Limits...
Novel synaptic memory device for neuromorphic computing.

Mandal S, El-Amin A, Alexander K, Rajendran B, Jha R - Sci Rep (2014)

Bottom Line: The devices are based on Mn doped HfO₂ material.The model was then utilized to show the application of these devices in speech recognition.A comparison between a 20 nm × 20 nm sized synaptic memory device with that of a state-of-the-art VLSI SRAM synapse showed ~10× reduction in area and >10(6) times reduction in the power consumption per learning cycle.

View Article: PubMed Central - PubMed

Affiliation: Department of Electrical Engineering and Computer Science, University of Toledo, University of Toledo, OH 43606, USA.

ABSTRACT
This report discusses the electrical characteristics of two-terminal synaptic memory devices capable of demonstrating an analog change in conductance in response to the varying amplitude and pulse-width of the applied signal. The devices are based on Mn doped HfO₂ material. The mechanism behind reconfiguration was studied and a unified model is presented to explain the underlying device physics. The model was then utilized to show the application of these devices in speech recognition. A comparison between a 20 nm × 20 nm sized synaptic memory device with that of a state-of-the-art VLSI SRAM synapse showed ~10× reduction in area and >10(6) times reduction in the power consumption per learning cycle.

Show MeSH

Related in: MedlinePlus

(a) CV measurement is shown for the sample. Assuming a dielectric constant ~24, the thickness was estimated to be 9.93 nm. The dielectric constant was verified by the transport study. Ln(I/V) vs sqrt(V) plots are shown here for (b) positive and (c) negative bias. Straight line fittings are obtained beyond 0.15 V which indicate that the dominant current conduction mechanism is F-P. At low bias and/or high temperatures (>330 K) different conduction mechanisms set in. Ln(I/V) vs 1/kT plot is shown for different bias points for (d) positive bias and (e) negative bias. The activation energy for carriers in F-P emission is extracted from the slopes of these plots. The intercepts provide the average carrier concentration in the device, assuming a mobility of 0.15 cm2/V-s. Ln(I/T2) vs 1/kT plots for Schottky emission are also shown in the insets of (d) for positive bias and (e) for negative bias. The fits obtained for F-P were better than that of Schottky which confirmed the dominant conduction mechanism in the dielectric to be F-P. (f) The activation energy is plotted as a function of sqrt(V). The slope provides the dielectric constant of the material. For a thickness of 9.93 nm the dielectric constant was estimated to be 24. The average trap depth was 0.207 eV for positive bias and 0.232 eV for negative bias.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4061545&req=5

f2: (a) CV measurement is shown for the sample. Assuming a dielectric constant ~24, the thickness was estimated to be 9.93 nm. The dielectric constant was verified by the transport study. Ln(I/V) vs sqrt(V) plots are shown here for (b) positive and (c) negative bias. Straight line fittings are obtained beyond 0.15 V which indicate that the dominant current conduction mechanism is F-P. At low bias and/or high temperatures (>330 K) different conduction mechanisms set in. Ln(I/V) vs 1/kT plot is shown for different bias points for (d) positive bias and (e) negative bias. The activation energy for carriers in F-P emission is extracted from the slopes of these plots. The intercepts provide the average carrier concentration in the device, assuming a mobility of 0.15 cm2/V-s. Ln(I/T2) vs 1/kT plots for Schottky emission are also shown in the insets of (d) for positive bias and (e) for negative bias. The fits obtained for F-P were better than that of Schottky which confirmed the dominant conduction mechanism in the dielectric to be F-P. (f) The activation energy is plotted as a function of sqrt(V). The slope provides the dielectric constant of the material. For a thickness of 9.93 nm the dielectric constant was estimated to be 24. The average trap depth was 0.207 eV for positive bias and 0.232 eV for negative bias.

