H33
Last updated2026-06-15Methodologyrev 3Benchmarkv12.0Sourcesha256:e713e324…

Benchmark v12 · post-quantum authentication

Two million, two hundred and ninety-three thousand,
seven hundred and sixty-six.

Full post-quantum authentication lifecycle per second. FHE biometric match plus STARK plus ML-DSA-65 (Dilithium) signature. All measurements below are from Benchmark v12.

Auth / sec
2,293,766
Full PQ lifecycle · not just hashing.
Latency P50
38µs
±0.24% variance across 198.4 B ops/day.
Cost per auth
$3.8×10⁻¹⁰
Amortized · commodity cloud hardware.
Ops per day
198.4B
Full-mesh, sustained · not burst.

In 5 sentences

What is itBenchmark v12: 2,293,766 post-quantum authentications per second sustained · 38 microseconds latency · ±0.24% variance · fully reproducible.
Why it mattersPost-quantum authentication has historically been slow. Benchmark v12 shows it does not have to be, and publishes the methodology so it can be independently reproduced.
When to useWhen someone claims 'post-quantum is too slow for real workloads.' Or when comparing throughput of an authentication service.
How it differsFull-lifecycle measurement (enroll + verify + revoke, not verify-only). Sustained mean, not burst peak. Commodity hardware, not custom silicon.
EvidenceAWS c7g.16xlarge · rustc 1.79 with -C target-cpu=neoverse-v1 · tokio 1.38 · workload 5/90/5 · 15-run rolling median · Dockerfile + benchmark script published.
Simulated live counter · Benchmark v12 rate
2,293,766 · at 2,293,766/sec

If this page had been open for the whole benchmark day, the counter would read 198.4 billion.

Latency · µs · rolling 15-second window
P50 · 38.0µs · ±0.24%

Methodology

What Benchmark v12 actually measured.

Auth-per-second means the full round-trip · not a hash operation.

1 · Biometric match under FHE

Fully Homomorphic Encryption match against enrolled template. Template never decrypts. This is the expensive part.

2 · STARK proof of biometric

Zero-knowledge attestation that the biometric matched, without revealing the match itself. Verifiable by relying party.

3 · Session signing

Session token signed with ML-DSA-65 (Dilithium, NIST FIPS 204). Every session is quantum-safe.

4 · Audit emission

Signed authentication receipt emitted to Object-Locked audit log. Timestamped, replayable.

Test rig

Commodity cloud hardware · ARM64. Not GPU-accelerated. Not specialized silicon. Fully reproducible.

Variance

±0.24% across 198.4 billion authentications per day, sustained. Not burst-peak — sustained-mean.

Comparison

Where the number lands.

Full post-quantum authentication throughput · publicly benchmarked implementations.

H33 Benchmark v12 2,293,766 / sec
Classical WebAuthn (widely deployed)~100,000 / sec
PQ + classical (hybrid, roadmap)~10,000-50,000 / sec

External comparison numbers are ballpark from published benchmarks · methodology matters more than absolute figures. Read the H33 methodology above.

Provenance

How Benchmark v12 was actually measured.

Every decision that changes the number is documented below. The whole run is reproducible from the published Dockerfile and benchmark script.

Hardware

AWS c7g.16xlarge · Graviton3

  • · 64 vCPU · Neoverse V1 (ARM64)
  • · 128 GB DDR5-4800
  • · No GPU · no specialized silicon · no HSM
  • · 200 Gbit network · not saturated by workload

Explicit choice: prove the number holds on commodity hardware, not on custom silicon.

Compiler

rustc 1.79 · release profile

-C target-cpu=neoverse-v1
-C opt-level=3
-C lto=fat
-C codegen-units=1
-C panic=abort

SIMD instructions are compiler-emitted from the Rust source · no hand assembly.

Concurrency

tokio 1.38 · 512 workers

  • · 512-thread work-stealing runtime
  • · Bounded per-worker channels (backpressure at 10k)
  • · Zero unbounded queues · zero Arc<Mutex> hot paths
  • · Every allocation preflighted at startup

Workload mix

Realistic auth traffic

  • · 5% enroll · new biometric registration
  • · 90% verify · full PQ lifecycle
  • · 5% revoke · session invalidation

Not just the fast path · matches production traffic distribution.

Warmup

10 s JIT + cache warm

  • · 10 second warmup phase (discarded)
  • · 5 minute measurement phase
  • · 15-run rolling median at 60 s intervals
  • · Steady-state only · no burst-peak

Variance

±0.24% · 15-run median

  • · P50 · 2,293,766 / s
  • · P95 · 2,285,200 / s
  • · P99 · 2,281,540 / s
  • · CI 95% · [2,288,150 · 2,299,382]

Reproducibility

Containerized run

  • · Dockerfile · Rust 1.79-slim on debian:bookworm
  • · run-benchmark.sh · 220 lines
  • · Test vectors · deterministic seeds
  • · Publish path: benchmark-v12/ · ships with docs.h33.you

Anyone with a c7g.16xlarge can reproduce within variance.

Date & version

v12.0 · methodology rev 3

  • · Run date · 2026-06-15
  • · Rev 3 changes: added revoke path to workload mix
  • · Rev 2 (2026-05-01): STARK proof included in lifecycle
  • · Rev 1 (2026-03-12): initial v12 · 2.19M base
Excerpt · run-benchmark.sh
# Set up · deterministic seeds, isolated cgroup, disabled TurboBoost swing
export H33_BENCH_SEED=0xdeadbeef
export RUSTFLAGS="-C target-cpu=neoverse-v1 -C lto=fat -C codegen-units=1"
cargo build --release --bin bench-pq-auth
cgexec -g cpu:bench ./target/release/bench-pq-auth \
  --workers 512 --warmup 10s --duration 300s \
  --mix enroll=5,verify=90,revoke=5 \
  --emit-percentiles --emit-median --emit-ci
# Repeat 15× at 60s intervals · take rolling median

Trust status

Numbers verified. Everything else on the page traces to code.

Benchmark v12 was run on production hardware · methodology below.

Runs today

  • Vault workflows running in production
  • Cryptographic receipts emitted per action
  • Immutable audit trail · Object-Locked
  • Auth1 · Wallet · Document · Audit APIs
  • Post-quantum primitives (Dilithium + Kyber)
  • Benchmark v12 · 2,293,766 auth/sec

Coming next

  • Public /verify// receipt viewer
  • Public /replay// decision viewer
  • docs.h33.you developer documentation
  • SOC 2 Type II Certified
  • Payment Requests + Rewards adapters

The benchmark is a claim you can reproduce.

Every methodology decision is documented above. Every result is replayable from the receipt.

Start free See the proof surface