FHE Research Platform · v1.4.2 · Feb 2026

Homomorphic encryption that runs in production. Benchmark CKKS, BGV, and TFHE against your actual latency constraints — without ever exposing the plaintext.

View Benchmarks
CKKS Multiply
0.4ms
n=16384, L=20
BGV Key-Switch
0.2ms
128-bit security
TFHE Bootstrap
12ms
per gate
SIMD Slots
16Kslots
CKKS / BGV
Scroll to benchmarks
Live Benchmarks

Is it fast enough?

Every number below was recorded on an Intel Xeon W-3375 under 128-bit security parameters. No cherry-picking — these are the latencies your production system will see.

ct × ct latency

Ciphertext Multiplication

The most expensive FHE primitive — determines your circuit depth budget.

CKKSn=16384, L=20
0.4 ms
BGVn=16384, L=20
0.3 ms
TFHEn=630
8.2 ms

Measured with SEAL 4.1 / OpenFHE 1.1.2 · Xeon W-3375 · 32-core · 256 GB RAM

Level refresh

Bootstrapping

Level refresh cost. Determines how deep your computation can go before needing a reset.

CKKSSEAL 4.1, L-reset
3.2 s
BGVHElib, L-reset
2.1 s
TFHEper gate
12 s

CKKS/BGV: full level reset. TFHE: per-gate bootstrap. Amortized over 16K slots.

KS overhead

Key-Switching

Relinearization and rotation key costs. Dominates in SIMD-heavy workloads.

CKKSKS to fresh key
0.8 ms
BGVKS to fresh key
0.5 ms
TFHEKey extraction
14.5 ms

Includes key generation amortization. GaloisKeys precomputed for all rotation steps.

Production Use Cases

What can you build with it?

Three architectures shipping in production today — each one proves FHE has crossed the threshold from research curiosity to engineering reality.

CKKS · Neural Networks

Private ML Inference

Run a neural network over encrypted patient data. The model sees ciphertexts — never plaintext inputs or outputs. Deployed at two EU healthcare networks.

Enc. InputCKKS LayerReLU (approx)Enc. Output
ResNet-20 latency
4.2 s
Accuracy loss
< 0.3%
Scheme
CKKS
Depth
L = 25
BGV · SQL over ciphertext

Encrypted DB Queries

Execute range queries and aggregations over encrypted transaction records. The query processor never decrypts — compliance architects at three Tier-1 banks use this today.

Enc. QueryBGV FilterEnc. Result
Query latency
18 ms
Throughput
5.4K qps
Scheme
BGV
Plaintext
Exact
CKKS + BGV · Bioinformatics

Secure Genomic Analysis

Sequence alignment and variant scoring over encrypted genomes. NIH-funded pipeline — raw FASTQ files never leave the patient's encrypted domain.

Enc. GenomeCKKS AlignBGV ScoreEnc. Report
Alignment time
2.1 min
Variant recall
99.1%
Scheme
CKKS+BGV
SIMD slots
16 384
Scheme Selection Guide

Which scheme fits your constraints?

The wrong scheme choice costs you 10–100× in latency. Toggle each scheme to see where it wins — and where it doesn't.

Scheme Comparison

TFHE · CKKS · BGV

Toggle a scheme to highlight its performance profile across critical parameters.

ParameterTFHECKKSBGV
Supported Operations
Boolean gates, LUTApprox. arithmeticExact arithmetic
Noise Model
TLWE / TGSWBFV-style rescalingModulus switching
Bootstrapping Cost
~12 ms / bit~2–8 s (level reset)~1–5 s (level reset)
Multiplication Depth
Unlimited (slow)L = 20–40 levelsL = 20–40 levels
Plaintext Space
Single bit / small intComplex float vectorsInteger polynomial
Key-Switch Latency
8–15 ms0.3–2 ms0.2–1.5 ms
SIMD Slots
1N/2 (up to 16 384)N/2 (up to 16 384)
Error Precision
Exact (up to noise)~20–40 bitsExact (mod p)
Ideal For
Logic circuits, ML inferenceStatistics, ML, genomicsDB queries, voting
Cipher Toolkit Support
v1.4+v1.0+v1.2+

Data sourced from Cipher v1.4 internal benchmarks · 128-bit security · Intel Xeon W-3375 · Feb 2026

186K
Benchmark runs logged
3
Production deployments
12
Published preprints
99.1%
Genomic variant recall
Open Research License

The safe is already
solving from the inside.

Download the Cipher Toolkit and run your first FHE operation in under 10 minutes. CKKS, BGV, and TFHE — all under one API surface.

Read the Full Report →

Linux · macOS · Docker · Apache 2.0 for research use