Content area
Fully Homomorphic Encryption (FHE) allows computation on encrypted data, enabling Private Inference (PI) in Machine Learning models. This comes at a high computational cost, which prohibits widespread adoption of PI despite its benefits. This paper presents a programmable hardware-software stack to accelerate FHE kernels, extending the work of a custom Vector Processor named the Ring Processing Unit (RPU). In contrast to prior work, which focused on fixed-function accelerators, the RPU presents an efficient and highly programmable ASIC that leverages high parallelism through an optimized OOO compiler to provide performance competitive with SOTA.
Experimental results prove that RPU achieves a significant speedup over CPU for FHE kernels, and the OOO compiler further boosts its performance up to 1.4×. Moreover, the RPU achieves competitive results to BASALISC with 12× fewer multipliers and higher programmability. These results show that performance comparable to ASICs can be achieved by simply leveraging the highly parallel nature of FHE kernels.