MNEMOS: A GPU-Based TFHE Acceleration Framework with Memory Access Optimization

Fully Homomorphic Encryption over Torus (TFHE) provides a promising approach for privacy-preserving computing by enabling computation directly on encrypted data. However, this strong security guarantee comes at the cost of enormous computational overhead compared with plaintext computation. Modern Graphics Processing Units (GPUs), equipped with thousands of parallel computing cores and high memory bandwidth, offer an attractive platform for accelerating TFHE workloads. By exploiting their massive parallelism, the latency of TFHE primitives can be significantly reduced, making privacy-preserving computing practical. Nevertheless, executing the TFHE applications on the GPU remains limited. In this paper, we propose MNEMOS, a TFHE acceleration framework for GPU platform, optimizing the memory access during TFHE execution. In our study, we observe that severe pipeline stalls occur during TFHE kernel execution, primarily caused by frequent memory accesses and cache misses, which significantly degrade overall performance. Moreover, the utilization of Tensor Cores (TCUs) is far from optimal. Due to the excessive memory access latency caused by frequent memory accesses, the computational throughput of TCUs cannot be fully exploited. To address these issues, we propose a memory-aware algorithmic optimization that improves reuse efficiency through data re-layout and access scheduling. In addition, we introduce a Tensor-Core-optimized FFT mapping strategy that mitigates performance degradation caused by cache misses and enhances the effective utilization of TCUs during PBS computation. Experiments demonstrate that our optimizations highly enhance the performance of TFHE-based applications by $\mathbf{1. 9 6} \times$ on average and up to $\mathbf{2. 2 3} \times$ in the best case.