Just now, NVIDIA launched the most powerful "brain" for humanoid robots, with its AI performance skyrocketing by 7.5 times and computing power soaring to 2070 TFLOPS.

The most powerful robot "brain" on the planet has changed hands again!

According to a report by Zhidx on August 25th, today, NVIDIA launched the NVIDIA Jetson Thor, specifically designed for physical AI and humanoid robots. It is hailed by Jensen Huang, the founder and CEO of NVIDIA, as "the ultimate supercomputer to drive the era of physical AI and general-purpose robots."

Jetson Thor is equipped with an NVIDIA Blackwell GPU, a 14-core Arm Neoverse CPU, and 128GB of video memory. The video memory bandwidth is 273 GB/s. The peak AI computing power reaches 2070 TFLOPS at FP4 precision and 1035 TFLOPS at FP8 precision. It can accelerate generative AI and large Transformer models at the edge.

It supports various generative AI models, including VLA (Visual Language Action) models, LLM (Large Language Models), and VLM (Visual Language Models). It can handle real-time video data streams and AI inference, making it suitable for building AI agents that can perform visual search and summarization tasks at the edge.

The power of the entire Jetson Thor computer, including CPU, GPU, SLC, DRAM connection, network, and power management, can be configured between 40W and 130W.

With 4 x 25 GbE networks, a camera offload engine, and a Holoscan sensor bridge, Jetson Thor can extract high-speed sensor data and achieve real-time performance.

The key feature of the new robot chip is to run multiple AI workflows, enabling robots to interact with humans and the physical world in real-time and intelligently, thus promoting the development of visual AI agents and complex robot systems.

Compared with the previous-generation Jetson Orin, Jetson Thor offers up to 7.5 times improvement in AI computing performance, up to 3.5 times in energy efficiency, up to 3.1 times in CPU performance, and up to 10 times in I/O throughput.

Compared with 10 years ago, the performance improvement is even more remarkable - the AI performance has increased by up to 7000 times.

Jetson Thor, when paired with the robot AI software platform, supports various mainstream AI frameworks and generative AI models from companies such as ByteDance, DeepSeek, Alibaba Qwen, Google Gemini, Meta, Mistral AI, OpenAI, and Physical Intelligence (π).

It is also fully compatible with NVIDIA's software stack from the cloud to the edge, including the Isaac platform for robot simulation and development, the Isaac GR00T humanoid robot foundation model, NVIDIA Metropolis for visual AI, and NVIDIA Holoscan for real-time sensor processing.

Robots need to be equipped with a variety of sensors to perceive the world and achieve low-latency AI processing. Real-time control frameworks typically run at frequencies between 100Hz - 1kHz, perception and planning usually run at 30Hz, and advanced inference typically runs at 1 - 5Hz, just like human thinking, which may take a few seconds.

When processing 16 sensor inputs in parallel and running the Llama 3B and Qwen 2.5 VL 3B models, Jetson Thor can generate the first token within 200ms and output each token within 50ms. This means it can generate more than 25 tokens per second when running these models, doubling the performance compared to the previous generation.

Designed for general inference, when running inference models such as Alibaba Qwen 3 - 30B - A3B, NVIDIA Cosmos Reason 1 7B, and DeepSeek - R1 - Qwen - 32B, Jetson Thor has up to 3 - 5 times performance improvement at FP8 precision, and up to 10 times at FP4 precision.

The software in NVIDIA's CUDA ecosystem is continuously optimized throughout the entire lifecycle of Jetson. For example, over its lifecycle, software upgrades have increased the performance of Xavier by 50% and that of Orin by 100%.

With continuous software optimization in the future, Jetson Thor will achieve even greater performance improvements.

Since 2014, NVIDIA's Jetson platform and robot software stack have attracted an ecosystem of over 2 million developers and over 150 hardware system, software, and sensor partners. Over 7000 customers have adopted Jetson Orin.

Star humanoid robot companies such as Zhongqing Robot, Galaxy Universal, Ubtech, and Unitree Technology, medical enterprises like United Imaging Healthcare, and intelligent transportation companies such as Wanji Technology have all been the first to adopt Jetson Thor.

The NVIDIA Jetson AGX Thor developer kit is now available worldwide, starting at $3499 (approximately RMB 25,000).

The Jetson T5000 and Jetson T4000 modules can be obtained from global distribution partners. The Jetson T5000 starts at $2999 (approximately RMB 21,500), and the Jetson T4000 starts at $1999 (approximately RMB 14,300).

The specific specifications are as follows:

The NVIDIA DRIVE AGX Thor development kit is a development platform designed by NVIDIA for safe autonomous driving vehicles. It has passed safety certification and is also equipped with a Blackwell GPU with a built - in generative AI engine, along with a rich set of SDKs and libraries. This development kit is now available for pre - order.

Leading intelligent driving vehicle companies such as BYD, DeepRoute.ai, GAC, IM Motors, Li Auto, WeRide, Xiaomi, Zeekr, and Zhuoyu are actively embracing DRIVE AGX Thor.

For humanoid robot development, NVIDIA provides basic systems, blueprints, tools, services, algorithms, and other robot technologies. Collaborating with the ecosystem, it offers an end - to - end complete workflow for the four important steps (data generation, model training, simulation testing, and deployment inference) of building robot products and bringing them into the real world.

NVIDIA's robot business is growing rapidly. This year, NVIDIA combined its automotive and robot businesses in its financial reports. The revenue in the first quarter was $567 million, a year - on - year increase of 72%.

Currently, NVIDIA is focused on building three computers for physical AI and robots, including the NVIDIA DGX AI supercomputer for model training, the NVIDIA OVX computer for synthetic data generation and simulation testing, and real - time computers installed on the robot body (such as Jetson Thor).

From perception AI, generative AI, Agentic AI to future - oriented physical AI, NVIDIA is expanding its computing territory across the entire lifecycle of AI.

This article is from the WeChat official account "Zhidx" (ID: zhidxcom), author: ZeR0, editor: Moying. It is published by 36Kr with authorization.