C-Transformer: A 2.6-18.1μJ/Token Homogeneous DNN-Transformer/Spiking-Transformer Processor with Big-Little Network and Implicit Weight Generation for Large Language Models

📋 论文概要

该论文提出了一种名为C-Transformer的同质DNN-Transformer/Spiking-Transformer处理器，通过大-小网络架构实现了2.6-18.1μJ/Token的高能效处理，解决了传统Transformer模型在边缘设备上计算和能耗过高的问题。

💡 主要创新点

• 提出大-小网络架构，结合DNN和Spiking神经网络，实现灵活高效的混合模型处理。
• 采用同质处理器设计，支持DNN-Transformer和Spiking-Transformer两种模式的无缝切换。
• 通过优化数据流和计算单元，在低功耗下实现高吞吐量，适用于实时边缘AI应用。

🏷 关键词

📄 原文摘要

20.5.1, are widely used, and even on-device LLM systems with real-time responses are anticipated

👥 作者与机构

Sangyeob Kim, Sangjin Kim, Wooyoung Jo, Soyeon Kim, Seongyon Hong, Hoi-Jun Yoo

Korea Advanced Institute of Science and Technology, Daejeon, Korea Recently, transformer-based large language models (LLMs), shown in Fig