NVIDIA Corporation has introduced the Jetson Thor computing architecture, a hardware platform designed to power the next generation of humanoid robots capable of understanding complex human instructions and executing physical tasks. According to NVIDIA Chief Executive Officer Jensen Huang, speaking at the company’s GTC conference, the new system-on-a-chip brings advanced artificial intelligence capabilities directly to robotic hardware.
Architecture and Performance Capabilities
The Jetson Thor architecture features a transformer engine built into the Blackwell GPU architecture, delivering 2,000 teraflops of FP8 floating-point performance. According to NVIDIA product specifications, the processor enables humanoid robots to run multimodal generative AI models, such as large language models and vision-language models, in real time. The platform integrates a functional safety processor, a secure cluster engine, and a high-performance CPU complex to handle demanding robotic control algorithms.
Robotics developers use NVIDIA Isaac GR00T, a foundational model for humanoid robots announced alongside the hardware platform. According to company announcements, Isaac GR00T allows robots to take multimodal inputs—including text, speech, video, and human demonstration—and generate precise physical actions. This integration lets developers train robots in simulated environments using NVIDIA Omniverse before deploying trained models directly to the Jetson Thor hardware.
Industry Adoption and Deployment Timeline
Several robotics companies have announced plans to adopt the Jetson Thor platform for their upcoming commercial systems. According to NVIDIA, companies including Boston Dynamics, Figure AI, Sanctuary AI, 1X Technologies, and Unitree Robotics are developing systems utilizing the architecture. These manufacturers build general-purpose humanoid robots intended for manufacturing floors, logistics hubs, and potentially consumer environments.
The release of Jetson Thor addresses a longstanding bottleneck in embodied artificial intelligence: the physical space and power constraints of running large neural networks onboard a mobile robot. Traditional autonomous systems required tethered connections to remote data centers for complex reasoning. By packing data-center-grade inference capabilities into a compact robotic module, NVIDIA provides the onboard compute required for robots to operate independently in dynamic, unstructured human spaces.
Frequently Asked Questions
What is NVIDIA Jetson Thor?
According to NVIDIA, Jetson Thor is a computing architecture and system-on-a-chip designed specifically for humanoid robots, providing the massive AI compute required to run multimodal models onboard.
Which companies are using Jetson Thor?
According to NVIDIA announcements, robotics developers such as Boston Dynamics, Figure AI, Sanctuary AI, 1X Technologies, and Unitree Robotics plan to incorporate the hardware into their robotic platforms.
How does Jetson Thor differ from previous chips?
Unlike earlier edge-computing modules focused primarily on computer vision and basic navigation, Jetson Thor includes a transformer engine delivering 2,000 teraflops of FP8 performance to process generative AI models in real time.
Keep reading