H33 Full Pipeline Benchmark v10.0 — FHE + Real STARK Proofs + Dilithium + ML

01

Pipeline Breakdown

Each authentication passes through four stages in a single API call. All operations are post-quantum secure. Biometric data never leaves FHE encryption.

Component	Latency	% of Pipeline	PQ-Secure	Notes
FHE Batch BFV	939 µs	76.2%	Yes (lattice)	Montgomery NTT, Harvey lazy reduction, 32 users/CT
Dilithium Attestation ML-DSA	291 µs	23.6%	Yes (ML-DSA)	1 sign+verify per 32-user batch
STARK ZKP Cached	0.059 µs	<0.01%	Yes (SHA3-256)	DashMap in-process, 7-column AIR
ML Threat Agents 3 agents	~2.35 µs	0.19%	N/A	Harvest + side-channel + crypto health
Total Pipeline	1,232 µs	100%	Full PQ	38.5 µs per auth (32 users/batch)

Pipeline composition: FHE inner product on encrypted biometric templates (BFV, N=4096) → STARK proof lookup (cached DashMap) → Dilithium batch attestation (1 signature per 32 users) → ML threat classification. All stages are fused in a single Rayon task per batch.

02

STARK Proof Performance

Real STARK zero-knowledge proofs with a 7-column algebraic intermediate representation (AIR). Cold proofs are generated once per enrollment; production auth hits the DashMap cache at sub-microsecond latency.

Metric	Value
Generate (cold)	68.093052 ms
Verify (cold)	14.366931 ms
Cache cold	14.400565 ms
Cache hot (DashMap)	1.159 µs
Production lookup	0.059 µs

AIR Configuration	Detail
Columns	7
Column names	enrolled, fresh, dot_acc, norm_a, norm_b, poseidon, step
Transition constraints	5 per row
Public inputs	7
Hash function	SHA3-256

Cache strategy: STARK proofs are generated at enrollment time (68ms cold, amortized over account lifetime). During auth, proofs are retrieved from an in-process DashMap at 0.059µs. This eliminates the TCP serialization bottleneck that caused an 11x regression when using external RESP cache at 96 workers.

03

Throughput Results

Sustained throughput measured over a 120-second window with per-second sampling. Peak and low values represent single-second snapshots within the window.

Window	Auth/Sec	Batches/Sec	Batch Latency	Variance
Peak (1s)	2,190,496	~1,778	1,228 µs	—
Sustained (120s)	2,172,518	~1,763	1,232 µs	±0.71%
Low (1s)	2,159,776	~1,753	1,236 µs	—
Spread	30,720	~25	8 µs	—

Throughput Notes

96 Rayon workers across 192 vCPUs (2 HW threads/core)
System allocator — jemalloc causes 8% regression on Graviton4
NUMA-aware thread pinning for cache locality
Variance collapsed from ±6% (v9) to ±0.71% (v10) via thermal management

Scaling Context

2.17M auth/sec = 187.5 billion auth/day on a single node
Each auth includes full FHE + ZKP + PQ signature + ML
Cost: ~$1.80–$2.30/hr (spot pricing)
Per-auth cost: < $0.000001 at sustained throughput

04

ML / AI Agent Performance

Three native Rust AI agents run inline on every authentication for real-time threat intelligence. Zero external dependencies — no Python, no ONNX, no GPU. Pure Rust compiled to ARM.

Agent	Function	Latency	Method	Source
Harvest Detection Threat	Detects harvest-now-decrypt-later attack patterns	0.69 µs	Bayesian classifier	ai_harvest.rs
Side-Channel Analysis Monitor	Detects timing and power analysis attack vectors	1.14 µs	Statistical anomaly detection	ai_sidechannel.rs
Crypto Health Monitor Health	Runtime health scoring of FHE/ZK/PQ parameters	0.52 µs	Parameter drift detection	ai_crypto_health.rs
Total ML Overhead	All 3 agents combined	~2.35 µs	Included in pipeline total (0.19% of batch)

Harvest Detection

Monitors for store-now-decrypt-later (SNDL) patterns
Bayesian prior updated from traffic distribution
Flags suspicious bulk-capture behavior in real time
Outputs threat score [0.0, 1.0] per batch

Side-Channel Analysis

Statistical detection of timing correlations
Monitors variance in FHE operation durations
Flags potential power/EM side-channel probes
Constant-time verification of critical paths

Crypto Health Monitor

Validates FHE noise budget hasn't drifted
Checks ZKP constraint satisfaction rates
Monitors Dilithium key freshness and rotation schedule
Composite health score per authentication

05

FHE Engine Comparison

H33 ships four FHE engine configurations, each optimized for different security levels and use cases. All engines share the same Montgomery NTT core with Harvey lazy reduction.

