A Multi-Bit ECRAM-Based Analog Neuromorphic System With High-Precision Current Readout Achieving 97.3% Inference...

Area of Science:

• Neuromorphic Engineering
• Solid-State Circuits
• Artificial Intelligence Hardware

Background:

• Neuromorphic systems aim to mimic the human brain's efficiency and structure.
• Analog memory devices like electro-chemical random-access memory (ECRAM) offer potential for low-power AI hardware.
• Challenges remain in achieving high precision, linearity, and endurance in analog neuromorphic circuits.

Purpose of the Study:

• To propose and validate an analog neuromorphic system utilizing a high-precision current readout circuit.
• To enhance symmetry, linearity, and endurance in multi-bit nonvolatile ECRAM.
• To enable efficient on-chip training and inference for AI applications.

Main Methods:

• Development of a 250nm CMOS neuromorphic chip featuring a 32x32 ECRAM synaptic array.
• Integration of activation modules and matrix processing units for managing analog paths.
• Implementation of feedback-based current scaling for precise output sensing.

Main Results:

• Achieved linear and symmetric weight updates across 100 levels in the ECRAM array.
• Demonstrated accurate read operations with an output error rate below 2.59% per column.
• Attained 97.3% inference accuracy on the MNIST dataset, closely matching software performance.

Conclusions:

• The proposed analog neuromorphic system effectively enhances ECRAM performance for AI tasks.
• The high-precision current readout circuit is crucial for achieving linearity, symmetry, and endurance.
• This work represents a significant step towards efficient, on-chip AI hardware.