Cheddar: A Swift Fully Homomorphic Encryption Library Designed for GPU Architectures¶
Authors: Wonseok Choi; Jongmin Kim; Jung Ho Ahn
Year: 2024
Venue: arXiv 2024
type: library collection: catalogue status: auto
Links: DOI · arXiv:2407.13055
Summary¶
Auto-extracted summary (unreviewed)
Cheddar is a GPU library for CKKS FHE that systematically enables 32-bit execution through a novel '25-30 prime' RNS construction, architecture-aware kernel optimizations using the GPU's 32-bit integer datapath, and extensive kernel fusion to cut memory-bandwidth pressure. It executes representative FHE workloads 2.18-4.45x faster than WarpDrive, outperforms custom FPGA designs, and is 1.53-1.76x faster than Phantom on basic mechanisms.
Abstract¶
Fully homomorphic encryption (FHE) frees cloud computing from privacy concerns by enabling secure computation on encrypted data. However, its substantial computational and memory overhead results in significantly slower performance compared to unencrypted processing. To mitigate this overhead, we present Cheddar, a high-performance FHE library for GPUs, achieving substantial speedups over previous GPU implementations. We systematically enable 32-bit FHE execution, leveraging the 32-bit integer datapath within GPUs. We optimize GPU kernels using efficient low-level primitives and algorithms tailored to specific GPU architectures. Further, we alleviate the memory bandwidth burden by adjusting common FHE operational sequences and extensively applying kernel fusion. Cheddar delivers performance improvements of 2.18--4.45$\times$ for representative FHE workloads compared to state-of-the-art GPU implementations.
Source: arxiv
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | systematic 32-bit FHE execution on GPUs | llm | 0.85 |
| contribution | 25-30 prime RNS construction method | llm | 0.85 |
| contribution | architecture-aware low-level GPU kernel optimizations | llm | 0.85 |
| contribution | adjusted operational sequences and extensive kernel fusion to reduce DRAM traffic | llm | 0.85 |
| contribution | full-fledged GPU FHE library with a high-level interface | llm | 0.85 |
| schemes | CKKS | llm | 0.85 |
| libraries | Phantom | llm | 0.85 |
| libraries | WarpDrive | llm | 0.85 |
| tasks | GPU acceleration of FHE | llm | 0.85 |
| hardware | NVIDIA RTX 4090 GPU | llm | 0.85 |
| security level | 128-bit | llm | 0.85 |
| interaction | non-interactive FHE | llm | 0.85 |
| code url | https://github.com/Desilo/liberate-fhe (URL found in the full text) | regex | 0.50 |
| llm relevance | low | llm | 0.85 |
| key result | workload speed-up: 2.18x to 4.45x (representative FHE workloads vs WarpDrive GPU implementation) | llm | 0.85 |
| key result | basic mechanism speed-up: 1.53x to 1.76x (vs Phantom) | llm | 0.85 |
| key result | comparison with FPGA: outperforms custom FPGA designs (FHE workloads) | llm | 0.85 |
System card: Cheddar¶
| Group | 7. FHE libraries, GPU back-ends and compilers |
| Family | CKKS |
| Arithmetic | CUDA |
| Interaction | GPU library |
| Model scale | - |
| Status | Active SOTA |
| Relation to the SoK | Among the fastest reported GPU CKKS bootstrapping implementations. |
| Note | Uses 32-bit word arithmetic. Code availability limited. |
| Systems page | Cheddar |
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: CKKS 18 [p.1], RSA 2 [p.13], Leveled / somewhat HE 2 [p.13]
- library: Cheddar 71 [p.1], HEaaN (library) 6 [p.9], Phantom 6 [p.11], HEonGPU 4 [p.11], Liberate.FHE 4 [p.11], OpenFHE 1 [p.12], Microsoft SEAL 1 [p.13], HElib 1 [p.13], Lattigo 1 [p.14], EVA / CHET 1 [p.13]
- model: ResNet 19 [p.1], CNN 10 [p.1], Logistic regression 2 [p.9], RNN / LSTM 1 [p.14]
- dataset: CIFAR-10 1 [p.1]
- hardware: GPU 103 [p.1], NVIDIA RTX 24 [p.2], NVIDIA A100 21 [p.1], FPGA 14 [p.2], NVIDIA H100 12 [p.9], NVIDIA V100 6 [p.9], CUDA 5 [p.8], ASIC 3 [p.1], AMD EPYC 1 [p.10], AVX-512 1 [p.13]
- technique: Rescaling 64 [p.2], RNS 57 [p.1], NTT 37 [p.2], Bootstrapping 13 [p.1], Fine-tuning 8 [p.2], Convolution 7 [p.1], ReLU 2 [p.9], Rotations 2 [p.2]
- assumption: 128-bit security 3 [p.9], Lattice-based 2 [p.13]
- topic: Homomorphic encryption 68 [p.1], Open source 5 [p.11]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/catalogue/7-100_cheddar-a-swift-fully-homomorphic-encryption-library-designe.pdf |
catalogue | 15 | 1.1 MB | 3f5bd059a308 |
Database key: cheddar-a-swift-fully-homomorphic-encryption-library-designed-for-gpu · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set cheddar-a-swift-fully-homomorphic-encryption-library-designed-for-gpu COLUMN VALUE