A Low-Power Analog Processor-in-Memory-Based Convolutional Neural Network for Biosensor Applications.

Area of Science:

• Electronics
• Biomedical Engineering
• Computer Science

Background:

• Convolutional Neural Networks (CNNs) are crucial for biosignal processing.
• Integrating CNNs directly onto biosensors offers potential for low-power, real-time analysis.
• Existing solutions often face power consumption and efficiency challenges.

Purpose of the Study:

• To present an on-chip, low-power analog processor-in-memory (PIM) implementation of a CNN for biosensor applications.
• To demonstrate the feasibility of integrating complex neural network operations directly within a biosensor.
• To achieve high energy efficiency for biosignal feature extraction and classification.

Main Methods:

• An analog 10T SRAM-based PIM architecture was employed for multiplication and accumulation (MAC) operations.
• A 32x32 biosensor array was utilized, with data processed via an input matrix formed by digital controller scanning.
• Memory reuse techniques were applied to the analog SRAM filter for low-power implementation.
• CNN models were trained using biosignal data to confirm classification accuracy.

Main Results:

• The analog PIM-based CNN achieved a power consumption of 19 mW during MAC operations.
• Demonstrated an energy efficiency of 5.38 TOPS/W.
• Successfully implemented 8-bit high-resolution processing within a 180 nm CMOS process.

Conclusions:

• The developed on-chip analog PIM-CNN offers a power-efficient solution for biosensor data analysis.
• This integration enables sophisticated on-device processing for biosignal classification.
• The high energy efficiency and resolution pave the way for advanced wearable and implantable biosensing systems.