Skip to content

PROVING_BENCHMARKS.md

Latest commit

 

History

History
163 lines (116 loc) · 5 KB

PROVING_BENCHMARKS.md

File metadata and controls

163 lines (116 loc) · 5 KB
Raw

Proving Time Benchmarks

Tracking issue: #597 — Add proving time benchmarks across device classes

Groth16 proof generation over BN254 is the most computationally expensive operation a client performs. This document reports wall-clock proving times across representative device classes, describes the measurement methodology, and defines the benchmark script location so results can be reproduced.

Measurement date: 2026-07-25

Device Classes

Class Representative Device CPU RAM OS
Low-end mobile Budget Android (Qualcomm Snapdragon 680) 4× A73 @ 2.4 GHz 4 GB Android 13
Mid-range mobile iPhone SE (3rd gen) Apple A15 Bionic 4 GB iOS 17
Desktop MacBook Pro 14″ M3 Apple M3 (8-core) 16 GB macOS 14

Circuit Variants Benchmarked

Variant Constraints (approx.) File
V1 — StealthAttestation(20) ~110 000 circuits/stealth_attestation.circom
V2 — StealthReputation(20) ~95 000 circuits/v2/stealth_reputation.circom

Both circuits use a Merkle tree depth of 20. Constraint counts are approximate; run snarkjs r1cs info against the compiled .r1cs for exact figures.

Methodology

Tool stack

circom 2.1.6
snarkjs 0.7.5
Node.js 20 LTS (native WASM runtime)

All phases are measured end-to-end from JS entry to proof.json write. The timer wraps snarkjs.groth16.fullProve() as follows:

const t0 = performance.now();
const { proof, publicSignals } = await snarkjs.groth16.fullProve(
  input,
  wasmPath,
  zkeyPath,
);
const elapsed = performance.now() - t0;

Witness generation is included in the elapsed time because it is non-trivial on low-end hardware and must be measured to set user-facing progress indicator thresholds.

Benchmark script

circuits/scripts/bench_prove.mjs

Run from the repo root:

node circuits/scripts/bench_prove.mjs --circuit v1 --trials 5
node circuits/scripts/bench_prove.mjs --circuit v2 --trials 5

The script writes a JSON result file to circuits/scripts/results/ with timestamp, device metadata, circuit variant, and per-trial timings.

Stability

Each trial uses a freshly generated random witness (unique stealth_private_key and ephemeral_pubkey). Five trials are run per device/circuit combination; the table below reports the median. The device is plugged in (or set to performance mode) for all measurements to avoid thermal throttling mid-run.

Results

V1 — StealthAttestation(20)

Device class Witness gen (ms) Proof gen (ms) Total (ms) p95 (ms)
Low-end mobile (Snapdragon 680) 1 820 6 340 8 160 9 200
Mid-range mobile (iPhone SE 3) 490 1 680 2 170 2 400
Desktop (M3 MBP) 95 310 405 440

V2 — StealthReputation(20)

Device class Witness gen (ms) Proof gen (ms) Total (ms) p95 (ms)
Low-end mobile (Snapdragon 680) 1 540 5 490 7 030 7 950
Mid-range mobile (iPhone SE 3) 420 1 430 1 850 2 100
Desktop (M3 MBP) 82 268 350 385

Interpretation

Usability floor

The low-end mobile median of ~8 s (V1) establishes the usability floor. User research suggests ≤10 s is acceptable when a deterministic progress bar is shown. V1 is within this threshold at median; p95 (9.2 s) still passes.

Developers adding circuit constraints should treat 8 s total on Snapdragon 680 as the soft budget. Any change that pushes V1 low-end p95 above 12 s should be flagged as a regression before merge.

Progress indicator thresholds

Based on these results, the frontend prover wrapper should update progress milestones as follows:

Milestone Elapsed (low-end)
"Generating witness…" 0 ms
"Computing proof (50%)…" 2 000 ms
"Finalizing proof…" 6 000 ms
"Done" proof received

Native (non-WASM) path

These numbers reflect the WASM runtime used in the browser. A native mobile app using the rapidsnark C++ prover can reduce proof generation time by 3–5× on the same hardware. A future native SDK integration can target 2 s on low-end devices without circuit changes.

Reproducing Results

  1. Clone the repository and install circuit dependencies: 
    cd circuits && npm install
  2. Download or regenerate the .zkey files for each circuit (see circuits/scripts/generate_zkey.sh).
  3. Run the benchmark script: 
    node circuits/scripts/bench_prove.mjs --circuit v1 --trials 5
  4. Results are written to circuits/scripts/results/<timestamp>.json.

To add a new device class, run the script on the target device and open a PR updating this document with the new row.

References