Nvidia Expands Thor Modules for Scalable Robotics and Edge AI

Addressing the challenge to deliver compact, power-efficient AI supercomputers optimized for scalable robotics deployment, Nvidia has expanded its Jetson Thor family with the introduction of the T3000 and T2000 modules. These Blackwell-powered modules target mainstream robotics and edge AI applications at scale.

The Jetson AGX Thor family of robotics supercomputers has been adopted by leading robotics companies for robot, factory, and industrial equipment applications. With the new Jetson Thor modules, Nvidia now offers a scalable edge AI platform that spans from 70 TOPS to 2,000 TFLOPS.

System-on-modules (SOMs) for embedded robotics development are designed to simplify hardware design with integrated devices such as compute, memory, and power management. In addition, these SOMs are certified, offer development kits for fast evaluation and customization, and often come with a large ecosystem of partners.

The T3000 module takes aim at enabling intelligent humanoid robots and autonomous systems at scale. Key features include a Blackwell GPU with 865 FP4 TFLOPS, an eight-core Neoverse Arm CPU, 32-GB LPDDR5X memory, 273-GB/s memory bandwidth, and high-speed I/O (25 GbE connectivity). The IGX T3000 module delivers the same performance with integrated functional safety, running the Nvidia Halos for Robotics full-stack safety system for robots operating alongside humans.

The T3000 also is about half the size and power of the T5000, Nvidia said. It offers similar inference performance of the T5000, including large language models, vision language action models, vision language models, and world foundation models, with the same high memory bandwidth and optimized tokens per watt. Nvidia positions the T3000 as an energy efficient and cost-effective solution when migrating from the T5000, particularly with the rising cost of memory.

The T2000 opens up a broader range of edge AI applications to the Thor architecture. Described as an entry point to the Thor product family, it enables applications such as visual AI agents, autonomous mobile robots, and manipulators. Key features include 400 FP4 TFLOPS of AI compute and 16 GB of memory.

Also available is the Jetson AGX Thor Developer Kit through channel partners, which will offer emulation and migration services for the T3000 and T2000 modules. It runs the complete Nvidia software stack, supporting open models such as Nemotron, Cosmos 3, and Isaac GR00T, and new Jetson agent skills, which enable developers to optimize the entire software stack and to achieve memory savings in days instead of weeks, according to Nvidia.

Cosmos 3, an open omnimodal world model for physical AI, is now available to edge devices, as a lightweight model compatible with the Thor platforms. It enables real-time livestream and multi-view vision understanding for smart infrastructure, robots, and autonomous systems. Nvidia said the Cosmos 3 Edge 4-billion parameter model is a "highly-performant starting point that's fast, cost-efficient, and easy to customize into world-action models for specific robot embodiments."

The T3000 emulation mode will be available later in July with JetPack 7.2.1. The T2000 emulation mode is planned for a future release. The modules will be available in the first quarter of 2027.

Related articles

Infineon, Nvidia Advance Humanoid Robotics with Digital Twins

ST Partners with Nvidia on Physical AI

Nvidia GTC: More Support for 800-VDC Data Centers

The post Nvidia Expands Thor Modules for Scalable Robotics and Edge AI appeared first on Embedded.