| name | memristor-snn-interception-task |
| description | Memristor-based spiking neural network accelerator for bio-inspired interception tasks - achieving 12.7x energy reduction vs digital SNN (arXiv:2605.31299v1, May 2026). |
| version | 2.0.0 |
| category | neuromorphic |
| tags | ["spiking-neural-network","memristor","neuromorphic-hardware","analog-computation","energy-efficient","edge-intelligence","interception"] |
| arxiv_id | 2605.31299v1 |
| authors | ["Qianhou Qu","Sheng Lu","Liuting Shang","Jaihan Utailawon","Sungyong Jung","Qilian Liang","Chenyun Pan"] |
| published | 2026-05-29T00:00:00.000Z |
| conference | IEEE Dallas Circuits and Systems Conference (DCAS 2026) |
Memristor-Based SNN Accelerator for Interception Tasks
Overview
This paper presents an analog memristor-based spiking neural network (SNN) accelerator that integrates in-memory synaptic computation with analog integrate-and-fire neurons, achieving significant energy efficiency gains over digital implementations.
Key Achievement: 12.7x lower energy consumption and 1.26x lower latency compared to digital SNN baseline at 5nm technology node.
Hardware Architecture
Core Components
-
In-Memory Synaptic Computation
- Uses memristor crossbar arrays for synaptic weight storage
- Eliminates multi-transistor CMOS synapse circuits
- Performs analog matrix-vector multiplication (MVM) in memory
-
Analog Integrate-and-Fire (IF) Neurons
- Implemented with analog circuits (not digital counters)
- Threshold detection via analog comparator
- Spike generation through analog pulse circuits
-
Event-Driven Operation
- Asynchronous spike processing
- No global clock required
- True neuromorphic computing paradigm
Technology Comparison
| Metric | Analog SNN (45nm) | Digital SNN (5nm) |
|---|
| Technology | 45nm | 5nm (advanced) |
| Energy per inference | 12.7x lower | Baseline |
| Latency | 1.26x lower | Baseline |
| Synapse implementation | Memristor arrays | CMOS circuits |
| Neuron type | Analog IF | Digital IF |
Bio-Inspired Interception Task
Predator-Prey Tracking
- Task: Simulate pursuit behavior (predator tracking prey)
- Input: Position and velocity of prey
- Output: Pursuit trajectory of predator
- Network: Feedforward SNN with trained weights
Performance Results
- Mean Squared Error (MSE): 0.004 (very close to ideal software inference)
- : Superior to digital baseline despite older technology node
