DEV Community

H33.ai
H33.ai

Posted on • Originally published at h33.ai

TFHE: Encrypted Decisions at 768 Transactions Per Second on ARM CPUs

BFV Adds and Multiplies. CKKS Infers. TFHE Decides.

Fully homomorphic encryption has three engines, each optimized for a different class of computation. TFHE handles the operations that BFV and CKKS cannot: comparisons, thresholds, equality checks, and Boolean decisions — all on encrypted data.

Measured Performance (Graviton4, 96 Channels)

Operation TPS Per-Channel
8-bit Greater-Than 768 125ms
16-bit Greater-Than 372 258ms
32-bit Greater-Than 182 526ms
64-bit Greater-Than 91 1,058ms
16-bit Equality 769 125ms

No GPU. No accelerator. Commodity ARM CPUs.

Use Cases

  • Fraud threshold: Is the encrypted transaction amount above the reporting limit? FraudShield
  • Credit decision: Is the encrypted credit score in the approval range? Banking
  • Access control: Does the encrypted access level meet the required tier? H33-Upstream
  • Session expiry: Has the encrypted timestamp passed the timeout? Identity

Automatic Routing via FHE-IQ

You do not choose between BFV, CKKS, and TFHE. The FHE-IQ router inspects the operation and selects the right engine automatically. Polynomial operations go to BFV. Real-number ML goes to CKKS. Comparisons go to TFHE. <500ns routing overhead.

GPU Acceleration

For higher throughput: 1,129 TPS on a single NVIDIA A10G. Multi-bit TFHE with production-grade noise margins.

Post-Quantum Secure

TFHE is lattice-based (LWE). Same mathematical hardness assumption as NIST ML-KEM. Every decision result attested via H33-74 with three PQ signature families.

TFHE Product Page · 96-Channel Benchmark · FHE Overview · Benchmarks · Schedule Demo

All NIST KATs passed. 20,000+ tests. Patent pending.

Top comments (0)