HPU Deepdive: How FHE Operations Run on the Homomorphic Processor

The Homomorphic Processing Unit (HPU), designed by Zama, is a hardware accelerator that executes computations directly on encrypted data using the TFHE (Fully Homomorphic Encryption over Torus) scheme, focusing exclusively on ciphertext operations while interacting with a host CPU that manages cryptographic keys and prepares data. The HPU uses a dedicated protocol for communication with the host and employs a radix representation to manage ciphertext sizes, allowing integers to be split into smaller, more manageable encrypted digits called elementary ciphertexts. It operates on these elementary ciphertexts by translating high-level integer operations (IOp) into low-level digit operations (DOp) for execution, while handling noise through techniques like Programmable Bootstrapping (PBS) and Key Switching (KS) to ensure the integrity of computations. The HPU's ability to execute unlimited computations on TFHE ciphertexts is complemented by its parallel processing capabilities and noise management, bridging advanced cryptography with hardware acceleration for efficient encrypted computations.