F1: A Fast and Programmable Accelerator for Fully Homomorphic Encryption¶
Authors: Axel Feldmann; Nikola Samardzic; Aleksandar Krastev; Srini Devadas; Ron Dreslinski; Karim Eldefrawy; Nicholas Genise; Chris Peikert; Daniel Sanchez
Year: 2021
Venue: MICRO 2021
type: accelerator collection: catalogue, sok-references status: auto
Links: arXiv:2109.05371
Summary¶
Auto-extracted summary (unreviewed)
F1 is the first programmable FHE accelerator able to run complete FHE programs: a wide-vector processor with functional units specialized for NTTs, automorphisms and modular arithmetic, an explicitly managed memory hierarchy, and a compiler that schedules on- and off-chip data movement to maximize reuse. Evaluated with cycle-accurate simulation and RTL synthesis (151 mm2, 180 W), F1 outperforms CPU software by a geometric-mean 5,432x (1,195x to 17,412x), bringing LoLa-CIFAR inference from 20 minutes to 241 ms.
Abstract¶
Fully Homomorphic Encryption (FHE) allows computing on encrypted data, enabling secure offloading of computation to untrusted serves. Though it provides ideal security, FHE is expensive when executed in software, 4 to 5 orders of magnitude slower than computing on unencrypted data. These overheads are a major barrier to FHE's widespread adoption. We present F1, the first FHE accelerator that is programmable, i.e., capable of executing full FHE programs. F1 builds on an in-depth architectural analysis of the characteristics of FHE computations that reveals acceleration opportunities. F1 is a wide-vector processor with novel functional units deeply specialized to FHE primitives, such as modular arithmetic, number-theoretic transforms, and structured permutations. This organization provides so much compute throughput that data movement becomes the bottleneck. Thus, F1 is primarily designed to minimize data movement. The F1 hardware provides an explicitly managed memory hierarchy and mechanisms to decouple data movement from execution. A novel compiler leverages these mechanisms to maximize reuse and schedule off-chip and on-chip data movement. We evaluate F1 using cycle-accurate simulations and RTL synthesis. F1 is the first system to accelerate complete FHE programs and outperforms state-of-the-art software implementations by gmean 5400x and by up to 17000x. These speedups counter most of FHE's overheads and enable new applications, like real-time private deep learning in the cloud.
Source: arxiv
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | first programmable FHE accelerator that executes full FHE programs | llm | 0.90 |
| contribution | architectural analysis of FHE workloads revealing acceleration opportunities | llm | 0.90 |
| contribution | vector functional units specialized for NTT, automorphisms and modular multiplication | llm | 0.90 |
| contribution | explicitly managed memory hierarchy with data movement decoupled from execution | llm | 0.90 |
| contribution | compiler that maximizes reuse and schedules off-chip and on-chip data movement | llm | 0.90 |
| schemes | CKKS | llm | 0.90 |
| schemes | BGV | llm | 0.90 |
| libraries | HElib | llm | 0.90 |
| libraries | HEAAN | llm | 0.90 |
| models | LoLa-MNIST (LeNet-style CNN) | llm | 0.90 |
| models | LoLa-CIFAR (6-layer CNN) | llm | 0.90 |
| models | logistic regression (HELR) | llm | 0.90 |
| datasets | MNIST | llm | 0.90 |
| datasets | CIFAR-10 | llm | 0.90 |
| tasks | private neural network inference | llm | 0.90 |
| tasks | encrypted logistic regression training | llm | 0.90 |
| tasks | encrypted database lookup | llm | 0.90 |