Engine	Scheme	Parameters	Security	Use Case	Cycle Time
H33-128 Production	BFV-64	N=4096, Q=56-bit, t=65537	NIST L1	Biometric auth, high throughput	~1.36 ms
H33-256	BFV-64	N=8192, multi-Q RNS	NIST L5	High-security, government	~5.98 ms
H33-CKKS	CKKS	N=4096, approximate	NIST L1	ML inference, float ops	~2.1 ms
H33-BFV32	BFV-32	N=4096, 32-bit modulus	NIST L1	ARM mobile / edge devices	~0.7 ms (ARM)

SIMD batching (all engines): 4096 polynomial slots ÷ 128 biometric dimensions = 32 users per ciphertext. CRT condition: t ≡ 1 (mod 2N) → t=65537, N=4096 (ψ=6561). Template storage reduction: ~32MB/user → ~256KB/user (128x).

06

Microsoft SEAL Comparison

Performance comparison against Microsoft SEAL, the most widely adopted open-source FHE library. SEAL provides FHE only; H33 includes the full post-quantum pipeline (FHE + STARK ZKP + Dilithium signatures + ML agents).

Metric	H33	Microsoft SEAL	Ratio
Single-thread FHE cycle	1.36 ms	2.85 ms	2.3x faster
Sustained throughput (96 cores)	2,172,518/sec	~92,000*/sec	23.6x
Pipeline scope	FHE + STARK + Dilithium + ML	FHE only	—
Post-quantum ZKP	STARK (SHA3)	None	—
Post-quantum signatures	Dilithium (ML-DSA)	None	—
Threat intelligence	3 ML agents	None	—
SIMD batch auth	32 users/CT	None	—

Caveat: *SEAL throughput is extrapolated from single-thread benchmarks; Microsoft does not publish a production batch-authentication benchmark. The 23.6x ratio reflects H33's full-stack optimization including Montgomery NTT, Harvey lazy reduction, NEON Galois ops, NTT-domain persistence, and in-process proof caching — none of which apply to SEAL's general-purpose design.

07

Product Suite Overview

18 products spanning identity, encryption, key management, data protection, and AI. All products share the same post-quantum cryptographic core and are accessible via a unified credit-based API.

Product	Category	Primary Crypto	Unit Cost	Description
H33-Vault	Storage	AES-256-GCM + Kyber	1 credit	Encrypted vault for secrets, keys, and sensitive data
H33-Share	MPC	Shamir + Dilithium	3 credits	Threshold secret sharing with PQ attestation
H33-Shield	Identity	BFV + STARK + Dilithium	5 credits	Full biometric auth pipeline (single API call)
H33-Key	KMS	Kyber + X25519	1 credit	Hybrid PQ key exchange and management
H33-Gateway	Network	Kyber TLS + Dilithium	2 credits	PQ-secure API gateway with mutual auth
H33-Health	Identity	BFV + CKKS	5 credits	HIPAA-grade encrypted health data processing
H33-128	FHE	BFV N=4096	1 credit	NIST L1 FHE compute (production engine)
H33-256	FHE	BFV N=8192	3 credits	NIST L5 FHE compute (high-security)
H33-CKKS	FHE	CKKS N=4096	2 credits	Approximate FHE for ML inference on encrypted data
H33-BFV32	FHE	BFV 32-bit	1 credit	Lightweight FHE for ARM mobile and edge
H33-MPC	MPC	Garbled circuits + OT	5 credits	Multi-party computation for joint analytics
H33-3-Key	KMS	Triple-wrap Kyber	3 credits	3-layer key wrapping for sovereign data
Biometrics	Identity	BFV + STARK	5 credits	Encrypted biometric enrollment and matching
Encrypted Search	FHE	BFV + PIR	3 credits	Private information retrieval on encrypted indexes
PQ Video	Network	Kyber + AES-256-GCM	2 credits/min	Post-quantum encrypted video streaming
Storage Encryption	Storage	Kyber + AES-256-GCM	1 credit/GB	At-rest PQ encryption for cloud storage
AI Detection	AI	Rust ML agents	1 credit	Harvest/side-channel/health threat detection
FHE-IQ	AI	CKKS + ML	5 credits	ML inference on fully encrypted data

08

Integration & API

Unified REST API with SDK support for four languages. Credit-based billing with per-call metering.

