Compared with Jetson AGX Orin, Jetson AGX Thor offers up to 7.5x the AI performance and around 3.5x better energy efficiency, making it better suited for very large generative and multimodal models or fleets of high-end robots. It also provides more memory (up to 128 GB vs. 64 GB), higher-bandwidth networking, and richer I/O, while still fitting into an embedded power envelope. For simpler vision or control workloads, Jetson Orin Nano/NX or AGX Orin modules may offer a better balance of cost, power, and performance.

Jetson Thor is NVIDIA's next-generation edge AI and robotics platform built on the Blackwell GPU architecture, delivering up to 2070 FP4 / 1035 FP8 TFLOPS of AI performance and significantly higher efficiency than Jetson Orin. It combines a powerful GPU, multi-core Arm CPU, and up to 128 GB LPDDR5X memory with high-speed networking (multiple 25 GbE lanes), rich I/O, and advanced video encode/decode to handle sensor-dense, multimodal AI workloads. Jetson Thor is tightly integrated with the NVIDIA AI software stack, including Isaac for robotics and Holoscan for sensor processing, so developers can deploy state-of-the-art generative and perception models directly on the robot or edge device.

To get started with NVIDIA Jetson AGX Thor, we recommend following NVIDIA's official Quick Start Guide for the Jetson AGX Thor Developer Kit, which walks you through BSP installation and initial setup. (https://docs.nvidia.com/jetson/agx-thor-devkit/user-guide/latest/quick_start.html) You can then explore the NVIDIA Developer Blog article introducing Jetson Thor as a platform for physical AI to better understand its capabilities and example use cases.(https://developer.nvidia.com/blog/introducing-nvidia-jetson-thor-the-ultimate-platform-for-physical-ai/) We will also continue to update relevant Jetson Thor resources and application notes on Seeed's documentation channels as new projects and products are released.

Jetson Thor boards target "physical AI" workloads such as humanoid and mobile robots, collaborative arms, warehouse and logistics robots, and complex autonomous machines that need large generative or multimodal models at the edge. They are also suitable for high-end industrial inspection, smart manufacturing cells, and other scenarios where you must fuse many sensors and run very large AI models in real time without relying on the cloud.