| tasks | bootstrapping | llm | 0.90 |
| hardware | F1 accelerator (cycle-accurate simulation and RTL synthesis, 16 compute clusters, 2x HBM2) | llm | 0.90 |
| hardware | Xeon E3-1240v5 4-core 3.5 GHz CPU baseline | llm | 0.90 |
| threat model | untrusted server; client encrypts data and offloads computation | llm | 0.90 |
| interaction | non-interactive FHE | llm | 0.90 |
| code url | https://github.com/homenc/HElib/tree/master/examples/BGV_country_db_lookup (URL found in the full text) | regex | 0.50 |
| limitations | does not support B/FV, so LoLa benchmarks were ported to CKKS | llm | 0.90 |
| limitations | results come from simulation and RTL synthesis rather than fabricated silicon | llm | 0.90 |
| limitations | CKKS bootstrapping is memory-bound and obtains the lowest speed-up | llm | 0.90 |
| llm relevance | medium | llm | 0.90 |
| key result | speed-up: 5,432x gmean (1,195x to 17,412x) (F1 vs 4-core Xeon E3-1240v5 on full FHE benchmarks) | llm | 0.90 |
| key result | latency: 241 ms (5,011x faster than 20 min on CPU) (LoLa-CIFAR inference, unencrypted weights, CKKS) | llm | 0.90 |
| key result | latency: 0.17 ms (17,412x) / 0.36 ms (15,086x) (LoLa-MNIST with unencrypted / encrypted weights) | llm | 0.90 |
| key result | latency: 1.15 ms (7,217x) (HELR logistic regression batch, 256 features, 256 samples) | llm | 0.90 |
| key result | bootstrapping latency: 2.40 ms BGV (1,830x), 1.30 ms CKKS (1,195x) (non-packed bootstrapping, Lmax=24) | llm | 0.90 |
| key result | area / TDP: 151.4 mm2 / 180.4 W (total F1 chip) | llm | 0.90 |
System card: F1¶
| Group | 8. FHE hardware accelerators |
| Family | ASIC (BGV/CKKS) |
| Arithmetic | - |
| Interaction | Hardware |
| Model scale | - |
| Status | Foundational |
| Relation to the SoK | First programmable FHE ASIC design. |
| Note | Projected thousands of times faster than CPU. Supports only shallow bootstrapping. |
| Systems page | F1 |
SoK reference [102]¶
N. Samardzic, A. Feldmann, A. Krastev, S. Devadas, R. Dreslinski, C. Peikert, and D. Sanchez, “F1: A fast and programmable accelerator for fully homomorphic encryption,” in MICRO-54: 54th Annual IEEE/ACM International Symposium on Microarchitecture, 2021, pp. 238–252.
Listed on the reference page of the site.
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: BGV 30 [p.2], CKKS 17 [p.2], GSW 5 [p.2], MPC 4 [p.13], BFV 2 [p.13]
- library: HElib 8 [p.7], EVA / CHET 2 [p.14]
- model: Logistic regression 8 [p.2], LeNet 1 [p.10], MobileNet 1 [p.10]
- dataset: MNIST 17 [p.10], CIFAR-10 3 [p.10]
- hardware: FPGA 8 [p.1], GPU 7 [p.10], ASIC 5 [p.1], NVIDIA A100 2 [p.10], Intel Xeon 1 [p.10]
- technique: NTT 71 [p.2], Bootstrapping 23 [p.3], Compiler 22 [p.1], RNS 16 [p.4], Key switching 12 [p.3], Modulus switching 8 [p.3], Rotations 3 [p.7], Bitslicing 3 [p.6]
- assumption: Lattice-based 4 [p.14]
- task: Genomics 2 [p.2]
- topic: Homomorphic encryption 134 [p.1], Benchmark 20 [p.4]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/catalogue/8-112_f1-a-fast-and-programmable-accelerator-for-fully-homomorphic.pdf |
catalogue | 15 | 1.3 MB | b637400c0a5e |
|
references/sok-references/102_f1-a-fast-and-programmable-accelerator-for-fully-homomorphic.pdf |
sok-references | 15 | 1.3 MB | b637400c0a5e |
byte-identical to references/catalogue/8-112_f1-a-fast-and-programmable-accelerator-for-fully-homomorphic.pdf |
Database key: f1-a-fast-and-programmable-accelerator-for-fully-homomorphic-encryption · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set f1-a-fast-and-programmable-accelerator-for-fully-homomorphic-encryption COLUMN VALUE