Compute · Fleet
Live telemetryThe Aurora simulation fleet.
675 GPUs — NVIDIA H100, H200, and B200 — spread across our offices in the United States, Europe, China, and Singapore. They simulate physics, render sensors, and pack ground truth in parallel. Figures below refresh live from the fleet API.
GPU pools
Capacity organized by GPU model.
Jobs are placed onto the GPU model that fits their workload — from H100 for standard rendering to B200 for the heaviest path-traced sensor synthesis.
NVIDIA H100
pool_h100
80 GB
320
300
NVIDIA H200
pool_h200
141 GB
250
238
NVIDIA B200
pool_b200
192 GB
105
97
The fleet is distributed across Aurora offices. The scheduler places each job by available capacity and locality.
- United States255 GPUs · 94%H100 · 120H200 · 90B200 · 45
- Europe190 GPUs · 69%H100 · 90H200 · 70B200 · 30
- China140 GPUs · 86%H100 · 70H200 · 50B200 · 20
- Singapore90 GPUs · 84%H100 · 40H200 · 40B200 · 10
- frg_92c4e1/ warehouse_07PACKING94%96 GPUs
- frg_5a1f7d/ factory_line_03RENDERING62%121 GPUs
- frg_c803b9/ amr_yard_11PACKING91%121 GPUs
- hc_2f7a04/ occlusion_ped_varSIMULATING44%116 GPUs
How the fleet scales
A generation job is split into independent shards — ranges of the scenario distribution that can be simulated with no coordination.
Shards are placed onto the right GPU model and office by the scheduler, following available capacity and locality.
Frames and ground truth stream off the workers and merge into a single sharded, versioned dataset — the same job on ten workers or ten thousand.
Read more about the compute model in Technology → Compute.
Scale your data
Put the fleet to work on your worlds.
Tell us about the perception models you are training and the environments they need to see. We will scope a synthetic-data plan with you.