NVIDIA Robotaxi Fleet Compute & Simulation Planner

Deployment Scenarios One-Click Architecture

Active Robotaxis 1,500

50 AVs 5,000 10,000 AVs

Perception Sensor Suite 4 LiDAR + 8 Cams

TensorRT Perception Precision FP8 Transformer

Omniverse Cloud Sim Ratio 50x real-time

Cloud synthetic simulation miles rendered per commercial road hour.

Cloud Training Supernodes 256 DGX Nodes

In-Vehicle Compute

3,000,000 TOPS

DRIVE Thor AGX Platform

End-to-End Latency

12.4 ms

Sensor Ingest → Actuation

Omniverse Sim Load

75,000 hrs/day

Synthetic Validation Stream

Total Power Draw

1.44 MW

Vehicle DC + Cloud Train

Fleet Capital Est.

$43.5M

Compute + DGX Infrastructure

Interactive Sensor Fusion & TensorRT Perception Pipeline

60 FPS

Hardware Subsystem Breakdown Real-Time Allocation

Subsystem Layer	Hardware Spec	Per-Unit Metric	Fleet Aggregate	Utilization
In-Vehicle Perception Core	NVIDIA DRIVE Thor SoC (2,000 TOPS)	2,000 TOPS / AV	3.00 PFLOPS	62% Headroom
Sensor Fusion & Planning	4x 128-Beam LiDAR + 8x 8MP HDR Cams	1.84 GB/s Ingest	2.76 TB/s Fleet IO	Deterministic PCIe Gen5
Omniverse Digital Twin Sim	Cloud GPU Nodes (Ray-Tracing & Physics)	50 hrs sim / vehicle-day	75,000 hrs/day	99.4% Sim Accuracy
Model Re-training Cluster	DGX H100 SuperPODs (Cloud)	256 Nodes	0.24 MW Cluster	Active Continuous Training

Autonomous Scale Architecture Note: As noted by NVIDIA, commercial robotaxi scaling requires closed-loop synthesis. In-vehicle DRIVE platforms operate safety-critical transformer models in real-time, while Omniverse simulates millions of rare edge cases in cloud digital twins to validate safety before deployment.