An ECRAM-Based Analog Compute-in-Memory Neuromorphic System with High-Precision Current Readout

This paper proposes a high-precision analog compute-in-memory (CIM) neuromorphic system that adopts a nonvolatile electro-chemical random-access memory (ECRAM) to improve linearity, symmetry, and endurance of the synapse array. For on-chip synapse training and inference, activation modules and matrix processing units adaptively form a neural network to perform analog-based update and read operations, respectively. The proposed neuromorphic system also utilizes current scaling and offset bias control to optimize the output sensing and matrix processing with ECRAM synapses. The 250-nm CMOS neuromorphic chip was fully verified with the 32 x 32 ECRAM synapse array, enabling linear update and accurate read operations. The proposed system can update and read the ECRAM synapse with 1000 weight levels, leading to high data throughput. The output error rates over 32 synapse read columns were measured within 2.59% when sweeping the weight level. The 32 x 32 ECRAM-based neuromorphic system consumes 5.9 mW when performing the inference.