Full Auth in 1.36ms:
220x Faster Than a Blink

A human blink takes 300,000 microseconds. H33 completes full post-quantum authentication in 1,360 microseconds. That means we can authenticate you 209 times before your eye finishes closing.

This isn't a theoretical benchmark. It's production performance, measured on standard cloud infrastructure (AWS c8g.metal-48xl with AWS Graviton4 (Neoverse V2)), processing real authentication requests.

Why This Matters

Zero-knowledge proofs have traditionally been considered "slow cryptography." Early ZK systems took seconds or even minutes to generate proofs. This made them impractical for real-time authentication.

H33 changes this equation entirely. At 1.36ms, full authentication is now faster than most database queries. It's faster than network round-trips. It's faster than you can perceive.

The Three-Stage Pipeline

Every H33 authentication executes three cryptographic stages in a single API call. Each stage is fully post-quantum secure, and the entire pipeline completes within a 32-user batch in approximately 1,356 microseconds — roughly 42 microseconds per individual authentication.

The dominant cost is the FHE biometric match in Stage 1. Biometric templates are encrypted under BFV (Brakerski/Fan-Vercauteren) with polynomial degree N=4096 and a single 56-bit modulus, then matched via an encrypted inner-product operation. Because BFV supports SIMD batching — 4,096 plaintext slots divided by 128 biometric dimensions — we pack 32 users into a single ciphertext and verify them all in one fused NTT-domain multiply-accumulate pass.

The Technical Achievement

Reaching 1.36ms required innovations across the entire stack:

• Montgomery NTT with Harvey lazy reduction: All Number Theoretic Transform butterflies operate in Montgomery form with values held in [0, 2q) between stages, eliminating modular division from the hot path entirely.
• NTT-domain fused inner product: Rather than performing an inverse NTT after every polynomial multiplication, H33 accumulates all products in the NTT domain and executes a single final INTT. For a 32-user batch, this removes 62 redundant transforms.
• Pre-NTT enrolled templates: Biometric templates are stored in NTT form at enrollment time. Each authentication skips the forward NTT on the reference vector, saving approximately 266 microseconds per batch.
• Batch Dilithium attestation: Instead of signing each authentication individually, H33 computes one SHA3-256 digest over the full 32-user batch and issues a single CRYSTALS-Dilithium signature plus verification — a 31x reduction in signing overhead compared to per-user attestation.
• Parallel NTT across moduli: All residue number system (RNS) channels are transformed concurrently via Rayon work-stealing, saturating all 96 cores during the encrypt phase.

// Rust: single API call — FHE + ZKP + Dilithium attestation
let result = h33::auth::verify_batch(&encrypted_templates, &probe, 32)?;

// result.fhe_us     = 1109   (BFV inner product, 32 users)
// result.zkp_us     = 0.085  (DashMap cache hit)
// result.attest_us  = 244    (SHA3 + Dilithium sign+verify)
// result.total_us   = 1356   (full pipeline)
// result.per_auth   = 42     (per-user amortized)

Real-World Impact

What does 1.36ms authentication mean in practice?

For users: Authentication feels instant. No loading spinners. No perceived delay. The security happens invisibly.

For developers: You can add ZK-based privacy to your authentication without performance concerns. The overhead is negligible — a single H33 auth call adds less latency than a typical DNS lookup.

For enterprises: At sustained throughput of 2,172,518 authentications per second on a single Graviton4 instance, H33 handles peak traffic events without horizontal scaling. Black Friday surges, concert ticket releases, election-day identity verification — one machine, one process, no queuing.

Benchmark Details

Our February 2026 production benchmarks were conducted on:

• Infrastructure: AWS c8g.metal-48xl (AWS Graviton4, Neoverse V2 cores)
• Configuration: 192 vCPUs, 377 GiB RAM, 96 Rayon workers
• FHE parameters: BFV with N=4096, single Q=56-bit modulus, t=65537
• Allocator: System (glibc malloc — verified faster than jemalloc on ARM64 for tight FHE loops)
• Measurement: End-to-end latency including all cryptographic operations, averaged over 10,000 batch cycles

// Single ZK proof generation
const result = await h33.auth.generateProof({
  userId: 'user_123',
  claims: ['identity', 'age_over_18'],
  mode: 'turbo'
});
// Completes in 1.36ms

// Batch processing (32 users per ciphertext)
const batchResult = await h33.auth.batchProof({
  requests: thirtyTwoUsers,
  mode: 'parallel'
});
// ~1,356µs total = ~42µs per auth

The Complete Performance Picture

ZK proofs are just one part of our performance story. Here's how the full authentication stack performs:

• Full Auth (Turbo mode): 1.36ms - Complete biometric + ZK + blockchain attestation
• Session Resume: 50µs - Returning users authenticate nearly instantly
• Incremental Auth: <50µs - When only 5% of context changes, only verify the delta
• Biometric Matching: 260µs - FHE-encrypted face/fingerprint verification

Try It Yourself

These aren't numbers you have to take on faith. Sign up for a free API key and run your own benchmarks. The performance is real, reproducible, and available today.