Jetson T3000 puts 865 FP4 TFLOPS inside smaller robot hardware
Original: NVIDIA Introduces New Jetson Thor Computers to Advance Mainstream Robotics and Edge AI View original →
The robotics AI race is moving from lab demos to a harsher constraint: whether a robot can run perception, reasoning, and action models inside its own power and memory budget. NVIDIA’s July 15 release adds two Jetson Thor modules, T3000 and T2000, aimed directly at that deployment problem. The headline number is 865 FP4 TFLOPS for T3000 in a package NVIDIA says is roughly half the size and power of the T5000 while delivering similar multimodal inference performance.
T3000 combines a Blackwell GPU, an eight-core Neoverse Arm CPU, 32GB of LPDDR5X memory, 273GB/s of memory bandwidth, and 25 GbE connectivity. The IGX T3000 variant adds integrated functional safety and runs the NVIDIA Halos for Robotics stack, which matters for machines operating near people rather than behind a lab barrier. NVIDIA also named 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi, and Techman Robot among companies building on Jetson AGX Thor, giving the release a stronger signal than a standalone developer-board refresh.
T2000 brings the same Thor architecture to lower-cost edge systems. With 400 FP4 TFLOPS and 16GB of memory, it gives developers a smaller entry point for visual AI agents, autonomous mobile robots, industrial manipulators, and other intelligent machines. NVIDIA frames the updated Jetson range as spanning 70 TOPS to 2,000 TFLOPS, a wide enough ladder for teams to move from prototypes to production SKUs without changing the whole software stack.
The less flashy but commercially important piece is memory. NVIDIA’s new Jetson agent skills automate memory optimization, system configuration, and deployment tasks that previously required specialized engineering time. The company cites up to 15GB of memory savings for UBTech, Agile Robots, and Connect Tech, enough to move from Jetson AGX Orin 64GB to the 32GB module in some cases. SandStar moved from an Orin NX 16GB target to 8GB, while NoTraffic reduced memory use by 30% on Jetson TX2 NX. Those reductions turn model optimization into bill-of-materials leverage.
The release also extends NVIDIA Cosmos 3 with Cosmos 3 Edge, a 4-billion-parameter world foundation model for embodied systems. NVIDIA says developers can post-train it for specific embodiments and sensors in about a day using the open Cosmos framework, then deploy it on Jetson Thor for real-time vision analysis and on-device robot policy. T3000 emulation mode is due later this month with JetPack 7.2.1, while T3000 and T2000 modules are scheduled for Q1 2027. The watch item is whether robotics teams can turn these smaller modules into fielded systems, not just cleaner demos.
Related Articles
NVIDIA said on March 20, 2026 that its Cosmos world foundation models have advanced again with Transfer 2.5, Predict 2.5, and Reason 2. The linked NVIDIA Technical Blog frames the update around higher-quality synthetic data, stronger long-tail scenario generation, and richer reasoning for robots and autonomous vehicles.
NVIDIA released Cosmos 3 as an open physical AI omnimodel with Super and Nano variants. Its technical post points to six synthetic datasets, Hugging Face checkpoints, and GitHub recipes for domain adaptation.
Robostral Navigate is an 8B embodied navigation model that uses a single RGB camera instead of LiDAR, depth sensors, or multiple cameras. Mistral says it reaches 76.6% success on unseen R2R-CE environments and beats multi-sensor systems by 4.5 points.