NVIDIA Expands Jetson Thor for Edge Robotics

As robots and autonomous machines are deployed in real-world environments, demand for compact, energy-efficient edge AI supercomputers is increasing. NVIDIA announced the Jetson T3000 and T2000 modules, based on the Thor architecture.

The T3000 delivers 865 FP4 teraflops of AI compute performance, and is equipped with a Blackwell GPU, an 8-core Neoverse Arm CPU, 32GB of LPDDR5X memory, and 273GB/s of memory bandwidth. It achieves similar multimodal inference performance to the T5000 at half the size and power consumption. The IGX T3000 integrates functional safety and runs the robot safety system NVIDIA Halos alongside it. Major companies including Boston Dynamics, Amazon Robotics, and 1X are adopting the Thor platform.

The T2000 offers 400 FP4 teraflops and 16GB of memory, providing an entry point for a broader range of edge AI systems such as visual AI agents, autonomous robots, and industrial manipulators. Jetson now supports a performance range from 70 TOPS to 2,000 teraflops.

Jetson Agent Skills is an AI agent tool that automates memory optimization, system configuration, and deployment tasks. UBTech and Agile Robots used it to reduce memory usage by up to 15GB, migrating from 64GB to 32GB modules. SandStar saved 4GB in smart retail, and NoTraffic achieved a 30% reduction in traffic systems.

Cosmos 3 Edge is a lightweight model with 4 billion parameters that helps robots see the world, reason in real time, and predict and generate actions. Developers can train it for a specific robot in about a day, then run it in real time on Jetson Thor.

The Jetson AGX Thor developer kit allows emulation of T3000 and T2000 performance. T3000 emulation will be supported starting with JetPack 7.2.1 later this month, while T2000 support will come in a later release. The actual modules are scheduled to launch in Q1 2027.