NeuDW-CIM: a 65-nm 0.8-pJ/Sop Reconfigurable Neuromorphic Compute-in-Memory Macro with Nonlinear Dendrites and K-Winners

2026-06-07

AI Summary

This work addresses the limited energy efficiency and insufficient capability for modeling biological plasticity in spiking neural networks deployed on in-memory computing architectures. To overcome these challenges, the authors propose a reconfigurable in-memory computing macro that supports nonlinear dendritic integration and K-winner-take-all sparse activation. Implemented in 65 nm CMOS technology, the design integrates a custom 9T ternary memory cell, a reconfigurable nonlinear in-memory ADC, and a dual-mode operation scheme, augmented with an early-termination strategy to reduce latency. Evaluated on the N-MNIST and DVS Gesture datasets, the system achieves accuracies of 97.2% and 95.5%, respectively, with an energy efficiency of 0.8 pJ/SOP, a 30% reduction in conversion latency, and a tenfold speedup in LIF neuron simulation. This is the first hardware implementation to jointly optimize dendritic functionality emulation and sparse activation.

Abstract

This work presents NeuDW-CIM, a highly efficient neuromorphic Compute-in-Memory (CIM) macro for Spiking Neural Networks (SNNs) implemented in 65 nm CMOS. The design introduces a custom twin 9T bit-cell for ternary in-puts/weights and a reconfigurable non-linear In-Memory ADC (IMA). The macro supports two specialized modes: 1) Nonlinear Dendrite (NLD) mode, which utilizes reconfigurable IMA to emulate biological dendritic functions, achieving measured accuracies of 97.2% on N-MNIST and 95.5% on DVS Gesture; and 2) Top-K Winner (KWN) mode, featuring an early-stopping mechanism that reduces IMA conversion latency by 30% and digital LIF latency by 10x. Benefiting from the sparse update in KWN mode, NeuDW-CIM achieves a measured energy efficiency (EE) of 0.8 pJ/SOP (1.6x improvement).

Problem

• Compute-in-Memory
• Spiking Neural Networks
• Nonlinear Dendrites
• Energy Efficiency
• Sparse Activation

Innovation

• Compute-in-Memory
• Spiking Neural Networks
• Nonlinear Dendrites
• Top-K Winner
• In-Memory ADC