Simvolt · Stage 03

Thousands of rollouts while you sleep

Simvolt is the compute engine. It runs your candidate policy across thousands of randomised environments at once on data-centre GPUs, so a full regression finishes in hours.

SweepExample

sweep #4812 · running

Runs
12,480
Completed
9,306
GPUs
32
Elapsed
1 h 12 m
GPU-hours so far
38.4

Capabilities

What Simvolt does

Scale

Massively parallel

Thousands of environments per sweep with Isaac Sim and Isaac Lab.

Inference

Your policy, unchanged

Runs ONNX, TorchScript or TensorRT builds exactly as they ship.

Determinism

Seeded runs

Every run has a seed, so any failure replays exactly.

Scheduling

Priority queues

Release gates jump the queue; research sweeps fill idle capacity.

Elastic

Burst capacity

Scale out for big sweeps and back down when done.

Cost

GPU-hour metering

See the GPU-hours each sweep used, per suite and per team.

How it works

Inside Simvolt

  1. Package

    Load the policy build and the twin version.

  2. Fan out

    Spread scenarios and variations across GPUs.

  3. Run

    Step physics, sensors and policy together.

  4. Collect

    Stream outcomes and traces to Veridex.

Specification

Inputs, outputs and hardware

ItemDetail
SimulatorNVIDIA Isaac Sim / Isaac Lab on Omniverse
Policy formatsONNX, TorchScript, TensorRT engines, ROS 2 nodes in a container
HardwareDGX / HGX (H100, B200 class) in Simprova Cloud or yours
Typical sweep10,000–50,000 runs in 1–4 hours
OutputsPer-run outcomes, metrics, traces and replay files

Questions

About Simvolt

Can we bring our own GPUs?

Yes. On Hybrid and On-prem, Simvolt schedules onto your DGX or cloud GPU quota.

Does it support ROS 2 stacks?

Yes. Containerised ROS 2 nodes run in the loop with simulated sensors and clocks.

How do you estimate GPU-hours?

The pricing page has an estimator. Actual use depends on scene complexity and run length.

See Simvolt on your site

We'll run it on your own floor plan and one recent policy build.