Mentions: Next, capacitance voltage (CV) characteristic of the device was obtained at several frequencies as shown in Fig. 2(a). From here, assuming a dielectric constant of 24, the film thickness was estimated to be 9.93 nm. To understand the mechanism of charge transport, I–V sweeps were performed at temperatures ranging from 260 K to 350 K. The conduction mechanism was found to be Frenkel-Poole (F-P) emission15 based on excellent r-square (R2) values obtained for the F-P fitting, shown in Figs. 2(a)–2(d). The equation for F-P can be given as: Here, μ is the mobility of dielectric, E is the electric field, A is the area of the device, n0 is defect concentration and ΦB is the depth of the trap from the conduction band of HfO2 which is corrected for the electric field in the exponential. Figures 2(b) and 2(c) show Ln(I/V) vs. sqrt(V) for different temperatures for positive bias and negative bias respectively. Beyond 0.2 V, a straight line fitting with R2 values between 0.998 and 0.999 is obtained for the plots at all temperatures indicating the conduction mechanism to be dominated by F-P. At low bias (<0.2 V) some other mechanism can be dominant, such as trap-assisted tunnelling17 at low temperatures and thermionic emission at higher temperatures (>330 K). The parameters for emission were determined by extracting the slope (Ea) of Ln (I/V) vs. 1/kT plot for different bias points as shown in figure 2(d) for positive bias and 2(e) for negative bias. For comparison, Schottky emission fittings for Ln(I/T2) vs 1/kT were also tried as shown in the inset of figures 2(d) and 2(e). However, R2 values for F-P (0.983–0.999 for negative biases and 0.998–0.999 for positive biases) were found to be better than Schottky fittings indicating the dominant mechanism of conduction to be F-P in these samples. The Ea was then plotted as a function of the square root of V for positive and negative bias as in Figure 2(f). The extracted ΦB for positive and negative bias were found to be 0.207 and 0.232 eV respectively. Assuming μ ~ 0.15 cm2/V-s16, n0 was estimated to be around 3 × 1011 cm−3. Using Mn:HfO2 thickness of 9.93 nm, extracted from CV, the dielectric constant of ~24 was extracted for both positive and negative bias using F-P fitting. F-P emission is usually associated with symmetric I-Vs29 due to bulk defects. However, asymmetric I-Vs in our devices could be a result of different ΦB observed for the positive and negative biases. It is possible that TiN, being an oxygen gettering layer, can getter oxygen from HfO2 near TiN/HfO2 interface which can lead to different oxidation states of Mn. As a result, defects of different depths in the band-gap of HfO2 can exist which can preferentially participate under positive and negative bias. It has also been reported that the dielectric constant extracted for F-P emission in HfO2 corresponds to optical frequencies30. However, the optical dielectric constant is valid for very high fields, while the E-field for our samples was much lower.


Novel synaptic memory device for neuromorphic computing.

Mandal S, El-Amin A, Alexander K, Rajendran B, Jha R - Sci Rep (2014)

(a) CV measurement is shown for the sample. Assuming a dielectric constant ~24, the thickness was estimated to be 9.93 nm. The dielectric constant was verified by the transport study. Ln(I/V) vs sqrt(V) plots are shown here for (b) positive and (c) negative bias. Straight line fittings are obtained beyond 0.15 V which indicate that the dominant current conduction mechanism is F-P. At low bias and/or high temperatures (>330 K) different conduction mechanisms set in. Ln(I/V) vs 1/kT plot is shown for different bias points for (d) positive bias and (e) negative bias. The activation energy for carriers in F-P emission is extracted from the slopes of these plots. The intercepts provide the average carrier concentration in the device, assuming a mobility of 0.15 cm2/V-s. Ln(I/T2) vs 1/kT plots for Schottky emission are also shown in the insets of (d) for positive bias and (e) for negative bias. The fits obtained for F-P were better than that of Schottky which confirmed the dominant conduction mechanism in the dielectric to be F-P. (f) The activation energy is plotted as a function of sqrt(V). The slope provides the dielectric constant of the material. For a thickness of 9.93 nm the dielectric constant was estimated to be 24. The average trap depth was 0.207 eV for positive bias and 0.232 eV for negative bias.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4061545&req=5

