FlashTFHE: A Scalable Architecture for Efficient Multi-Bit Fully Homomorphic Encryption
Fully Homomorphic Encryption (FHE) enables secure processing of encrypted data in untrusted cloud environments, but its computational overhead remains a critical bottleneck. The FHE over Torus (TFHE) scheme, originally designed for Boolean gates that each require a costly bootstrapping operation, has been extended to encrypt multi-bit integers, yielding large end-to-end speedups by amortizing bootstrapping cost across many fast linear operations. However, current implementations are constrained to narrow numeric representations because securely supporting wider representations (i.e., beyond 4 bits) demands significantly larger cryptographic parameter sets. We reveal that these new parameters expose fundamental scalability barriers in state-of-the-art spatial accelerators by reshaping internal data dimensions and reuse patterns: scaling computation throughput causes severe memory bandwidth bottlenecks, mismatched data reuse patterns reduce hardware utilization, and buffering massive intermediate states becomes area-prohibitive. To address this, we introduce FlashTFHE, an accelerator featuring a temporal data reuse strategy. By time-multiplexing on-chip key data across parallel bootstrappings, FlashTFHE decouples computational throughput from memory bandwidth constraints. This architecture enables the use of highly efficient functional units and leverages adaptive batching to maximize utilization. FlashTFHE efficiently supports up to 10-bit ciphertexts, achieving speedups of up to 2529× over CPUs, 1320× over GPUs, and 7 × over the state-of-the-art TFHE accelerator.
