2024.04.02 11:43
I'm PortAI, I can summarize articles.

Far more than the B200, this is the soul of NVIDIA

Huang Renxun released the world's most powerful AI chip Blackwell on GTC, promoting the development of generative AI. He also unveiled a multimodal humanoid robot model and robot chip, calling it "NVIDIA's soul". The humanoid robot is capable of sensing the world, making decisions, executing tasks, and engaging in conversations with humans. In addition, an artificial intelligence robot company named Figure has introduced a robot called Figure 01, which has a certain level of response speed and the ability to understand human intentions

On March 18th at GTC, NVIDIA founder Jensen Huang appeared in his iconic black leather jacket and unveiled the world's most powerful AI chip - Blackwell, pushing generative AI to a tipping point.

The GPU architecture of Blackwell has 208 billion transistors, using a custom, double reticle TSMC NP (an improved version of the 4N process) manufacturing process, with an interconnect speed of up to 10TBps between the two small chips, significantly enhancing processing capabilities.

Moore's Law is dead, but NVIDIA has given birth to a new Moore's Law. Last year, Jensen Huang declared the "iPhone moment" for AI has arrived. Now, Jensen Huang is also known as the "new Steve Jobs".

At the event, Jensen Huang also unveiled a multimodal humanoid robot model and robot chip, calling it "NVIDIA's soul".

01 The World of Humanoid Robots

Unlike traditional industrial robots that exist only to perform specific tasks, humanoid robots have carried the expectation of intelligence since their inception - to perceive the world like humans, make decisions, and execute tasks.

In 1973, Waseda University in Japan introduced the world's first full-size humanoid robot WABOT-1, which could communicate in Japanese, measure direction and distance, walk slowly, grab objects, and had the intellectual capacity of an eighteen-month-old child.

However, half a century later, humanoid robots still seem far from reach. Not to mention the core obstacles of performance and cost in commercialization.

Recently, a video of a humanoid robot went viral overnight on the internet.

In the video, this robot performed simple grabbing and placing actions based on verbal commands from humans, such as handing an apple to a human, tidying up plastic bags into a basket, and placing cups and plates on a draining rack.

Throughout the interaction, the robot conversed fluently with humans, although there was still a certain gap in response speed compared to real humans, it had clearly improved significantly.

At the same time, the robot could understand human intentions, while also explaining the reasons for carrying out the instructions. This robot is named Figure 01, and it is launched by an artificial intelligence robot company called Figure. Behind all this, there is not only Figure's proud "end-to-end" neural network technology but also the strong support of OpenAI's large models.

Now, Figure has sent the first humanoid robot Figure 01 to BMW's Spartanburg car factory to replace humans in high-risk operations.

Currently, it's not just Figure manufacturing humanoid robots. In fact, many technology companies have begun to showcase their own humanoid robots.

Amazon is testing a bipedal humanoid robot Digit in its warehouse operations, designed by Agility Robotics, in which Amazon has invested.

Tesla unveiled the humanoid robot Tesla Optimus Gen 2. The second-generation humanoid robot appears to have greater flexibility than the previous generation - including more elegant walking postures and greater freedom of movement in the hands, even capable of performing delicate tasks like cooking eggsIt has only been three years since Tesla launched its first generation humanoid robot.

In March, Samsung Electronics formulated a strategy to enter the robot market. They invested over 300 million RMB in the local robot manufacturer Rainbow Robotics. The company's robot, HUBO, is South Korea's first bipedal walking robot.

This has sparked a wave of financing for humanoid robots.

On the international front, humanoid robot company Figure A has raised 675 million USD. Besides OpenAI, the investors in this financing round include Microsoft, NVIDIA, Amazon founder Jeff Bezos, Intel, venture capital firms Parkway Venture Capital, Align Ventures, and "Wood Sister" Ark Fund, among others.

Apart from Figure AI, which is highly favored by tech giants, Norwegian robot startup 1X Technologies supported by OpenAI announced the completion of a 100 million USD Series B financing. This round of financing was led by Swedish venture capital firm EQT Ventures, with other investors including Samsung Electronics' investment company Samsung NEXT, Nistad Group, Sandwater, and Skagerak Capital, among other Norwegian venture capital firms.

In China, a total of 9 humanoid robot companies received over 1.9 billion RMB in financing in 2023. Among them, 3 companies raised over 100 million RMB in a single round of financing, and humanoid robot startup Zhiyuan Robot received five rounds of financing in nearly half a year.

Intelligent robot company Unitree Technology recently announced the completion of nearly 1 billion RMB in Series B2 financing. The investors in this round include Meituan, Jinshi Investment, Source Code Capital, and existing shareholders Deep Ventures, Zhongwang Investment, Rongyi, Dunhong, and Midas Stone. According to Unitree Technology, the company currently has two major product series: quadruped robots and general humanoid robots, with general humanoid robots being the new business direction the company embarked on in 2023.

02 NVIDIA's Layout in Humanoid Robots

NVIDIA's humanoid robot project, codenamed GROOT. The name is likely a tribute to the Marvel superhero Groot.

