A 16Mb 166.8TOPS/W Near-Memory Phase-Domain-Computing Ferroelectric NAND Flash for Approximate Nearest Neighbor Search on Edge Devices

📋 论文概要

本文提出并实现了一款16Mb近内存相位域计算（NM-PDC）铁电NAND闪存芯片，可用于近似最近邻搜索。该芯片通过在近内存区域进行相位域计算，解决了传统NAND近/存内计算IO宽度有限、能量延迟积大且无法支持多样向量格式的问题，单次搜索操作可计算512个256维4b向量的相似距离，能效达166.8TOPS/W，端到端延迟减少12.8倍。

💡 主要创新点

• 提出近内存相位域计算（NM-PDC）架构，将计算置于存储近端而非存储单元内，突破IO瓶颈并支持多样向量格式。
• 采用铁电NAND闪存（FeNAND）实现高能效近内存计算，利用相位域信号处理降低能耗和延迟。
• 在16Mb阵列中实现512路并行相似距离计算，支持256维4b向量，能效达166.8TOPS/W，适用于大规模近似最近邻搜索。

🏷 关键词

📄 原文摘要

2University of Chinese Academy of Sciences, Beijing, China Columbia University, New York, NY *Equally Credited Authors (ECAs) 1 3 Abstract Previous near-memory computing (NMC) or in-memory-computing (IMC) NANDs suffers from limited IO width, large energy-delay-product, and an inability to support diverse vector formats. This work presents a fabricated 16Mb near-memory phase-domain-computing (NM-PDC) FeNAND chip can compute the 512 similarity distances between 256-dimensional 4b vectors in a single search operation, achieving 166.8TOPS/W energy efficiency and a 12.8× reduction in end-to-end search latency. Approximate nearest neighbor search (ANNS) for sparse, high-dimensional data, such as text embeddings, is widely used in applications like natural language processing,

👥 作者与机构

Weizeng Li*1,2, Bohan Wang*1,2, Zhidao Zhou1,2, Junyu Zhu1,2, Zhi Li1,2, Junzhe Shen1,2, Wenfeng Zha1,2, Zhongze Han1,2, Yiman Wang1,2, Linfang Wang3,

Hongyang Hu1,2, Qing Luo1,2, Chunmeng Dou1,2, Ming Liu1 Institute of Microelectronics of the Chinese Academy of Sciences, Beijing, China,