonvolatile memories especially the ferroelectric (FE)-based ones such as ferroelectric tunnel junctions (FTJs) and ferroelectric field-effect transistors (FeFETs) have recently attracted a lot of attention. FTJs have been intensively researched for the last decade and found to be very promising memory devices due to their significant nondestructive readout advantage as compared to conventional ferroelectric random access memory (FRAM). However, more research is needed on FTJ devices to obtain reliable endurance and retention behavior. In this article, we demonstrate the characteristics and performance of zirconium-doped hafnium oxide-based FTJ devices in terms of FE switching and reliability. This is investigated for FTJ stack structure tuning as well as for the FE switching process in FTJ devices. The FTJ memory switching characteristics, the effects of polarization switching on the write conditions, and the impact of pulse width and pulse amplitude on switching are investigated. The impact of FE layer thickness and interface layer type/thickness are reported to obtain a maximum FTJ $I_{{on}}$ / $I_{{off}}$ ratio (memory window) and reliable performance. The maximum $I_{{on}}$ / $I_{{off}}$ ratio changes depending on the FE layer (zirconium-doped HfO₂ layer) thickness (12, 8, 6, and 4 nm), the interface layer type (SiO₂, Al₂O₃), and thickness (1 and 2 nm), indicating the maximum value of $I_{{on}}$ / $I_{{off}}$ ratio for a 1 nm SiO₂ interface layer stack. Moreover, a stable endurance of 10⁴ cycles is reported and extrapolated measurements suggest stable retention for more than ten years. Time-dependent breakdown analysis was performed to investigate the reliability of devices indicating a lifetime of ten years.
Full-text available
p>Interfacial Layer Engineering to Enhance Endurance and Noise Immunity of FeFETs for IMC Applications</p
Due to the slow-down of Moore’s Law and Dennard Scaling, new disruptive computer architectures are mandatory. One such new approach is Neuromorphic Computing, which is inspired by the functionality of the human brain. In this position paper, we present the projected SEC-Learn ecosystem, which combines neuromorphic embedded architectures with Federated Learning in the cloud, and performance with data protection and energy efficiency.
Today, a large number of applications depend on deep neural networks (DNN) to process data and perform complicated tasks at restricted power and latency specifications. Therefore, processing-in-memory (PIM) platforms are actively explored as a promising approach to improve the throughput and the energy efficiency of DNN computing systems. Several PIM architectures adopt resistive non-volatile memories as their main unit to build crossbar-based accelerators for DNN inference. However, these structures suffer from several drawbacks such as reliability, low accuracy, large ADCs/DACs power consumption and area, high write energy, etc. In this paper, we present a new mixed-signal in-memory architecture based on the bit-decomposition of the multiply and accumulate operations. Our in-memory inference architecture uses a single FeFET as a non-volatile memory cell. Compared to the prior work, this system architecture provides a high level of parallelism while using only 3-bit ADCs. Also, it eliminates the need for any DAC. In addition, we provide flexibility and a very high utilization efficiency even for varying tasks and loads. Simulations demonstrate that we outperform state-of-the-art efficiencies with 36.5 TOPS/W and can pack 2.05 TOPS with 8-bit activation and 4-bit weight precision in an area of 4.9 mm ² using 22 nm FDSOI technology. Employing binary operation, we obtain 1169 TOPS/W and over 261 TOPS/W/mm ² on system level.
