Area-Efficient FHE Acceleration via Ciphertext-Level Parallelism and Shared-Key Reuse

Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

FHE accelerators typically optimize single-ciphertext throughput using wide vector lanes, which is inefficient for batched workloads due to repeated transfers of large evaluation keys. We propose an accelerator/runtime co-design exploiting ciphertext-level parallelism for ciphertexts sharing keys. Our hardware replaces wide-lane units with many small-lane clusters to process ciphertexts concurrently and retain keys on device, while a parameterizable multi-stage automorphism module supports large polynomials on narrow lanes. Transparently to applications, our OpenFHE integration batches identical operations across threads via a centralized queue and retains reusable objects on device. End-to-end CKKS evaluations, factoring in host-side and communication overheads, demonstrate that under a 128 MB on-chip buffer, the optimal configuration reduces per-ciphertext execution time by up to 44%, lowers PCIe transfers by 65%, and improves normalized compute-only throughput per area by 24.6%. These results prove ciphertext-level parallelism effectively complements residue-polynomial-level parallelism for data-movement-efficient FHE. © 2026 Copyright held by the owner/author(s).

키워드

Fully Homomorphic Encryption; Accelerator
제목
Area-Efficient FHE Acceleration via Ciphertext-Level Parallelism and Shared-Key Reuse
저자
Ahn, Youngtae; Kim, Jaehyung; Cho, Hyungmin
DOI
10.1145/3816440.3818540
발행일
2026-08-10
유형
Proceedings Paper
저널명
Proceedings of the International Symposium on Low Power Electronics and Design