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
Synaptic weights.(a) The initial current level of the 16 × 16 array of synaptic devices is shown. The conductance is initialized to 7.5 nA at 0.5 V read. (b) The word “hello” when trained on to the 16 × 16 crossbar array of synaptic devices results in a conductance distribution as shown in the plot. The same word was repeated 10 times to ensure the distribution is similar. (c) When the word “apple” was input into the network, the distribution of conductance is markedly different. The weight evolution is obtained from the theoretical STDP fitting of figure 8(b). The colormap shows the current read in nA at 0.5 V after the training procedure. The conductance of each synaptic device along a row is the same since they were initialized to the same conductance level and faced the same STDP training.
© Copyright Policy - open-access
Related In: Results  -  Collection

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

f10: Synaptic weights.(a) The initial current level of the 16 × 16 array of synaptic devices is shown. The conductance is initialized to 7.5 nA at 0.5 V read. (b) The word “hello” when trained on to the 16 × 16 crossbar array of synaptic devices results in a conductance distribution as shown in the plot. The same word was repeated 10 times to ensure the distribution is similar. (c) When the word “apple” was input into the network, the distribution of conductance is markedly different. The weight evolution is obtained from the theoretical STDP fitting of figure 8(b). The colormap shows the current read in nA at 0.5 V after the training procedure. The conductance of each synaptic device along a row is the same since they were initialized to the same conductance level and faced the same STDP training.

Mentions: The initial synaptic weights before training are shown in Figure 10(a). The current levels at 0.5 V read are shown on the adjacent colour map. A unified colour refers to a constant conductance level for all the devices in the array. Figures 10(b) and (c) show the weight distribution of the synaptic array at the end of the simulation for words “apple” and “hello”, respectively. A clear distinction can be observed in the pattern of conductance levels for these two words. It is interesting to note that the current level along each row of the crossbar is equal as inferred from the colour map. Such a pattern was expected as each of the synaptic elements along a particular row undergoes the same STDP learning since the conductance change and conductance initialization for the synaptic devices were fixed. The impacts of device to device variability and statistical variation in STDP have been examined in a previous work which can be incorporated leading to a more diversified map26.


Novel synaptic memory device for neuromorphic computing.

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

Synaptic weights.(a) The initial current level of the 16 × 16 array of synaptic devices is shown. The conductance is initialized to 7.5 nA at 0.5 V read. (b) The word “hello” when trained on to the 16 × 16 crossbar array of synaptic devices results in a conductance distribution as shown in the plot. The same word was repeated 10 times to ensure the distribution is similar. (c) When the word “apple” was input into the network, the distribution of conductance is markedly different. The weight evolution is obtained from the theoretical STDP fitting of figure 8(b). The colormap shows the current read in nA at 0.5 V after the training procedure. The conductance of each synaptic device along a row is the same since they were initialized to the same conductance level and faced the same STDP training.
© Copyright Policy - open-access
Related In: Results  -  Collection

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

f10: Synaptic weights.(a) The initial current level of the 16 × 16 array of synaptic devices is shown. The conductance is initialized to 7.5 nA at 0.5 V read. (b) The word “hello” when trained on to the 16 × 16 crossbar array of synaptic devices results in a conductance distribution as shown in the plot. The same word was repeated 10 times to ensure the distribution is similar. (c) When the word “apple” was input into the network, the distribution of conductance is markedly different. The weight evolution is obtained from the theoretical STDP fitting of figure 8(b). The colormap shows the current read in nA at 0.5 V after the training procedure. The conductance of each synaptic device along a row is the same since they were initialized to the same conductance level and faced the same STDP training.
Mentions: The initial synaptic weights before training are shown in Figure 10(a). The current levels at 0.5 V read are shown on the adjacent colour map. A unified colour refers to a constant conductance level for all the devices in the array. Figures 10(b) and (c) show the weight distribution of the synaptic array at the end of the simulation for words “apple” and “hello”, respectively. A clear distinction can be observed in the pattern of conductance levels for these two words. It is interesting to note that the current level along each row of the crossbar is equal as inferred from the colour map. Such a pattern was expected as each of the synaptic elements along a particular row undergoes the same STDP learning since the conductance change and conductance initialization for the synaptic devices were fixed. The impacts of device to device variability and statistical variation in STDP have been examined in a previous work which can be incorporated leading to a more diversified map26.

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