PEFT: a near-memory processing-enabled heterogeneous accelerator for BatchPBS TFHE

Fully Homomorphic Encryption (FHE) is increasingly adopted for private deep neural network (DNN) inference due to its formal privacy guarantees. Torus-FHE (TFHE) supports accurate evaluation of nonlinear functions, but has to evaluate the intermediate activations one by one through bootstrapping. To address this challenge, the Batched Programmable Bootstrapping (BatchPBS) algorithm was introduced. Although it enables batched nonlinear function evaluation, we observe it still suffers from significant latency cost, primarily due to the off-chip memory access.To overcome this limitation, we propose PEFT, a heterogeneous accelerator for BatchPBS TFHE based on near-memory processing (NMP). PEFT features an optimized framework designed to minimize data movement for BatchPBS. At its core is a co-designed NMP unit that executes various modular integer arithmetic instructions, supported by an efficient data mapping scheme and execution flow. Our integrated approach significantly enhances the performance and energy efficiency of TFHE DNN inference workloads.