Achieving Low-Power Analog Resistive Switching in Filamentary Memristive Devices for Energy-Efficient Analog...

Area of Science:

• Materials Science
• Computer Engineering
• Artificial Intelligence

Background:

• Traditional Von Neumann architectures face energy inefficiencies, particularly in deep learning.
• Filamentary memristive devices struggle with high operating voltages and power consumption.
• Analog in-memory computing offers a promising alternative for energy efficiency.

Purpose of the Study:

• To engineer a novel dual-matrix memristive device to overcome limitations of existing filamentary memristive devices.
• To reduce operational voltage and analog switching current for lower power consumption.
• To demonstrate the effectiveness of the proposed device in deep learning applications.

Main Methods:

• Utilized a dual-matrix filamentary switching approach with GeSe2 (high-mobility) and densified amorphous silicon (low-mobility) matrices.
• Incorporated Ag and Pt nano-cluster layers for enhanced device performance.
• Performed simulations on the Spiking-VGG9 architecture with quantization-aware training and the tiki-taka method.

Main Results:

• Achieved over 50% reduction in forming and set voltages.
• Reduced reset current by more than 96% and analog energy consumption by 93% (0.98 pJ).
• Demonstrated stable retention (≈24 h) and endurance (≈50k cycles).
• Simulations showed 89.38% accuracy on CIFAR10 and 63.70% on CIFAR10-DVS with a 60.42% reduction in total energy consumption.

Conclusions:

• The proposed dual-matrix memristive device significantly enhances energy efficiency for analog in-memory computing.
• This approach effectively addresses power consumption challenges in deep learning hardware.
• The technology shows great potential for developing low-power AI accelerators.