Restoring-free Transposable eDRAM-based Computing-In-Memory System for SNN On-chip Unsupervised STDP Learning

초록

This paper proposes an embedded dynamic random-access memory (eDRAM) computing-in-memory (CIM)-based spiking neural network (SNN) processor for on-chip unsupervised learning. First, a novel transposable 3T-1C eDRAM bitcell was presented, offering a reduced area footprint with respect to previously reported CIM bitcells. The proposed bitcell enhances energy efficiency by leveraging its compatibility with the spike-timing-dependent plasticity (STDP) algorithm. Second, restoring-free access technique can be implemented through the instant weight-update operation, enabled by the read-calculate-write sequence in the learning operation. Third, to maximize the property of restoring-free accesses, a novel sense amplifier was employed, enhancing energy efficiency by 27% during the learning operation. Finally, a leaky integrate-and-fire bitline (LIF_BL) was proposed, eliminating the need for power-hungry ADCs by exploiting the characteristics of an SNN. The mixed-signal neuron model was implemented through a simple subtraction and counting method, increasing accuracy and energy efficiency. The proposed computing-in-eDRAM-based SNN processor was implemented using the 28 nm CMOS process, and 89.4% classification accuracy was achieved using the MNIST dataset, consuming 0.15 nJ/pixel of energy with an area of 0.57 mm2. © 2025 IEEE.

키워드