On August 26, Beijing time, NVIDIA released its robotic computing platform Jetson Thor. Just 12 hours later, the robot Psi V1 personally opened its ‘new brain’ at the Lingchu Intelligent office in Haidian, Beijing.
This is one of the first batches of Jetson Thor to arrive globally. Currently, this chip has officially been released in the form of a developer kit, priced at $3,499.
Regarding the ‘new brain’ NVIDIA has provided for robots, engineer Zhong Yifan from Lingchu Intelligent told First Financial News that Thor’s breakthroughs in computing power and data processing capabilities allow robots to run large-scale, high-performance models directly on the edge, enabling them to perform more general and complex tasks.
‘Currently, Lingchu Intelligent is using NVIDIA’s graphics card chips for training, and for model deployment on the edge, we are using NVIDIA’s Orin and Thor,’ Zhong Yifan said.
According to data disclosed by NVIDIA, the new generation Thor chip is based on the Blackwell architecture, providing up to 2070 TFLOPS at FP4 precision, which translates to a peak computing power of 2.07 quadrillion floating-point operations per second, a 7.5-fold increase over the previous generation Orin chip, with energy efficiency improved by 3.5 times.
Currently, robots generally adopt a hybrid deployment model of cloud + edge. An executive from a Shenzhen robotics company candidly told First Financial News: ‘More general robot models require massive computing power.’
Before the massive computing power challenge is resolved, a fast-slow system architecture may be a stopgap measure. Xu Huazhe, co-founder of Xinghai Map, explained that the fast system is mainly deployed on the edge, responsible for execution and immediate response, such as perception, control, and response, emphasizing real-time performance and stability. The slow system, on the other hand, is more cloud-based, handling understanding and reasoning tasks, especially in scenarios requiring long-chain logical deductions, which are typically complex tasks.
However, cloud deployment also brings real-world issues. ‘Latency is hard to avoid, and for some high-frequency scenarios, such as continuous grabbing and rapid judgment, latency can directly affect product safety and feasibility,’ the executive stated.
Zhong Yifan mentioned that the improvements in data processing capabilities and interface bandwidth of Thor allow robots to directly handle multimodal inputs from high-resolution, high-frequency sensors and complete processing in a timely manner. Therefore, more tasks that originally relied on cloud processing may gradually be brought back to be completed locally by the robot, which could accelerate the deployment of robots in high-frequency, complex interaction scenarios.
‘First, stack the computing power ceiling, then use full-stack software like Isaac, Cosmos, and GR00T to shape developer habits. NVIDIA has only one goal,’ said Zheng Yangyang, an AI robotics industry researcher at Samoyed Cloud. He noted that NVIDIA continues the logic it had before the AIGC explosion: to set up the infrastructure first and strive to influence industry standards before the industry explodes.
However, Zheng Yangyang pointed out that compared to the relatively centralized scenarios of large model training, robot applications are more fragmented, facing both cost pressures and the need for long-term scenario validation. ‘Competitors can still form differentiated advantages in low-power chips, niche scenarios, or open-source ecosystems, which is also an opportunity for Chinese manufacturers.’
Within the Chinese camp, several manufacturers have already begun deploying in the ‘brain’ segment.
In June this year, Digua Robotics released the RDK S100 development kit, which adopts a ‘big and small brain’ heterogeneous architecture design, focusing on balancing inference and real-time motion control, reducing the size and complexity of the control system.
Hezhima Intelligent also provided the Huashan A2000 chip and Wudang C1236 chip for Wuhan University’s ‘Tianwen’ humanoid robot, used respectively for the robot’s ‘big brain’ and ‘small brain.’
Rockchip has launched its flagship chip RK3588, tailored for the characteristics of robots, supporting multimodal data processing and high-performance computing, providing underlying computing power support for the robot’s perception, decision-making, and execution.
Zheng Yangyang stated that the advantage of domestic chips lies in their higher cost-performance ratio and more market-oriented customization services. ‘Because they are closer to domestic robot manufacturers, they also have certain differentiated advantages in scenario optimization.’ Therefore, Zheng Yangyang added that the ‘computing power ceiling’ built by NVIDIA may not be the only answer for the maturity of the robotics industry.
Source: First Financial News
Editor: Zheng Zhehao Review: Jiang Bo, Chen Jie