NVIDIA's new platform is called the "General Humanoid Robot Basic Model." In fact, NVIDIA is providing an AI platform for the emerging humanoid robot manufacturers.

These include 1X Technologies, Agility Robotics, Apptronik, Boston Dynamics, Figure AI, Fourier Intelligence, Sanctuary AI, Unitree Technology, and XPeng, among othersThis almost covers all the well-known humanoid robot manufacturers, except for a few exceptions like Tesla.

Robots powered by GR00T have two major features: the ability to understand natural language, observe human behavior, and mimic actions. These two features enable robots to quickly learn coordination, flexibility, and other skills to navigate, adapt, and interact with the real world.

Huang Renxun also stated: "Robots supported by this platform will never lead to a robot uprising."

At the conference, videos showed several robots completing various tasks. Some robots can mimic human drumming, some can learn human dance moves from videos and simulate them. Based on large language models, some humanoid robots can also understand verbal commands from humans and respond.

In addition, NVIDIA also released a new humanoid robot computer Jetson Thor based on the NVIDIA Thor System-on-Chip (SoC).

This SoC includes a next-generation GPU with a transformer engine, based on the NVIDIA Blackwell architecture, providing 8-bit floating-point AI performance of 8 trillion operations per second to run multimodal generative AI models like GR00T.

Currently, Disney's two small robots, Orange and Green, use NVIDIA's first AI chip Jetson designed for robots.

Huang Renxun said: "The next generation of robots is likely to be humanoid robots because we have more imitation training data to provide to robots that are similar to us. Humanoid robots can be more efficiently deployed in human-designed workstations, manufacturing, and logistics processes."

03 The Billion-Dollar Market Behind Humanoid Robots

A report bluntly states that humanoid robots are expected to become a billion-dollar blue ocean market, with the global humanoid robot manufacturing and home service market estimated to reach $110.3 billion by 2035 under neutral assumptions.

The industry chain that humanoid robots can involve is even larger. Overall, the core hardware and software systems in the upstream are crucial, with hardware mainly including servo motors, reducers, controllers, sensors, high-performance chips, etc.; software includes machine vision, human-computer interaction, machine learning, system control, etc.

Robots consist of three key technical modules: motion module, sensor module, and artificial intelligence module. For general traditional robots, having just one of these technologies often provides usability. For example, industrial robots mainly focus on motion control technology, while robotic vacuum cleaners focus on navigation and sensing technology.


For example, in the case of the second-generation Optimus mentioned earlier, if we carefully analyze its components, we will find: Tesla-designed actuators and sensors, 2-degree-of-freedom driven neck, faster response 11-degree-of-freedom dexterous hand, tactile sensors (ten fingers), actuator-integrated electronics and wiring harness, foot force/torque sensors, articulated toes, etcThe biggest incremental improvement in the second-generation Optimus product lies in the sensors. When Tesla showcased Optimus, Gen2 demonstrated better full-body control and body balance capabilities (deep squat), as well as more precise hand operations (handling eggs).

Humanoid robots require various sensors, including torque sensors, tension sensors, encoders, temperature sensors, six-axis force sensors, and inertial navigation sensors, among others. Among them, the inertial navigation sensor (IMU) is the core for achieving posture control of humanoid robots, and the six-axis force sensor is the core for force control and swing stability control, capable of accurately measuring random force changes.

Currently, related companies in the sensor field include Hanwei Technology, Suzhou Nengsda, Hongxin Electronics' subsidiary Ruihu Technology, Kelite Sensing, and Donghua Testing.

High Computing Power Chips

As highly integrated robots, humanoid robots have higher chip computing power requirements compared to industrial robots. For example, Xiaomi's latest release, the bionic quadruped robot CyberDog 2, is equipped with 21T of computing power; and the recently released expedition A1 humanoid robot by Zhiyuan Robot is also equipped with up to 200T of AI computing power. This aspect mainly depends on leading manufacturers such as NVIDIA, Intel Mobileye, and Qualcomm.


Open AI CEO Sam Altman tweeted on X: "This is the most interesting year in human history, but it will be the most boring year in the future."

This tweet may be a footnote to the era.

Currently, according to the "Action Plan for Further Promoting the Construction of New Infrastructure in Shanghai (2023-2026)", Shanghai will focus on five aspects: new networks, new computing power, new data, new facilities, and new terminals in 2024, with a focus on promoting 56 tasks. Shanghai will strive to establish the national manufacturing innovation center for humanoid robots.

Now, at the NIO Automobile Factory in Hefei, Anhui, the newly arrived Walker S industrial version humanoid robot, after inspecting the door locks, seat belts, and headlight covers of the new car, walks straight to the front of the car and attaches the car badge.

Scenes once written in elementary school compositions are now truly unfolding. Perhaps in the next decade, we can truly achieve harmony and collaboration between robots and humans.

Author: Jiulin, Source: Semiconductor Industry Observer (ID: ICViews), Original Title: "Far More Than B200, This Is NVIDIA's Soul"