ANP-I: A 28nm 1.5pJ/SOP Asynchronous Spiking Neural Network Processor Enabling Sub-0.1µJ/Sample On-Chip Learning for Edge-AI Applications

📋 论文概要

该论文提出了一款28nm工艺的异步脉冲神经网络处理器ANP-I，实现了1.5pJ/SOP的超低能耗推理和低于0.1µJ/样本的片上学习，解决了边缘AI应用中片上训练能耗过高的问题。

💡 主要创新点

• 采用异步电路设计，消除全局时钟，降低动态功耗，实现高能效脉冲神经网络推理。
• 提出低能耗片上学习机制，通过优化突触权重更新计算，使训练能耗降至亚0.1µJ/样本。
• 在28nm工艺下集成异步SNN处理核，支持在线学习，适用于边缘AI场景。

🏷 关键词

📄 原文摘要

With the development of on-chip learning processors for edge-AI applications, energy efficiency of NN inference and training is more and more critical. As on-chip training energy dominates the energy consumption of edge-AI processors [1,2,4,5], reduction is of paramount importance. Spiking neural networks (SNNs) offer energy-efficient inference and learning compared with convolutional neural networks (CNNs) or deepneural networks (DNNs), but SNN-based processors have three challenges that need to be addressed (Fig. 22.6.1). 1) During on-chip training, some factors involved in ∆W computation are zeros resulting in ∆W=0, leading to redundant ∆W computation and memory access for weight update. 2) After reaching a certain accuracy, more data cannot

👥 作者与机构

Jilin Zhang1, Dexuan Huo1, Jian Zhang1, Chunqi Qian1, Qi Liu1, Liyang Pan1,

Zhihua Wang1, Ning Qiao2, Kea-Tiong Tang3, Hong Chen1 Tsinghua University, Beijing, China, 2SynSense, Chengdu, China, National Tsing Hua University, Hsinchu, Taiwan 1 3