THE QUANTUM INTELLIGENCER

A paper deposited in the cryptology archive this day proposes a photonic accelerator for fully homomorphic encryption, and reports a speedup of between 2.14 and 5.10 times over a state-of-the-art electronic accelerator on four real-world workloads. The proposed machine, called PHAT, serves the scheme known as TFHE, fully homomorphic encryption over the torus, which permits arbitrary operations on encrypted data but has been limited in practice by the expense of bootstrapping, the periodic refresh that keeps a ciphertext computable. A large part of that expense lies in the Fast Fourier Transform. The authors argue that conventional electronic accelerators are checked by the limits of technology scaling and by the memory wall, and offer in their place an optically addressed phase-change memory operated as a processing-in-memory system, with a dataflow and a scheduler intended to keep the Fourier units occupied. The reported gain is measured against the state-of-the-art application-specific integrated circuit accelerator, across the four workloads the paper enumerates. The paper's account of its own difficulties is worth preserving, for the usual reasons. Directly mapping the Fourier transform onto optically addressed phase-change memory is not clean: the authors record the need for high-precision analog computation, and the high latency and energy cost of programming the optical cells. The architecture answers each in turn. It supplies OPCM-based FFT units, a twiddle-stationary dataflow arranged for the medium, and a scheduling mechanism whose object is to keep those units occupied. The reported speedup of 2.14 to 5.10 times against the state-of-the-art ASIC accelerator across four examined workloads is therefore not the whole finding. The whole finding is a claim that the dominant cost of TFHE, the refresh operation known as bootstrapping, can be moved onto a photonic memory plane without losing the precision the encryption scheme demands. —Inspector Grey Dispatch from The Prepared E0