SpikeRAM: A 48.1pW/Synapse/Bit Event-Driven Spiking Compute-Near/In-Memory Processor with Neuromorphic Sensor Enabling Life-Long On-Chip Learning

📋 论文概要

本文提出SpikeRAM，一种事件驱动的尖峰计算近/内存储器处理器，集成了神经形态传感器，实现了超低功耗（48.1pW/Synapse/Bit）和464M突触的高效率处理，并支持事件驱动的推理与少样本学习。通过灰码权重和三进制梯度优化，减少了86%以上的内存写入次数，解决了边缘感知系统中的实时处理和在线学习问题。

💡 主要创新点

• 提出事件驱动的尖峰计算近/内存储架构（SpikeRAM），结合神经形态传感器，实现异步感知与特征提取。
• 采用灰码权重和三进制梯度技术，减少内存写入次数超过86%，显著降低功耗。
• 实现事件驱动推理与在线学习（e-OTBP算法），支持少样本事件签名验证。

🏷 关键词

📄 原文摘要

Switzerland, 5SynSense, Ningbo, China 1 4 Abstract SpikeRAM is a high efficiency (48.1pW/Synapse/Bit) memory-centric neuromorphic system with perception, computing and on-chip learning, achieving 464M synapses and 8.28mW power in real-time processing, while realizing few-shot learning for event-based signature verification. The EVS-sCNN core enables asynchronous sensing and feature extraction. The sFC-OCL core enables event-driven inference and learning with the e-OTBP algorithm. Graycode weights and ternary gradients reduce memory write times by over 86%. Edge perceptual SoCs increasingly demand high energy efficiency, privacy and adaptability for real-world tasks as shown in Fig. 18.4.1 (top) [1–6]. For perception, an event-based

👥 作者与机构

Haotian Fu1, Yue Zhou1, Zhuo Zhang1, Hongzhao Zheng1, Renxu Yang1, Yulong Huang1, Dezhen Yang2, Yannan Xing3, Tugba Demirci4, Ning Qiao5, Bojun Cheng1

Hong Kong University of Science and Technology, Guangzhou, China, 2North China Research Institute of Electro-Optics, Beijing, China, 3SynSense, Chengdu, China, SynSense, Zurich,