Neuromorphic Computing Using NAND Flash Memory Architecture With Pulse Width Modulation Scheme.

Area of Science:

• Computer Science
• Electrical Engineering
• Materials Science

Background:

• Neuromorphic computing aims to mimic the human brain's structure and function.
• NAND flash memory offers high density and non-volatility, making it a potential candidate for neuromorphic hardware.
• Existing neuromorphic systems often require specialized hardware, increasing costs and complexity.

Purpose of the Study:

• To propose a novel operation scheme for high-density and robust neuromorphic computing using conventional NAND flash memory architecture.
• To implement a neuromorphic system without modifying the existing NAND flash memory architecture.
• To investigate the impact of quantization training (QT) versus post-training quantization (PTQ) on classification accuracy with 4-bit synaptic weights.

Main Methods:

• Analog input is encoded using a pulse width modulation (PWM) circuit.
• Synaptic weights are represented by the adjustable conductance of NAND cells with 4-bit precision.
• Multiply-accumulate (MAC) operations are performed in a single step using the proposed scheme.
• The read-verify-write (RVW) scheme is employed to achieve low current variance in NAND cells.

Main Results:

• The proposed scheme integrates seamlessly with conventional NAND flash memory architecture.
• Saturated current-voltage characteristics of NAND cells mitigate issues like serial resistance and IR drop.
• The system achieves high classification accuracies of 98.14% on MNIST and 89.6% on CIFAR10 datasets.
• Quantization training (QT) demonstrates effectiveness for 4-bit weight precision in neuromorphic applications.

Conclusions:

• The novel operation scheme enables efficient and robust neuromorphic computing on standard NAND flash memory.
• The integration of PWM for input encoding and adjustable conductance for weights simplifies neuromorphic system implementation.
• The study validates the feasibility of using NAND flash memory for advanced AI tasks, achieving competitive accuracy levels.