Nvidia is increasing its Jetson Thor platform with the 2 new computing modules Jetson T3000 and Jetson T2000. The techniques are supposed to allow high-performance AI purposes immediately in robots, autonomous automobiles, industrial machines, and visible edge AI techniques. In distinction to traditional AI servers, the computations are carried out immediately on the machine. In consequence, robots can analyze sensor information domestically and make choices with out requiring a connection to a cloud information heart for each processing step. Nvidia introduced the 2 modules on 15 July 2026. Market launch is scheduled for the primary quarter of 2027.
Jetson T3000 with 865 FP4 teraflops
The Jetson T3000 is the extra highly effective of the 2 new modules. Nvidia combines a Blackwell GPU with an Arm processor based mostly on the Neoverse structure.
The gear consists of:
-
eight Arm Neoverse CPU cores
-
32 GB LPDDR5X reminiscence
-
273 GB/s reminiscence bandwidth
-
25 Gigabit Ethernet
-
865 FP4 teraflops of AI compute efficiency
Based on Nvidia, the T3000 requires roughly half as a lot house and energy because the bigger Jetson T5000. In multimodal inference duties, it’s however anticipated to realize comparable efficiency. These embrace giant language fashions, vision-language fashions, robotic motion fashions, and so-called World Basis Fashions. The said FP4 efficiency describes computations with very low numerical precision. It’s significantly appropriate for contemporary AI inference, however can’t be immediately in contrast with standard FP32 graphics or compute efficiency. Along with the usual Jetson T3000, Nvidia is planning an IGX-T3000 model. It’s to supply the identical primary compute efficiency however embrace extra features for practical security. The system is meant for robots that work immediately alongside folks. Attainable areas of use embrace manufacturing strains, hospitals, warehouses, and autonomous transport techniques. The platform helps Nvidia’s Halos software program stack, which is meant to offer security features for robotic techniques. The Jetson T2000 is geared toward much less compute-intensive purposes. The module provides 400 FP4 teraflops and 16 GB of reminiscence.
Nvidia lists the next utility areas, amongst others:
With the T2000, the Thor structure is meant for use not solely in costly humanoid robots, but additionally in additional compact and cost-effective industrial merchandise. Along with the brand new modules, Nvidia is introducing so-called Jetson Agent Expertise. These software program instruments are supposed to assist builders with reminiscence optimization, system configuration, and deployment of AI fashions. Nvidia cites a number of examples from its companion community. UBTech, Agile Robots, and Join Tech are mentioned to have lowered the reminiscence consumption of sure purposes by as much as 15 GB. This enabled some techniques to be moved from a Jetson AGX Orin with 64 GB to a variant with 32 GB of reminiscence. The transport firm NoTraffic is alleged to have lowered the reminiscence necessities of an utility on the Jetson TX2 NX by round 30 %. Nonetheless, these figures come from Nvidia and taking part companions and haven’t been independently verified. One other element of the announcement is Cosmos 3 Edge. It is a World Basis Mannequin with 4 billion parameters. The mannequin is meant to guage sensor information, interpret a robotic’s setting, and predict or generate doable actions. Processing takes place immediately on a Thor-based system. Builders are anticipated to have the ability to adapt Cosmos 3 Edge to particular robots, cameras, and sensors. This might, for instance, permit a cellular robotic to study to detect obstacles, calculate actions upfront, and reply to them in actual time. Builders don’t have to attend till 2027 for the completed T3000 and T2000 modules. The already obtainable Jetson AGX Thor improvement system can partially emulate the efficiency of the brand new modules. The T3000 emulation mode is ready to look in July 2026 with JetPack 7.2.1. Emulation for the T2000 is to observe later. The precise modules are scheduled to turn out to be obtainable within the first quarter of 2027.
Conclusion
With the T3000 and T2000, Nvidia is transferring its established platform technique to the robotics market. The corporate offers not solely the processor, however combines {hardware}, AI fashions, simulation software program, developer instruments, and safety features. The T3000 is prone to be of explicit curiosity for high-performance humanoid and industrial robots. The smaller T2000, against this, may turn out to be extra vital for the mass deployment of edge AI techniques. Essential financial particulars stay open for now. Nvidia has not but named official end-customer costs. As well as, it stays to be seen how the modules carry out underneath real-world situations by way of energy consumption, cooling, and constantly obtainable AI efficiency.
