OmniCrypt: A 435.86M-GOPS/W Bootstrappable Multi-Scheme FHE Accelerator with On-Chip Data Generation for Privacy-Preserving Computation

📋 论文概要

本文提出OmniCrypt，一款28nm CMOS工艺下硅验证的多方案全同态加密(FHE)加速器，支持自举(bootstrapping)和方案切换。通过集成三个专用引擎和片上数据生成，有效减少内存和计算开销，实现高达435.9M-GOPS/W的能效和252.8M-GOPS的性能，在CKKS自举吞吐量上达到最高，相比CPU加速107倍。

💡 主要创新点

• 首次提出支持自举和方案切换的多方案FHE加速器架构。
• 集成三个专用引擎（外部乘积计算、片上数据生成、SAM3模乘器）以最小化内存和计算开销。
• 实现业界最高的CKKS自举吞吐量和能效。

🏷 关键词

📄 原文摘要

*Equally Credited Authors (ECAs) Abstract OmniCrypt is a silicon-proven multi-scheme FHE accelerator that supports both bootstrapping and scheme switching. Fabricated in 28nm CMOS (12.96mm2), it integrates three specialized engines for efficient external product computation, on-chip data generation, and the SAM3 modular multiplier to minimize memory and compute overhead. OmniCrypt achieves 252.8M-GOPS, 19.5M-GOPS/mm2, and 435.9M-GOPS/W, delivering the highest CKKS bootstrapping throughput and 107× CPU speedup in application benchmarks. Fully homomorphic encryption (FHE) is cryptographic technique that enables direct computation on encrypted data, preserving privacy throughout the process. Security in FHE relies on adding noise to ciphertexts [9], but this noise grows with each operation, eventually limiting the number of computations that can be performed. To enable deeper computations, a process called bootstrapping (BTS) is used to refresh the ciphertext and reset its noise

👥 作者与机构

Adiwena Putra*, Hyunjun Cho*, Sungwoong Yune, Cuong Manh Duong, Jaeho Jeon, Joo-Young Kim

KAIST, Daejeon, Korea