VeloPruf is a boutique GPU compute service for large particle simulations — gravity, electrostatics, fluids, swarms, and quantum-thermal jobs. Every large job ships with a physics validation certificate: we re-run a sample of your actual job against an exact reference and attach the verdict. If it can't be validated, we say so instead of stamping it.
Benchmarks below were run July 2026, warm-up excluded, best-of-N discipline, with pre-registered predictions. The validation methodology is part of the product.
| Workload | Scale | Measured |
|---|---|---|
| N-body (gravity + electrostatics + LJ), Barnes-Hut multi-level | 1,048,576 × 1,000 steps | ~24 s total |
| N-body, sustained | 52,428,800 particles | 1.69 s / step |
| Full job incl. validation certificate (verdict: PASS) | 200,000 × 500 steps | 3.7 s end-to-end |
| SPH-style fluid (cell-list vs naive) | 524,288 particles | ×246 speedup |
| Swarm / boids (cell-list vs naive) | 524,288 agents | ×272 speedup |
| Quantum-thermal (path-integral MD, PILE) | 65,536 atoms × 16 beads = 1,048,576 | 14.8 ms / step |
Head-to-head, same machine, same smart algorithm (measured July 15, 2026): the certified job above — 200,000 particles × 500 steps, certificate included — took 3.7 seconds on our GPU. In NumPy on that machine's own CPU: ~2.3 hours (measured step rates × step counts). Even granting an optimistic 8× for hand-tuned parallel native code: ~17 minutes. Logs available on request.
Approximate solvers can drift — silently. VeloPruf's engine takes a 20,480-particle sample of your job, runs it exactly, runs it exactly again with shuffled ordering (the chaos floor — the honest statistical bar), runs it fast, and compares. You get the verdict with the numbers, attached to your result.
{
"verdict": "PASS",
"sample_n": 20480,
"steps": 200,
"fast": { "ke_dpct": 0.00, "radial_tv_pct": 0.00, "mean_r_dpct": 0.00 },
"chaos_floor": { "ke_dpct": 0.00, "radial_tv_pct": 0.00, "mean_r_dpct": 0.00 }
}
An actual certificate from a 200,000-particle job (July 2026, RTX 4060 laptop). Quantum jobs get their own convergence certificate — a P-doubling ladder that has publicly rejected insufficient settings before approving valid ones. A certificate that can't say "no" is worth nothing; ours has said no, on the record.
Your input files are securely overwritten and deleted right after processing. Outputs retained 7 days, then purged. Your model setup never becomes our dataset.
During the pilot we run your job for you: email us the job file, we run it on the validated engine and email back results + certificate, typically within 24 hours. Self-serve API comes after the pilot.
Path-integral MD for nuclear quantum effects (light atoms, low temperatures), validated against exact references — with a convergence certificate per job.
This section is unusual for a landing page. It's here because the same discipline that produces the certificates produced this list.
Three steps, no software to install. (A self-serve API is planned after the pilot — we'd rather earn trust job-by-job first.)
Email us your job: a JSON file with your particles and settings, or just describe the system and we'll help you format it.
Your job runs on the validated engine (exact or multi-level Barnes-Hut, auto-selected). Typically back to you within 24 hours.
Results + the validation certificate, attached. Inputs are wiped after the run. NDA available on request.
// you send:
{ "particles": [ {"x":0.5,"y":0.5,"mass":1.0,"vx":0,"vy":0}, ... ],
"steps": 1000,
"physics": { "gravity":1.0, "coulomb":0.1, "lj_sigma":0.02 } }
// you get back:
{ "particles": [ ...final positions & velocities... ],
"compute_time_ms": 24051,
"certificate": { "verdict":"PASS", "fast":{"ke_dpct":0.00, ...} } }
During the two-week pilot, everything is free. Below is the planned pricing once the pilot ends. Raw GPU time is cheap; what costs you is days of setup, and never being sure the result is right. We price the validated answer. Each job covers up to 1B particle-steps (e.g. 1M particles × 1,000 steps) within the documented envelope; bigger jobs by quote.
The first 10 pilot users get certified jobs free for two weeks. In exchange: honest feedback about what you ran and whether the certificate mattered to you. If you run particle, fluid, swarm, or path-integral workloads — write us. Tell us what you simulate and at what scale.
N-body / gravityelectrostaticsLennard-Jones SPH fluidsswarms / boidspath-integral MD validation certificates