Author:21 Reporter Ni YuqingSource: 21st Century Economic Report (ID: jjbd21)
At the 2025 World Robot Conference (WRC), NVIDIA’s Omniverse and Simulation Technology Vice President Rev Lebaredian, Yuzhu Technology Founder Wang Xingxing, and Galaxy General Founder and CTO Wang He made a rare joint appearance.
During a media communication meeting, they engaged in in-depth discussions on physical AI, simulation, robotic platforms and commercial implementation, synthetic data, and the industrialization path of embodied intelligent models.
Rev Lebaredian first presented a core judgment: Over the past three to four decades, the computer and IT industries have amplified the capabilities of various sectors, but the impact has largely remained in the “information space,” which refers to content that can be digitized. The internet has brought computers into daily life and led to decades of growth, but the total scale of the IT industry is about $5 trillion, which is just a small part compared to the global total of over $100 trillion across all industries. He pointed out that the greater value in other industries lies in reaching the “atoms” of the physical world—transportation, manufacturing, logistics, healthcare, etc.
“Today, with the emergence of artificial intelligence, we finally have the ability to give machines ‘physical intelligence,’ allowing us to truly connect the physical world and the information world.” He emphasized that this means the power of computers will enter the $100 trillion physical world market, and robots are the bridge to achieve this leap.
In his view, China has unique conditions to stand out in this leap. For example, nearly half of the world’s AI researchers and developers are in China, which also possesses unmatched electronic manufacturing capabilities and a vast manufacturing base available for large-scale deployment and testing.
Regarding NVIDIA’s role, he pointed out that the company’s mission is to build computers specifically for the “toughest problems” and to promote the development of robots and the implementation of physical AI. NVIDIA proposes that three types of computers must be built: first, computers embedded in the robot body, such as the Jetson Thor carried by humanoid robots; second, AI factory computers used to process massive data and train models before deployment through DGX and HGX systems; third, simulation computers that generate data through physical laws and test robots in advance, with testing speeds faster than real-time.
Currently, robotic companies like Yuzhu Technology and Galaxy General are collaborating with NVIDIA. For instance, Galaxy General’s G1 Premium humanoid robot is one of the first humanoid robots equipped with NVIDIA Jetson Thor, demonstrating fluidity and operational speed in complex scenarios such as industrial palletizing, depalletizing, and material box handling. Yuzhu Technology has deployed NVIDIA’s full-stack robotic technology on its new humanoid robot R1, optimizing movement and control capabilities through the high-fidelity simulation platform Isaac Sim, and achieving rapid strategy iteration with the Isaac Lab system.
Wang Xingxing stated: “In a sense, I view humanoid robots as an important carrier for general robots.” In the design of humanoid robots, he believes the structure is not as complex as imagined, “essentially consisting of several joint motors connected in series,” so in the future, when general AI matures, “anyone can easily manufacture a humanoid robot, just as people can today buy computer components to assemble a computer. As AI becomes powerful enough, the requirements for hardware will decrease.”
From Yuzhu Technology’s product history, last year, Yuzhu launched a humanoid robot priced at approximately 99,000 RMB, whose architecture has become one of the more mainstream designs globally. This year’s new version is priced at about 39,000 RMB, supports appearance customization, and “is currently in stock, with mass production expected to be completed by the end of the year,” Wang Xingxing said.
On other product lines, Yuzhu Technology recently released the A2 robotic dog, weighing about 37 kilograms, with a continuous load capacity of 30 kilograms and an empty load range of 20 kilometers. At the same time, Yuzhu Technology emphasizes the development of robotic upper limbs and hands. Wang Xingxing introduced: “We have independently developed a dexterous hand with about 20 degrees of freedom, aiming to enable robots to perform daily tasks, not just showcase actions. We hope to achieve natural interaction in the next one to two years, such as instructing a robot to pour water for someone without prior adaptation.”
Wang He approached the topic from the underlying logic of embodied intelligence. He pointed out that general robots will become a key and revolutionary product in the next multi-trillion dollar market. This revolutionary product includes several core elements: the robot body itself, and the embodied intelligent model that drives its operation. Behind the model is data support. Wang He stated: “We, along with NVIDIA, believe that synthetic data is key to rapidly implementing embodied intelligence. Currently, real-world data accounts for only 1% of our training data, with the remaining 99% being synthetic data.”
Meanwhile, Galaxy General and NVIDIA announced the next-generation humanoid robot project based on the Isaac platform. “Currently, Galaxy General is in a wheeled form, while the next generation will adopt a purely bipedal design, using OpenWBT_Isaac for data collection and remote control. This platform can train and deploy various task capabilities, such as pushing carts and picking up objects from the ground, in both simulated and real environments,” Wang He said.
He believes that in the next decade, the overall share of humanoid robots will not be small. “I expect that the output value of humanoid robots will multiply by ten every three years. So if we sell 1,000 units now, in three years it will be 10,000, and in another three years it will be 100,000. If each unit sells for several hundred thousand, it will reach 100 billion, exceeding the total output value of all industrial robotic arms.”
Wang He further stated: “In the next ten years, we will see a robot market that can surpass the current volume of all industrial robots. In another ten years, it may exceed the trillion-dollar market of automobiles and smartphones, so we cannot underestimate it, but it won’t happen as quickly as everyone thinks.”
From the on-site communication, to truly bring robots out of the laboratory and into large-scale deployment, it requires both a top-level computing and simulation capability to build a technical foundation, as well as cost-effective and mass-producible hardware engineering, along with a large-scale training system driven by synthetic data. As Wang Xingxing said, “AI and robotic technology will elevate human civilization to new heights, just like the invention of electricity and the steam engine.”