Core Endpoints

# Biometric auth (single API call)
POST /api/v1/auth/verify

# Enrollment
POST /api/v1/auth/enroll

# FHE compute
POST /api/v1/fhe/encrypt
POST /api/v1/fhe/compute
POST /api/v1/fhe/decrypt

# Key management
POST /api/v1/keys/generate
POST /api/v1/keys/exchange

# Vault
POST /api/v1/vault/store
GET  /api/v1/vault/retrieve

# Health & billing
GET  /api/v1/health
GET  /api/v1/billing/balance
        

Authentication

# API key in header
Authorization: Bearer h33_pk_...

# Example: verify biometric
curl -X POST https://api.h33.ai/api/v1/auth/verify \
  -H "Authorization: Bearer h33_pk_..." \
  -H "Content-Type: application/json" \
  -d '{"template": "base64...", "user_id": "..."}'
        

SDK Languages

Rust (native) Python JavaScript / TypeScript Go

Rate Limits by Tier

Tier	Rate Limit
Free	5 req/sec
Starter	100 req/sec
Pro	1,000 req/sec
Enterprise	Unlimited

Webhook support: Async operations (large FHE computations, batch enrollments) deliver results via configurable webhooks. Configure at /api/v1/webhooks.

09

Test Suite

Comprehensive test coverage across all cryptographic primitives, with deterministic unit tests, randomized property testing, and integration tests for the full pipeline.

2,227

Deterministic tests passing

300,000+

Proptest random inputs

5

Test categories

100%

Critical path coverage

Category	Count	Scope	Notes
Unit	1,247	Individual function correctness	NTT, BFV ops, Galois, modular arithmetic
Integration	389	Cross-module pipelines	FHE → ZKP → Dilithium → ML full flow
Fuzz (proptest)	300,000+	Random input generation	Polynomial arithmetic, encrypt/decrypt round-trip
Benchmark	142	Performance regression detection	Criterion + custom Graviton4 harness
Cross-module	150	Module boundary correctness	Domain form transitions, CRT combine, batch verify

10

Version History

Throughput progression from v7.0 baseline to v10.0 with real STARK proofs and variance collapse.

Version	Date	Key Change	Sustained Auth/Sec	Improvement
v7.0	Feb 14, 2026	Baseline Montgomery NTT + Harvey lazy reduction	1,291,207	—
v8.0	Feb 26, 2026	NTT-form multiply_plain, skip INTT per call	1,595,071	+23.5%
v9.0	Mar 5, 2026	In-process DashMap cache, batch attestation	1,714,496	+7.5%
v10.0	Mar 9, 2026	Real STARK proofs + variance collapse (±0.71%)	2,172,518	+26.7%

Cumulative improvement: v7.0 → v10.0 represents a 68.3% throughput increase (1,291,207 → 2,172,518 auth/sec) while adding real STARK proofs, 3 ML agents, and collapsing variance from ±6% to ±0.71%. Per-auth latency dropped from ~55µs to 38.5µs.

1.29M

v7.0

1.60M

v8.0

1.71M

v9.0

2.17M

v10.0

Sustained throughput progression

Graviton4 c8g.metal-48xl, 96 workers, 120s window

+68.3% cumulative improvement

11

Methodology & Reproduction

All benchmarks are reproducible on identical hardware. Thermal management is critical for sustained results on bare-metal Graviton4.

Configuration

Workers: 96 Rayon threads
Allocator: System (glibc), NOT jemalloc
FHE mode: biometric_fast() — N=4096, Q=56-bit, t=65537
Batch size: 32 users per ciphertext
ZKP cache: In-process DashMap (CACHEE_MODE=inprocess)
NTT: Montgomery radix-4 with Harvey lazy reduction
NEON: Galois rotation + key-switching

Measurement

Sustained window: 120 seconds
Sampling: Per-second throughput counters
Warm-up: 10 seconds excluded from results
Thermal: CPU freq pinning, NUMA-aware thread placement
Variance: Reported as (max-min)/mean over 120s
Instance: c8g.metal-48xl (192 vCPU, 377 GiB)

Reproduction

# 1. Deploy to Graviton4
./scripts/graviton4_deploy.sh

# 2. Run benchmark (120s sustained, 96 workers)
CACHEE_MODE=inprocess RAYON_NUM_THREADS=96 \
  cargo run --release --example graviton4_bench

# 3. Expected output
# Sustained (120s): 2,172,518 auth/sec ±0.71%
# Batch latency: 1,232µs (32 users)
# Per-auth: 38.5µs
    

Known anti-patterns (do not retry): jemalloc on Graviton4 (8% regression), fused NTT twiddles (L1 cache pollution at 96 workers), TCP RESP proxy at high concurrency (11x regression), NEON 64-bit butterfly (no native vmull_u64), arena pooling for Vec<Vec<u64>> (zeroing overhead).