NVIDIA announced the launch ofNext-generation Jetson Thor series embedded computing platform, also focusing on various scenarios such as AI inference, robot control and edge computing. By replacing the Blackwell display architecture GPU and combining it with the Arm Neoverse-V3AE processor, Jetson Thor can achieve up to 2070 TFLOPS of AI performance on a mini platform. NVIDIA even calls it "the most powerful and smallest mini PC in history", and the price starts at US$3499.
A major leap forward for embedded AI platforms
From Jetson Nano and Xavier to AGX Orin, NVIDIA has continuously miniaturized data center-grade AI computing power. The newly launched Jetson Thor is the first to bring the Blackwell display architecture to an edge platform.
This product series has applications in areas such as autonomous robots, driverless vehicles, smart manufacturing, medical image analysis, and edge data centers. Emphasizing high performance, energy efficiency, and scalability, it enables developers to perform real-time inference and decision-making without relying on the cloud.
Positioning differences among the three models
This time, NVIDIA launched three models: Jetson AGX Thor Developer Kit, Jetson Thor T5000, and Jetson Thor T4000.
The Jetson AGX Thor Developer Kit for developers and the Jetson Thor T5000 for design applications are both equipped with a 14-core Arm Neoverse-V3AE CPU and a 2560-core Blackwell GPU, capable of 2070 TFLOPS of AI inference capabilities. They are also equipped with 128GB of LPDDR5X memory, and their power consumption ranges from 40W to 130W. Their performance is almost comparable to that of a small AI supercomputer, and they are capable of high-intensity tasks such as visual recognition, natural language processing, and machine learning training.
The newly released Jetson Thor T4000 utilizes a 12-core Arm Neoverse CPU and a 1536-core Blackwell GPU, delivering 1200 TFLOPS of performance and 64GB of LPDDR5X memory, while consuming only 40W to 75W of power. While slightly inferior to the Jetson Thor T5000, it remains a high-end choice among similar embedded platforms, suitable for mid-to-high-end applications such as autonomous mobile robots and edge medical image processing.
Blackwell shows the significance of decentralization of architecture
Since its launch at GTC 2024, the Blackwell display architecture has become a core technology for NVIDIA data center GPU acceleration, doubling training and inference performance while significantly improving energy efficiency and memory bandwidth. Jetson Thor brings this display architecture to a small platform for the first time, enabling edge devices to directly handle tasks previously only handled by servers.
For example, in self-driving car scenarios, Jetson Thor supports multi-camera and radar fusion, enabling real-time road condition assessment. In smart factories, it enables multi-machine collaboration, quality inspection, and edge learning, raising the intelligence level of production automation.
Pricing and Market Outlook
With a starting price of US$3499 (approximately NT$11), Jetson Thor is significantly more expensive than currently available embedded platforms. However, for businesses and developers seeking high-performance edge computing deployments, the Jetson Thor series offers a more flexible option compared to traditional servers with limited space and heat dissipation, while also being relatively cost-effective.
In the future, the launch of Jetson Thor is expected to become a key enabler for edge AI deployment in applications ranging from smart checkout and customer analytics in retail settings, to medical image interpretation, traffic flow prediction, and even military and unmanned systems. NVIDIA also provides a complete software stack, including the JetPack SDK, CUDA, and TensorRT, to accelerate developers' progress from prototyping to commercialization.
