en.Wedoany.com Reported - NVIDIA has further expanded its Thor-based autonomous machine product line with the introduction of the Jetson T3000 and T2000 modules. These two modules are primarily aimed at developers of robotics, vision AI agents, and industrial systems, enabling them to run AI models locally rather than in the cloud. This launch extends the performance range of the Jetson series from 70 TOPS to 2000 teraflops.

The Jetson T3000 is the more powerful of the two products. According to NVIDIA, it delivers 865 FP4 teraflops of AI compute power in a compact form factor, with approximately half the size and power consumption of the T5000. The module integrates a Blackwell GPU, an eight-core Neoverse Arm CPU, 32GB of LPDDR5X memory, 273GB/s of memory bandwidth, and 25 GbE connectivity. Additionally, the IGX T3000 variant adds functional safety support and features the Halos for Robotics safety system for human-robot collaboration scenarios.

NVIDIA notes that the T3000 is particularly well-suited for handling multimodal reasoning tasks, encompassing large language models (LLMs), vision language models (VLMs), vision language action models (VLAMs), and world foundation models (WFMs). The company states that under the same workload, its inference performance is comparable to the T5000, but with a smaller footprint. The Jetson T2000, on the other hand, is positioned as a lower entry point for the Thor series, offering 400 FP4 teraflops of compute power and 16GB of memory, suitable for development scenarios involving autonomous mobile robots, industrial robotic arms, and vision AI agents.

On the software front, NVIDIA has simultaneously launched Jetson agent skills—a set of tools designed to automate tasks such as memory optimization, system configuration, and deployment for Jetson Thor and Jetson Orin devices. The company claims this software can compress optimization cycles from weeks to days and allows customers to opt for lower memory configurations without sacrificing performance. This capability is particularly significant in the context of high memory costs and stringent hardware constraints in edge systems.

NVIDIA cited several early adopters to illustrate the effectiveness of these tools. UBTech, Agile Robots, and Connect Tech successfully migrated from the Jetson AGX Orin 64GB module to the 32GB module by reducing memory usage by up to 15GB. In the retail sector, SandStar reduced memory consumption by up to 4GB, enabling deployment on the Jetson Orin NX 8GB module instead of the 16GB version. In the transportation sector, NoTraffic cut memory requirements by 30% on the Jetson TX2 NX, freeing up space to run more AI functions on the same hardware. Companion robot manufacturer GROOVE X also utilized Jetson's hybrid AI accelerator to reallocate workloads and alleviate memory pressure. NVIDIA stated that these cases demonstrate that optimized software can achieve cost reductions comparable to hardware upgrades.

To complement the Thor platform, NVIDIA has also expanded the Cosmos 3 series with the introduction of Cosmos 3 Edge. This is a 4-billion-parameter world foundation model designed for embodied AI systems that require real-time environmental analysis and on-device generation of action commands. Developers can use NVIDIA's open-source Cosmos framework to adapt Cosmos 3 Edge to specific robots and sensors in approximately one day, then deploy it on Jetson Thor for on-device visual analysis and robot policy execution. This release reflects NVIDIA's overall strategy of bundling hardware sales with a software stack for robotics, simulation, and model deployment, which includes Isaac for robot simulation and perception, as well as proprietary models like Nemotron, Cosmos 3, and Isaac GR00T.

According to NVIDIA, the Jetson AGX Thor has been adopted by robotics teams including 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi, and Techman Robot. A broader list of hardware partners includes ADLINK, Advantech, AAEON, Aetina, Auvidea, AVerMedia, Connect Tech, ForeCR, JWIPC, NEXCOM Robotic Solutions, Realtimes, Seeed Studio, Twowin, TZTEK, and YUAN, all of which have begun offering Thor-based systems. Software partners include Antmicro, Neurealm, REBOTNIX, and RidgeRun, providing simulation and migration support for customers transitioning to the new modules. Developers can use the existing Jetson AGX Thor developer kit to simulate the performance of the T3000 and T2000, as these modules share the same chip architecture and software stack. The T3000 simulation mode will be available later this month with JetPack 7.2.1, while T2000 support will be provided in a subsequent release. The Jetson T3000 and T2000 modules are scheduled to be available in the first quarter of 2027.