Real-time signal processing enabled by fused networks on a memristor-based system on a chip

Area of Science:

• Computer Science
• Electrical Engineering
• Materials Science

Background:

• The von Neumann bottleneck increases energy consumption in computing, especially for data-intensive signal processing.
• Memristor-based in-memory computing offers an energy-efficient solution by integrating computation and memory.
• Analog computing with memristors presents a promising avenue for overcoming traditional digital system limitations.

Purpose of the Study:

• To demonstrate real-time signal processing using a novel memristor-based analog system on a chip (SoC).
• To integrate real-time discrete Fourier transform (DFT) and convolutional neural network (CNN) functionalities on a single chip.
• To evaluate the energy efficiency and performance of the proposed SoC for signal processing tasks.

Main Methods:

• A 128x128 memristor crossbar array was utilized for analog computation.
• The system implemented a fused network combining real-time DFT and CNN for signal analysis.
• The system was tested on audio classification using the AudioMNIST dataset and video edge detection.

Main Results:

• The memristor-based DFT achieved a peak signal-to-noise ratio of 33.49 dB for audio signals.
• The integrated CNN demonstrated 94.72% accuracy in classifying audio spectrograms.
• Convolution-based edge detection was successfully applied to real-time video frames.

Conclusions:

• Memristor-based SoCs enable efficient real-time signal processing by performing computations in analog memory.
• The demonstrated fused DFT-CNN network offers significant energy savings compared to traditional digital systems.
• This technology holds substantial potential for future low-power, high-performance signal processing applications.