PBSAcc: A High-Performance Hardware Accelerator for FHE With Improved Batch Programmable Bootstrapping

As an important privacy computing method, fully homomorphic encryption (FHE) faces a serious computational bottleneck. To accelerate it, there are three kinds of accelerators: CKKS/BGV accelerators, TFHE/FHEW accelerators, and cross-scheme accelerators. However, none of them supports SIMD-style nonlinear function operations or batched Boolean operations. To solve this problem, this paper proposes a high-performance FHE accelerator (named PBSAcc) with batch programmable bootstrapping (Batch-PBS), which provides batch processing capabilities for TFHE/FHEW PBS. For Batch-PBS, we replace the highly computational steps in the original algorithm with RingSwitch and rotation-free homomorphic matrix-vector multiplication to reduce the amount of computation. In PBSAcc, we propose a modified Montgomery modular multiplication algorithm for specific FHE-friendly primes with Karatsuba multiplication to reduce the underlying multiplier requirement of the modular multiplier, and a 3D automorphism method to realize the vectorization of automorphism. The experimental results show that, in typical FHE workloads, such as PBS, Zama-NN-20, private information retrieve (PIR), and Pack, PBSAcc achieves a 1.16× to 1,193× performance improvement and a 1.46× to 833.33× efficiency improvement over other FHE accelerators.