Helios: Melting Kernel Boundaries for GPU-Accelerated HE via Graph Rewriting and Microarchitecture-Aware Mapping
Recommended citation: Helios: Melting Kernel Boundaries for GPU-Accelerated HE via Graph Rewriting and Microarchitecture-Aware Mapping. Yi Chen, Ziyu Tang, Chao Yang, Guang Fan, Mingzhe Zhang, Meng Li. 59th IEEE/ACM International Symposium on Microarchitecture. 2026.
Abstract
Homomorphic encryption (HE) enables computation on encrypted data, but its practical deployment remains limited by the high cost of real workloads. GPUs are an attractive platform for HE acceleration, yet modern FHE workloads remain strongly memory-bound because even a single ciphertext often exceeds GPU on-chip storage by one to two orders of magnitude, forcing frequent DRAM traffic. Kernel fusion is therefore essential, but existing GPU FHE systems still fail to realize its full benefit. Two obstacles dominate: Number Theoretic Transform (NTT)-containing cryptographic operators fragment the graph into small fusible regions, while naive fusion over enlarged regions often causes severe hardware inefficiency.
We show that FHE programs should not be optimized through opportunistic fusion over a fixed operator graph. Instead, end-to-end execution is largely concentrated in a small set of recurring hot modules, which expose opportunities beyond local fusion. Based on this view, we present Helios, a compiler-driven GPU optimization framework for FHE that fits naturally in existing FHE compilation stacks. Helios combines FHE-specific graph rewriting with hardware-aware mapping: it treats hot modules as the optimization unit, rewrites them to defer, fold, or eliminate movable barriers and restructure dependency-heavy subgraphs, and maps the resulting fused regions efficiently to GPUs. Across representative CKKS workloads, Helios achieves 1.54x–2.26x speedup on A100-80GB and up to 2.49x on H100 over strong compiler-generated GPU baselines. On rotation-accumulation-dominated workloads, Helios on A100-80GB already outperforms a prior expert-tuned H100 result.
