NVIDIA Introduces Blackwell GPU for the Humanoid Robot ‘Brain’.Author | ZeR0 Editor | Mo YingThe strongest robot ‘brain’ has changed hands again!According to a report by Zhixiaodong on August 25, NVIDIA today launched the NVIDIA Jetson Thor, designed specifically for physical AI and humanoid robots, which NVIDIA founder and CEO Jensen Huang described as the “ultimate supercomputer driving the era of physical AI and general robotics“.Jetson Thor features NVIDIA’s Blackwell GPU, a 14-core Arm Neoverse CPU, and 128GB of video memory, with a memory bandwidth of 273 GB/s. The peak AI computing power at FP4 precision is 2070 TFLOPS, and at FP8 precision, it reaches 1035 TFLOPS, capable of accelerating 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), and can handle real-time video data streams and AI inference, suitable for building AI agents that can perform visual search and summarization tasks at the edge.
The entire Jetson Thor computer’s power can be configured between 40W and 130W, including CPU, GPU, SLC, DRAM connections, networking, and power management.With the help of four 25 GbE networks, camera offload engines, and Holoscan sensor bridges, Jetson Thor can extract high-speed sensor data to achieve real-time performance.The new robot chip focuses on running multiple AI workflows, enabling robots to interact intelligently and in real-time with humans and the physical world, promoting the development of visual AI agents and complex robotic systems.Compared to the previous generation Jetson Orin, Jetson Thor’s AI computing performance has increased by up to 7.5 times, energy efficiency has improved by up to 3.5 times, CPU performance has increased by up to 3.1 times, and I/O throughput has improved by up to 10 times.
When compared to ten years ago, the performance has increased dramatically—AI performance has improved by as much as 7000 times.
Jetson Thor, in conjunction with the robot AI software platform, supports various mainstream AI frameworks, as well as 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 cloud to edge, including the Isaac platform for robot simulation and development, the Isaac GR00T humanoid robot base 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 operate at frequencies of 100Hz-1kHz, perception and planning usually run at 30Hz, and advanced inference typically runs at 1-5Hz, similar to human thought processes that may take a few seconds.
When processing inputs from 16 sensors in parallel and running the Llama 3B and Qwen 2.5 VL 3B models, Jetson Thor generates the first token in under 200ms, with an output time of 50ms per token, meaning it can generate over 25 tokens per second, doubling the performance of the previous generation.
Jetson Thor is designed for general inference, achieving performance improvements of up to 3-5 times when running inference models such as Alibaba Qwen 3-30B-A3B, NVIDIA Cosmos Reason 1 7B, and DeepSeek-R1-Qwen-32B at FP8 precision, and even more at FP4 precision, with improvements of up to 10 times.
The software in NVIDIA’s CUDA ecosystem is continuously optimized throughout the lifecycle of Jetson. For example, performance improvements of 50% for Xavier and 100% for Orin have been achieved through software upgrades over their lifecycles.With ongoing software optimizations in the future, Jetson Thor will achieve even greater performance enhancements.
Since 2014, the NVIDIA Jetson platform and robot software stack have attracted over 2 million developers and an ecosystem of over 150 hardware systems, software, and sensor partners. More than 7000 customers have adopted Jetson Orin.
Leading humanoid robot companies such as Zhongqing Robotics, Galaxy General, UBTECH, and Yushu Technology, as well as medical enterprises like United Imaging Healthcare and smart transportation companies like Wanjitech, have already adopted Jetson Thor.The NVIDIA Jetson AGX Thor developer kit is now available globally, priced at $3499 (approximately ¥25,000).Jetson T5000 and Jetson T4000 modules can be obtained from global distribution partners. The price for Jetson T5000 is $2999 (approximately ¥21,500) for orders over 1000 units.The specific specifications are as follows:
The NVIDIA DRIVE AGX Thor development kit is a development platform designed for safe autonomous vehicles, certified for safety, and also equipped with the Blackwell GPU with a built-in generative AI engine, featuring a rich SDK suite and libraries. This development kit is now available for pre-order.Leading smart driving companies such as BYD, Yuanrong Qixing, GAC, Zhiji, Li Auto, Wenyan Zhixing, Xiaomi, Zeekr, and Zhuoyu are actively embracing DRIVE AGX Thor.For humanoid robot development, NVIDIA provides foundational systems, blueprints, tools, services, algorithms, and other robotic technologies, collaborating with the ecosystem to provide an end-to-end complete workflow for building robotic products and bringing them to the real world through four important steps (data generation, model training, simulation testing, deployment inference).NVIDIA’s robotics business is growing rapidly.This year, NVIDIA merged its automotive and robotics businesses in its financial report, with first-quarter revenue reaching $567 million, a 72% year-on-year increase.Currently, NVIDIA is focused on building three computers for physical AI and robotics, including the NVIDIA DGX AI supercomputer for training models, the NVIDIA OVX computer for synthetic data generation and simulation testing, and real-time computers installed on robots (such as Jetson Thor).From perception AI, generative AI, Agentic AI to future-oriented physical AI, NVIDIA is expanding its computational territory across the entire lifecycle of AI.
(This article is original content from NetEase News • NetEase account feature content incentive program signed account 【Zhixiaodong】. Unauthorized reproduction is prohibited.)