f2: (a) CV measurement is shown for the sample. Assuming a dielectric constant ~24, the thickness was estimated to be 9.93 nm. The dielectric constant was verified by the transport study. Ln(I/V) vs sqrt(V) plots are shown here for (b) positive and (c) negative bias. Straight line fittings are obtained beyond 0.15 V which indicate that the dominant current conduction mechanism is F-P. At low bias and/or high temperatures (>330 K) different conduction mechanisms set in. Ln(I/V) vs 1/kT plot is shown for different bias points for (d) positive bias and (e) negative bias. The activation energy for carriers in F-P emission is extracted from the slopes of these plots. The intercepts provide the average carrier concentration in the device, assuming a mobility of 0.15 cm2/V-s. Ln(I/T2) vs 1/kT plots for Schottky emission are also shown in the insets of (d) for positive bias and (e) for negative bias. The fits obtained for F-P were better than that of Schottky which confirmed the dominant conduction mechanism in the dielectric to be F-P. (f) The activation energy is plotted as a function of sqrt(V). The slope provides the dielectric constant of the material. For a thickness of 9.93 nm the dielectric constant was estimated to be 24. The average trap depth was 0.207 eV for positive bias and 0.232 eV for negative bias.
Mentions: Next, capacitance voltage (CV) characteristic of the device was obtained at several frequencies as shown in Fig. 2(a). From here, assuming a dielectric constant of 24, the film thickness was estimated to be 9.93 nm. To understand the mechanism of charge transport, I–V sweeps were performed at temperatures ranging from 260 K to 350 K. The conduction mechanism was found to be Frenkel-Poole (F-P) emission15 based on excellent r-square (R2) values obtained for the F-P fitting, shown in Figs. 2(a)–2(d). The equation for F-P can be given as: Here, μ is the mobility of dielectric, E is the electric field, A is the area of the device, n0 is defect concentration and ΦB is the depth of the trap from the conduction band of HfO2 which is corrected for the electric field in the exponential. Figures 2(b) and 2(c) show Ln(I/V) vs. sqrt(V) for different temperatures for positive bias and negative bias respectively. Beyond 0.2 V, a straight line fitting with R2 values between 0.998 and 0.999 is obtained for the plots at all temperatures indicating the conduction mechanism to be dominated by F-P. At low bias (<0.2 V) some other mechanism can be dominant, such as trap-assisted tunnelling17 at low temperatures and thermionic emission at higher temperatures (>330 K). The parameters for emission were determined by extracting the slope (Ea) of Ln (I/V) vs. 1/kT plot for different bias points as shown in figure 2(d) for positive bias and 2(e) for negative bias. For comparison, Schottky emission fittings for Ln(I/T2) vs 1/kT were also tried as shown in the inset of figures 2(d) and 2(e). However, R2 values for F-P (0.983–0.999 for negative biases and 0.998–0.999 for positive biases) were found to be better than Schottky fittings indicating the dominant mechanism of conduction to be F-P in these samples. The Ea was then plotted as a function of the square root of V for positive and negative bias as in Figure 2(f). The extracted ΦB for positive and negative bias were found to be 0.207 and 0.232 eV respectively. Assuming μ ~ 0.15 cm2/V-s16, n0 was estimated to be around 3 × 1011 cm−3. Using Mn:HfO2 thickness of 9.93 nm, extracted from CV, the dielectric constant of ~24 was extracted for both positive and negative bias using F-P fitting. F-P emission is usually associated with symmetric I-Vs29 due to bulk defects. However, asymmetric I-Vs in our devices could be a result of different ΦB observed for the positive and negative biases. It is possible that TiN, being an oxygen gettering layer, can getter oxygen from HfO2 near TiN/HfO2 interface which can lead to different oxidation states of Mn. As a result, defects of different depths in the band-gap of HfO2 can exist which can preferentially participate under positive and negative bias. It has also been reported that the dielectric constant extracted for F-P emission in HfO2 corresponds to optical frequencies30. However, the optical dielectric constant is valid for very high fields, while the E-field for our samples was much lower.

Bottom Line: The devices are based on Mn doped HfO₂ material.The model was then utilized to show the application of these devices in speech recognition.A comparison between a 20 nm × 20 nm sized synaptic memory device with that of a state-of-the-art VLSI SRAM synapse showed ~10× reduction in area and >10(6) times reduction in the power consumption per learning cycle.

View Article: PubMed Central - PubMed

Affiliation: Department of Electrical Engineering and Computer Science, University of Toledo, University of Toledo, OH 43606, USA.

ABSTRACT
This report discusses the electrical characteristics of two-terminal synaptic memory devices capable of demonstrating an analog change in conductance in response to the varying amplitude and pulse-width of the applied signal. The devices are based on Mn doped HfO₂ material. The mechanism behind reconfiguration was studied and a unified model is presented to explain the underlying device physics. The model was then utilized to show the application of these devices in speech recognition. A comparison between a 20 nm × 20 nm sized synaptic memory device with that of a state-of-the-art VLSI SRAM synapse showed ~10× reduction in area and >10(6) times reduction in the power consumption per learning cycle.

Show MeSH
Related in: MedlinePlus