CoFHEE: A Co-processor for Fully Homomorphic Encryption Execution

The migration of computation to the cloud has raised privacy concerns as sensitive data becomes vulnerable to attacks since they need to be decrypted for processing. Fully Homomorphic Encryption (FHE) mitigates this issue as it enables meaningful computations to be performed directly on encrypted data. Nevertheless, FHE is orders of magnitude slower than unencrypted computation, which hinders its practicality and adoption. Therefore, improving FHE performance is essential for its real world deployment. In this paper, we present a year-long effort to design, implement, fabricate, and post-silicon validate a hardware accelerator for Fully Homomorphic Encryption dubbed CoFHEE. With a design area of $12mm^2$, CoFHEE aims to improve performance of ciphertext multiplications, the most demanding arithmetic FHE operation, by accelerating several primitive operations on polynomials, such as polynomial additions and subtractions, Hadamard product, and Number Theoretic Transform. CoFHEE supports polynomial degrees of up to $n = 2^{14}$ with a maximum coefficient sizes of 128 bits, while it is capable of performing ciphertext multiplications entirely on chip for $n \leq 2^{13}$. CoFHEE is fabricated in 55nm CMOS technology and achieves 250 MHz with our custom-built low-power digital PLL design. In addition, our chip includes two communication interfaces to the host machine: UART and SPI. This manuscript presents all steps and design techniques in the ASIC development process, ranging from RTL design to fabrication and validation. We evaluate our chip with performance and power experiments and compare it against state-of-the-art software implementations and other ASIC designs. Developed RTL files are available in an open-source repository.