The NVIDIA
AV Stack

Deconstructing the Autonomy Blueprint

NVIDIA’s offering is broken down into a few distinct, interconnected layers, allowing automakers to deploy Level 4 autonomy years faster than if they had built the tech themselves.

arrow_downward

memory

The Hardware

DRIVE Hyperion & Thor

This is the physical nervous system and brain. Hyperion is the standardized blueprint for where to place sensors (14 cameras, 9 radars, 1 lidar, and 12 ultrasonics).

bolt DRIVE Thor

The incredibly powerful in-vehicle AI supercomputer chip that processes all of that sensor data in real-time.

radar

Hyperion Nervous System

Cameras

14 exterior high-definition cameras (creating a 360-degree overlapping field) plus 3-4 interior cameras for driver and occupant monitoring.

Radar & LiDAR

9 exterior surround radars alongside 1-3 high-resolution LiDAR nodes.

Ultrasonics

20 sensors for ultra-precise, short-range curbing & debris detection.

Audio Array

An exterior microphone array so the vehicle can "hear" approaching emergency sirens or honking before it can even see them.

speed

Thor Deep Dive (1/2)

Architecture

Built on the NVIDIA Blackwell GPU architecture paired with a 14-core Arm Neoverse V3AE CPU.

Compute Power

Delivers 2,000 FP4 TFLOPS or 1,000 INT8 TOPS per chip. Roughly 8x more powerful than the previous generation.

Memory

128 GB LPDDR5X system RAM with a blistering 273 GB/s memory bandwidth.

speed

Thor Deep Dive (2/2)

Transformer Engine

Thor is the first AV platform to incorporate a dedicated transformer inference engine. This is critical because it allows the car to run massive Vision-Language-Action (VLA) models entirely "at the edge" inside the car without needing a constant internet connection to the cloud.

NVLink-C2C

If an automaker wants even more power (e.g., for a heavy-duty Level 5 autonomous semi-truck), they can use this seamless interconnect technology to chain two Thor chips together so they act as one monolithic super-brain.

neurology

The "Brain"

Alpamayo 1.5

This is NVIDIA's open Vision-Language-Action (VLA) AI model. Instead of rigid, manually coded "if/then" rules for driving, Alpamayo takes in video feeds and actually reasons through complex scenarios.

"It can even take natural language instructions to adjust its driving behavior on the fly."

account_tree

Alpamayo Deep Dive

Glass Box Reasoning

Generating human-readable logic traces alongside commands, ending the "black box" era. This provides clear, auditable decision trails for insurance and safety regulators.

Teacher-Student Distillation

A massive 10B parameter model trains in the cloud, then "distills" its knowledge into lightning-fast 2B parameter student models that run locally on the car.

AlpaSim Closed-Loop Reality

Dropped directly into an interactive physics engine/city digital twin. Alpamayo practices surviving rare anomalies (like dropped ladders on highways) without physical risk.

shield

The Safety Net

Halos OS

This is the underlying operating system that acts as an unshakeable guardrail.

System Override Protocol

It runs continuous safety checks from the cloud to the car to ensure the AI's decisions never violate core safety standards (like crashing or running a red light).

view_in_ar

Digital Twins

Omniverse & Cosmos

Before a car ever hits real asphalt, it is trained in the matrix. NVIDIA uses its Omniverse simulation platform to create physics-accurate "digital twins" of the real world.

"Automakers can generate synthetic data to train the AI on rare edge cases (e.g., a pedestrian dropping a mattress on a highway in a blizzard) without having to wait for it to happen in real life."

directions_bus

Real-World Execution

Case Study: Isuzu Autonomous Bus

hardware

1. Bolting in the Hardware

Isuzu takes the Hyperion blueprint, installs the 14 cameras, lidars, and radars onto the bus frame, and drops the Thor supercomputer under the hood.

public

2. Virtual Miles

They boot up Omniverse, create Tokyo's streets, and run millions of synthetically generated weather / traffic anomalies to teach the bus.

download

3. Loading the Brain

Alpamayo is loaded onto the bus. Live camera feeds allow the bus to recognize complex scenes (like swerving cyclists) and avoid them.

safety_check

4. Deploying Safely

Halos OS constantly monitors calculations, overriding any unsafe decisions with forced stops.

The Conclusion

By leveraging this open framework, Isuzu deploys a fully autonomous Level 4 bus years faster and vastly cheaper than inventing the wheel from scratch.

Return to Research

The NVIDIA AV Stack