H33-TFHE evaluates encrypted comparisons, equality tests, and threshold decisions on commodity ARM CPUs. BFV adds and multiplies. TFHE decides. The FHE-IQ router picks the right engine automatically.
All numbers measured on AWS c8g.metal-48xl (Graviton4, 192 vCPUs, 96 channels). 30 seconds sustained per test. No GPU. Correctness verified before each run.
| Operation | Bit Width | AND Gates | TPS (96 ch) | Per-Channel | Hardware |
|---|---|---|---|---|---|
| Greater-Than | 8-bit | 15 | 768 | 125 ms | Graviton4, no GPU |
| Greater-Than | 16-bit | 31 | 372 | 258 ms | Graviton4, no GPU |
| Greater-Than | 32-bit | 63 | 182 | 526 ms | Graviton4, no GPU |
| Greater-Than | 64-bit | 127 | 91 | 1,058 ms | Graviton4, no GPU |
| Equality | 16-bit | 15 | 769 | 125 ms | Graviton4, no GPU |
| Raw AND gate | 1-bit | 1 | 11,526 | 8.3 ms | Graviton4, no GPU |
Linear scaling. Every width doubling doubles the gate count and halves throughput. 128-bit comparison: ~45 TPS. 256-bit: ~23 TPS. You can predict the cost of any width without running another benchmark.
BFV and CKKS can add and multiply encrypted data. They cannot compare, branch, sort, or match. Those operations are non-polynomial and require TFHE's boolean gate architecture.
Is the encrypted fraud score above the cutoff? Is the credit score in range? 8-bit comparison at 768 TPS per node.
Does the encrypted account identifier match the watchlist entry? 16-bit equality at 769 TPS. Cross-bank fraud detection without exposing data.
Is the encrypted transaction above the reporting threshold? 32-bit comparison covers amounts up to $21M in cents.
Has the encrypted session token expired? Full 64-bit Unix timestamp comparison at 91 TPS. Temporal access control without decryption.
| Width | TPS | Use Cases |
|---|---|---|
| 8-bit | 768 | Fraud score thresholds (0-255), credit risk bands, eligibility flags, compliance zone transitions, categorical decisions |
| 16-bit | 372 / 769 eq | Credit scores (300-850), identity attribute matching, KYC field comparison, insurance qualification, medical eligibility |
| 32-bit | 182 | Transaction amounts in cents (up to $21M), session-age validation, rate limiting, bounded scoring at full precision |
| 64-bit | 91 | Full Unix timestamps (ms precision), monetary micro-units, encrypted indexed lookups, high-frequency settlement |
You don't choose between BFV and TFHE. The FHE-IQ router inspects your operation and selects the right engine.
142 routing tests. 100 realistic scenarios across banking, healthcare, legal, cybersecurity, insurance, IoT, and governance. Every non-polynomial operation routes to TFHE. Every polynomial operation stays on BFV.
TFHE is not a general-purpose compute layer. It is a predictable, linearly scaling, encrypted decision engine. Encrypted comparisons double in cost per width doubling. Encrypted equality checks at small widths cluster at the same throughput as small-width comparisons because both compile to a constant number of programmable bootstraps. Hash functions and deep Boolean circuits are architecturally outside TFHE's economic envelope regardless of implementation.
What TFHE cannot do efficiently. Full SHA3-256 evaluation on encrypted data: 0.30 TPS. That is 640x below production viability. This is a fundamental limitation of TFHE's gate-evaluation model applied to high-gate-count circuits, not an implementation gap. We publish this number because the FHE industry has a credibility problem with unpublished limitations.
TFHE is lattice-based (Learning With Errors). Same mathematical hardness assumption that protects BFV. Every decision result is attested via the three-family post-quantum signature bundle and committed to the H33-74 substrate.