Full-text available
This work presents 2 bits/cell operation in deeply scaled ferroelectric finFETs (Fe-finFET) with 1µs write pulse of maximum ±5V amplitude and WRITE endurance above 109 cycles. Fe-finFET devices with single and multiple fins have been fabricated on SOI wafer using a gate first process, with gate lengths down to 70 nm and fin width 20 nm. Extrapolated retention above ten years also ensures stable inference operation for ten years without any need for re-training. Statistical modeling of device-to-device and cycle-to-cycle variation is performed based on measured data and applied to neural network simulations using the CIMulator software platform. Stochastic device-to-device variation is mainly compensated during online training and has virtually no impact on training accuracy. On the other hand, stochastic cycle-to-cycle threshold voltage variation up to 400mV can be tolerated for MNIST handwritten digits recognition. Substantial inference accuracy drop with systematic retention degradation was observed in analog neural networks. However, quaternary neural networks (QNN) and binary neural networks (BNN) with Fe-finFETs as synaptic devices demonstrated excellent immunity towards the cumulative impact of stochastic and systematic variations
With the exponential increase in the quantity of information to be stored and processed, an important issue that must be urgently resolved for the advancement of modern society is to decrease the power consumed by semiconductor devices with high operation speeds. Logic-in-memory (LiM) and neuromorphic devices were proposed as promising solutions to improve the operation speed and energy efficiency by merging logic and memory devices that are classically separated in von Neumann computing systems. Numerous emerging memories were proposed for the LiM and neuromorphic devices of which ferroelectric memories were considered to be one of the most promising candidates since the discovery of unexpected ferroelectricity in complementary metal–oxide–semiconductor compatible binary oxides such as doped HfO2. Therefore, a review of binary ferroelectric oxides, from materials to devices, for logic-memory hybrid systems is presented herein.Graphic abstract
Full-text available
Analogue in-memory computing using memristors could alleviate the performance constraints imposed by digital von Neumann systems in data-intensive tasks. Conventional linear memristors typically operate at high currents, potentially limiting power efficiency and scalability in practical applications. Here, we show that nonlinear ferroelectric tunnel junction memristors can perform linear computation at ultralow currents. Using logarithmic line drivers, we demonstrate that analogue-voltage-amplitude vector–matrix multiplication (VMM) can be performed in selectorless ferroelectric tunnel junction crossbars by exploiting a device nonlinearity factor that remains constant for multiple conductive states. We also show that our ferroelectric tunnel junction crossbars have the attributes required to scale analogue VMM-intensive applications, such as neural inference engines, towards energy efficiencies above 100 tera-operations per second per watt. Nonlinear ferroelectric tunnel junction memristors can be used to perform linear vector–matrix multiplication operations at ultralow currents.
Full-text available
In this manuscript, recent progress in the area of resistive random access memory (RRAM) technology which is considered one of the most standout emerging memory technologies owing to its high speed, low cost, enhanced storage density, potential applications in various fields, and excellent scalability is comprehensively reviewed. First, a brief overview of the field of emerging memory technologies is provided. The material properties, resistance switching mechanism, and electrical characteristics of RRAM are discussed. Also, various issues such as endurance, retention, uniformity, and the effect of operating temperature and random telegraph noise (RTN) are elaborated. A discussion on multilevel cell (MLC) storage capability of RRAM, which is attractive for achieving increased storage density and low cost is presented. Different operation schemes to achieve reliable MLC operation along with their physical mechanisms have been provided. In addition, an elaborate description of switching methodologies and current voltage relationships for various popular RRAM models is covered in this work. The prospective applications of RRAM to various fields such as security, neuromorphic computing, and non-volatile logic systems are addressed briefly. The present review article concludes with the discussion on the challenges and future prospects of the RRAM.
Full-text available
The microstructure of ferroelectric hafnium oxide plays a vital role for its application, e.g., non-volatile memories. In this study, transmission Kikuchi diffraction and scanning transmission electron microscopy STEM techniques are used to compare the crystallographic phase and orientation of Si and Zr doped HfO2 thin films as well as integrated in a 22 nm fully-depleted silicon-on-insulator (FDSOI) ferroelectric field effect transistor (FeFET). Both HfO2 films showed a predominately orthorhombic phase in accordance with electrical measurements and X-ray diffraction XRD data. Furthermore, a stronger texture is found for the microstructure of the Si doped HfO2 (HSO) thin film, which is attributed to stress conditions inside the film stack during crystallization. For the HSO thin film fabricated in a metal-oxide-semiconductor (MOS) like structure, a different microstructure, with no apparent texture as well as a different fraction of orthorhombic phase is observed. The 22 nm FDSOI FeFET showed an orthorhombic phase for the HSO layer, as well as an out-of-plane texture of the [111]-axis, which is preferable for the application as non-volatile memory.
