A 28nm 0.48mJ/boot Torus FHE Processor for Arbitrary Computation on Encrypted Data

📋 论文概要

本文提出了一款基于28nm CMOS工艺的3.24mm² torus全同态加密(FHE)处理器，通过安全保留的低比特宽度torus量化、分治多项式乘法引擎以及无冲突原子swizzling与全深度流水线技术，分别降低了62%的延迟、提升29%的能效并减少96%的内存需求。该处理器在bootstrapping操作中相比CPU实现了196倍的能效提升，支持对加密数据的任意计算。

💡 主要创新点

• 提出安全保留的低比特宽度torus量化与密钥提示，将密钥提示延迟减少62%。
• 采用分治策略的创新多项式乘法引擎，提升能量效率29%。
• 实现无冲突原子swizzling与全深度流水线，使内存大小减少96%。

🏷 关键词

📄 原文摘要

Abstract This work presents a 3.24mm2 torus fully homomorphic encryption (FHE) processor fabricated in 28nm CMOS, featuring: 1) a security-preserving low-bit-width torus quantization with keys hints reducing latency by 62%, 2) a divide-and-conquer innovative polynomial multiplication engine improving energy efficiency by 29%, and 3) a conflict-free atomic swizzling with full-depth pipelining reducing memory size by 96%. It delivers 196× higher energy efficiency in bootstrapping compared to a CPU. Fully homomorphic encryption (FHE) has recently emerged as a promising technology for privacy-preserving applications, as it enables computations directly on encrypted data without requiring decryption. State-of-the-art FHE schemes can be broadly classified into two types: arithmetic FHE, exemplified by CKKS, which enables homomorphic arithmetic operations [2-4], and logic FHE, represented by TFHE, which supports homomorphic

👥 作者与机构

Xinglong Yu, Yi Sun, Yifan Zhao, Honglin Kuang, Song Wang, Zhen Yang, Jun Han

Fudan University, Shanghai, China