JSSC 2021第7期Memory65nm

A 044-μJdec 399-μsdec Recurrent Attention In-Memory Processor for Keyword Spotti

65nm CMOS工艺下基于深度学习的低功耗关键词识别芯片

65nm CMOS, 39.9µs决策延迟, <0.5µJ/dec决策能耗

关键词识别内存计算深度学习能效优化硬件算法协同设计

▸提出KeyRAM算法，通过置信度计算降低计算复杂度

▸采用内存计算架构与数字协处理器结合，提升能效

▸提出稀疏感知求和方案，优化稀疏激活计算

Abstract

This article presents a deep learning-based classiﬁer IC for keyword spotting (KWS) in 65-nm CMOS designed using an algorithm-hardware co-design approach. First, a recur- rent attention model (RAM) algorithm for the KWS task (the KeyRAM algorithm) is proposed. The KeyRAM algorithm enables accuracy versus energy scalability via a conﬁdence-based computation (CC) scheme, leading to a 2 .5× reduction in compu- tational complexity compared to state-of-the-art (SOTA) neural networks, and is well-suit