A Multi-bit ECRAM-Based Analog Neuromorphic System with High-Precision Current Readout Achieving 97.3% Inference Accuracy

• Published in: IEEE transactions on biomedical circuits and systems 2024-09, Vol.PP, p.1-14
• Main Authors: Um, Minseong, Kang, Minil, Eom, Kyeongho, Kwak, Hyunjeong, Noh, Kyungmi, Lee, Jimin, Son, Jeonghoon, Kwon, Jiseok, Kim, Seyoung, Lee, Hyung-Min
• Format: Article
• Language: English
• Subjects: Arrays; CMOS; current scaling; ECRAM; Field programmable gate arrays; matrix processing; Neural networks; neuromorphic system; Neuromorphics; non-volatile memory; Nonvolatile memory; Synapses; Training
• ISSN: 1932-4545; 1940-9990; 1940-9990
• DOI: 10.1109/TBCAS.2024.3465610

This article proposes an analog neuromorphic system that enhances symmetry, linearity, and endurance by using a high-precision current readout circuit for multi-bit nonvolatile electro-chemical random-access memory (ECRAM). For on-chip training and inference, the system uses activation modules and matrix processing units to manage analog update/read paths and perform precise output sensing with feedback-based current scaling on the ECRAM array. The 250nm CMOS neuromorphic chip was tested with a 32 x 32 ECRAM synaptic array, achieving linear and symmetric updates and accurate read operations. The proposed circuit system updates the 32 x 32 ECRAM across 100 levels, maintaining consistent synaptic weights, and operates with an output error rate of up to 2.59% per column. It consumes 5.9 mW of power excluding the ECRAM array and achieves 97.3% inference accuracy on the MNIST dataset, close to the software-confirmed 97.78%, with only the final layer (64 x 10) mapped to the ECRAM.