Full-text available
The discovery of ferroelectricity in thin doped hafnium oxide films revived the interest in ferroelectric memory concepts. Zirconium doped hafnium oxide (HZO) crystalizes at low rapid thermal annealing (RTA) temperatures (e.g. 400°C), which makes this material interesting for the implementation of ferroelectric functionalities into the back‐end‐of‐line (BEoL) of modern integrated circuits. So far, the ferroelectric phase of prior amorphous HZO films is achieved by a dedicated RTA treatment. However, this article shows that such a dedicated anneal step is not needed. A sole furnace treatment within the thermal budget present during the interconnect‐formation is sufficient to functionalize even ultrathin 5 nm HZO films. This result will help to optimize the integration sequence of HZO films (e.g. involving a minimum number of BEoL process steps), which saves process time and fabrication costs. For this study, metal‐ferroelectric‐metal capacitors with HZO films of different thicknesses were annealed at 400 °C for various durations within different types of ovens (RTP, furnace). Structural and electrical characterization confirmed that all furnace‐annealed samples had similar X‐ray diffraction patterns, remanent polarization, endurances, and thickness‐dependencies as RTP‐annealed ones. The HZO film of 10 nm is most promising for the integration into the BEoL. This article is protected by copyright. All rights reserved.
Full-text available
Deep neural networks are efficient at learning from large sets of labelled data, but struggle to adapt to previously unseen data. In pursuit of generalized artificial intelligence, one approach is to augment neural networks with an attentional memory so that they can draw on already learnt knowledge patterns and adapt to new but similar tasks. In current implementations of such memory augmented neural networks (MANNs), the content of a network’s memory is typically transferred from the memory to the compute unit (a central processing unit or graphics processing unit) to calculate similarity or distance norms. The processing unit hardware incurs substantial energy and latency penalties associated with transferring the data from the memory and updating the data at random memory addresses. Here, we show that ternary content-addressable memories (TCAMs) can be used as attentional memories, in which the distance between a query vector and each stored entry is computed within the memory itself, thus avoiding data transfer. Our compact and energy-efficient TCAM cell is based on two ferroelectric field-effect transistors. We evaluate the performance of our ferroelectric TCAM array prototype for one- and few-shot learning applications. When compared with a MANN where cosine distance calculations are performed on a graphics processing unit, the ferroelectric TCAM approach provides a 60-fold reduction in energy and 2,700-fold reduction in latency for a single memory search operation. A compact ternary content-addressable memory cell, which is based on two ferroelectric field-effect transistors, can provide memory augmented neural networks with improved energy and latency performance compared with traditional approaches based on graphics processing units.
Discovery of robust yet reversibly switchable electric dipoles at reduced dimensions is critical in advancing nanoelectronics devices. Energy bands flat in momentum space generate robust localized states that are activated independently of each other. We determined flat bands exist and induce robust yet independently switchable dipoles exhibiting a unique ferroelectricity in HfO 2 . Flat polar phonon bands in HfO 2 cause extreme localization of electric dipoles within its irreducible half-unit-cell-widths (~3 Å). Contrary to conventional ferroelectrics with spread dipoles, those intrinsically localized dipoles are stable against extrinsic effects such as domain walls, surface exposure, and even down-to-angstrom-scale miniaturization. Moreover, the sub-nm-scale dipoles are individually switchable without creating any domain-wall energy cost. This offers unexpected opportunities for ultimately-dense unit-cell-by-unit-cell ferroelectric switching devices directly integrable into silicon technology.
Non-linear ferroelectric tunnel junctions could be used to create low-power in-memory computing systems.
The recent demand for analog devices for neuromorphic applications requires modulation of multiple nonvolatile states. Ferroelectricity with multiple polarization states enables neuromorphic applications with various architectures. However, deterministic control of ferroelectric polarization states with conventional ferroelectric materials has been met with accessibility issues. Here we report unprecedented stable accessibility with robust stability of multiple polarization states in ferroelectric HfO2. Through the combination of conventional voltage measurements, hysteresis temperature dependence analysis, piezoelectric force microscopy, first-principles calculations, and Monte Carlo simulations, we suggest that the unprecedented stability of intermediate states in ferroelectric HfO2 is due to the small critical volume size for nucleation and the large activation energy for ferroelectric dipole flipping. This work demonstrates the potential of ferroelectric HfO2 for analog device applications enabling neuromorphic computing.