TensorFHE: Achieving Practical Computation on Encrypted Data Using GPGPU¶
Authors: Shengyu Fan; Zhiwei Wang; Weizhi Xu; Rui Hou; Dan Meng; Mingzhe Zhang
Year: 2023
Venue: HPCA 2023
type: library collection: catalogue status: auto
Links: arXiv:2212.14191
Summary¶
Auto-extracted summary (unreviewed)
TensorFHE accelerates CKKS on commodity GPGPUs by reformulating the NTT as matrix multiplications (via a Cooley-Tukey decomposition into three matrix products with precomputed twiddle matrices) so it can run on Tensor Core Units with far fewer modular reductions, and by batching FHE operations at the operation level to maximize throughput rather than single-operation latency. On an NVIDIA A100 it reaches 913 KOPS for NTT and 88 KOPS for homomorphic multiplication, 2.61x faster than the prior best GPU implementation and comparable to ASIC accelerators, including 2.9x faster than F1+ on a specific workload.
Abstract¶
In this paper, we propose TensorFHE, an FHE acceleration solution based on GPGPU for real applications on encrypted data. TensorFHE utilizes Tensor Core Units (TCUs) to boost the computation of Number Theoretic Transform (NTT), which is the part of FHE with highest time-cost. Moreover, TensorFHE focuses on performing as many FHE operations as possible in a certain time period rather than reducing the latency of one operation. Based on such an idea, TensorFHE introduces operation-level batching to fully utilize the data parallelism in GPGPU. We experimentally prove that it is possible to achieve comparable performance with GPGPU as with state-of-the-art ASIC accelerators. TensorFHE performs 913 KOPS and 88 KOPS for NTT and HMULT (key FHE kernels) within NVIDIA A100 GPGPU, which is 2.61x faster than state-of-the-art FHE implementation on GPGPU; Moreover, TensorFHE provides comparable performance to the ASIC FHE accelerators, which makes it even 2.9x faster than the F1+ with a specific workload. Such a pure software acceleration based on commercial hardware with high performance can open up usage of state-of-the-art FHE algorithms for a broad set of applications in real systems.
Source: arxiv
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | NTT formulated as matrix-matrix multiplication executed on GPU Tensor Core Units | llm | 0.80 |
| contribution | operation-level batching to exploit GPGPU data parallelism for throughput | llm | 0.80 |
| contribution | GPU implementation of CKKS kernels including bootstrapping (slim bootstrapping, BSGS, faster homomorphic DFT) | llm | 0.80 |
| contribution | demonstration that GPGPU can match ASIC FHE accelerators | llm | 0.80 |
| schemes | CKKS | llm | 0.80 |
| libraries | CUTLASS | llm | 0.80 |
| libraries | CUDA | llm | 0.80 |
| tasks | FHE acceleration (NTT, homomorphic multiplication, bootstrapping) | llm | 0.80 |
| hardware | NVIDIA A100 GPU | llm | 0.80 |
| interaction | non-interactive FHE | llm | 0.80 |
| code url | https://github.com/NVIDIA/cutlass (URL found in the full text) | regex | 0.50 |
| llm relevance | low | llm | 0.80 |
| key result | NTT throughput: 913 KOPS (NVIDIA A100) | llm | 0.80 |
| key result | HMULT throughput: 88 KOPS (NVIDIA A100) | llm | 0.80 |
| key result | speed-up over state-of-the-art GPU FHE implementation: 2.61x (NVIDIA A100) | llm | 0.80 |
| key result | speed-up over F1+ ASIC accelerator: 2.9x (specific workload) | llm | 0.80 |
System card: TensorFHE¶
| Group | 7. FHE libraries, GPU back-ends and compilers |
| Family | CKKS |
| Arithmetic | CUDA (tensor cores) |
| Interaction | GPU technique |
| Model scale | - |
| Status | Active |
| Relation to the SoK | Uses GPU tensor cores for NTT. |
| Note | Shows AI accelerators can be repurposed for FHE. |
| Systems page | TensorFHE |
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: CKKS 49 [p.1], BFV 3 [p.3], BGV 3 [p.3], Leveled / somewhat HE 3 [p.12]
- library: PyTorch 1 [p.8]
- model: ResNet 8 [p.8], RNN / LSTM 7 [p.8], Logistic regression 3 [p.8], CNN 1 [p.13]
- hardware: ASIC 20 [p.1], FPGA 12 [p.1], GPU 10 [p.1], CUDA 9 [p.5], NVIDIA A100 6 [p.1], NVIDIA V100 4 [p.8], Intel Xeon 1 [p.8]
- technique: NTT 118 [p.1], Rescaling 14 [p.3], Bootstrapping 12 [p.8], RNS 6 [p.3], Matrix multiplication 6 [p.3], Convolution 6 [p.2], Key switching 2 [p.3]
- assumption: LWE 3 [p.3], RLWE 2 [p.12], Lattice-based 2 [p.13]
- task: Genomics 2 [p.1]
- topic: Homomorphic encryption 112 [p.1], Benchmark 3 [p.4], Open source 2 [p.7]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/catalogue/7-111_tensorfhe-achieving-practical-computation-on-encrypted-data.pdf |
catalogue | 13 | 1.7 MB | 464554cc0edb |
Database key: tensorfhe-achieving-practical-computation-on-encrypted-data-using-gpgpu · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set tensorfhe-achieving-practical-computation-on-encrypted-data-using-gpgpu COLUMN VALUE