The following is the full media Q&A with NVIDIA’s Omniverse and Simulation Technology Vice President Rev Lebaredian, Yuzhu Technology Founder Wang Xingxing, and Galaxy General Founder and CTO Wang He (organized by the reporter without altering the original meaning):
1. In high-precision application scenarios like healthcare and elderly care, how does NVIDIA Omniverse ensure that robots trained with simulation data are reliable and safe?
Rev: If you want to build a robot system that can act in the real world and be safe and reliable, the only choice is to use simulation. For example, you want the system to be smart enough to handle rare special situations, even emergencies that you hope will never happen.
Take autonomous driving as an example; we absolutely do not want a car to hit a person or a child, but when a child appears on the street, how should the vehicle’s brain react, and how do you obtain training data for such systems? The only way is through simulation. Because we can never place a child in front of a car as a training sample; that would be unethical, dangerous, and time-consuming and expensive.
Even after training the system, it still needs to be tested in those same scenarios before being deployed in the real world to ensure it can react correctly when similar situations actually occur. Therefore, the real choice is to achieve this through simulation, as real-world testing is too slow, costly, and dangerous; we do not want robots to fail in the real world first.
In other words, if we cannot make the simulation accurate enough to test the robots, we will not be able to manufacture reliable robots. Fortunately, there are already very accurate simulators available. However, these simulators require significant computational resources and are costly, and the real challenge we face is how to improve simulation speed to make it cost-effective in large-scale system construction, which is what we have been striving for.
2. What key technological trends do you think will drive the simulation field for AI robots in the coming years, and can you share some cases of Chinese ecosystem partners using NVIDIA simulation technology to accelerate product innovation?
Rev: I believe the biggest trend right now is that all the technologies and developments that have emerged in conventional AI are being applied to physical AI. The most important breakthrough is the improvement in reasoning capabilities. For example, DeepSeek has brought reasoning capabilities into the open-source domain, and now we are seeing various other models.
We have made leaps in intelligence levels, applying the same technology to physical AI, which is a significant breakthrough in the robotics field. If we want to create such robots—like a robot that can fetch you water—based on translation understanding, we may see our robots interacting more naturally with people and completing complex multi-step tasks by the end of this year or next. This is a critical capability. Combining this capability with simulation, I believe this is a point that may not yet be widely understood but will become a significant breakthrough.
Currently, the main problem we face is that AI is extremely data-dependent, and obtaining suitable data is very challenging. Our existing reasoning models, especially for reasoning about the physical world, can now help us improve the data generation and creation process.
Today, the data we generate, even synthetic data, requires a lot of human involvement, with humans constructing virtual worlds and simulation environments, determining what data to generate to make intelligent systems smarter. However, if we apply the AI technologies we are developing to the data generation process, we can automate it, creating an “autonomous driving” synthetic data generation.
If we have an autonomous driving synthetic data generation factory, we can directly integrate it into the training process, automating the training process, reducing human intervention, and making the robot brain smarter. As for how Chinese companies are applying simulation technology and its impact, I believe they are actively exploring and applying these technologies.
Wang He: I think because everyone’s simulation engines and parallel renderers have greatly reduced the difficulty of generating synthetic data, whether through reinforcement learning or generating data and then doing imitation learning, the entire difficulty has significantly decreased. Indeed, whether it is the walking or dancing skills of humanoid robots, or the data behind tasks like grasping, folding clothes, and navigation, it relies heavily on a very good simulator and a highly parallel renderer. So we are very grateful to NVIDIA as an ecosystem partner for supporting the entire ecosystem from chips to the entire simulation platform.
3. With the Yuzhu R1 priced at 39,900 RMB, significantly lowering the threshold for consumer humanoid robots, does this mean that hardware costs are no longer a barrier to the commercialization of robots? What challenges remain in promoting the large-scale commercial use of humanoid robots at this stage?
Wang Xingxing: This is a viewpoint I have always shared. For grassroots commercialization, including the commercialization of humanoid robots, cost and hardware are not the key issues. In a sense, if a machine can be used for 100,000 or even 1 million RMB, there are still many scenarios where it can be utilized.
The biggest problem currently is that the entire embodied intelligence model is still not versatile enough, and its practicality still needs significant improvement. This is actually the most challenging issue at present. Hardware, to some extent, has been sufficient for the past year or two; it has always been adequate. Of course, to do better, you need to improve reliability and cost, but it is not a limiting factor.
4. There is ongoing debate among domestic and foreign experts and companies about the necessity of humanoid robots. Some believe humanoid robots are the ultimate carrier of AGI, while others think that in the next decade, humanoid robots will only account for 10%. How does Galaxy General view the necessity of humanoid robots?
Wang He: Looking at humanoid robots today, in the long run, they must be able to integrate into our human lives. From an endgame perspective, for example, the seats everyone is currently making, besides humanoid robots, if they can also work, reach heights of 1 or 2 meters, and touch the ground, they can only move in such environments; there is no other form besides humanoid.
So, looking ahead to the next few years, I feel that humanoid robots are continuously advancing from a mobile composite robot towards the endgame. Because if it is a fixed-point robot, the tasks it can perform are limited to what is in front of it, right? Therefore, its limitations are significant. So mobility is inevitable. Just having a mobile vehicle, in various factories, these mobile carts can only carry goods, right? How do the goods come down? They fall from a slot above; their limitations are that they cannot perform any operations. Therefore, a robot that can move and has a mechanical arm to operate is what we are building today, on a mobile platform, making it elevatable and foldable. With two mechanical arms, because one hand cannot carry a box, it can only grasp one thing, right? Two hands can look up, down, left, and right, so it is essentially a semi-circular shape.
Thus, I believe that in the next decade, the overall share of humanoid robots will not be small, but it depends on who you compare it to. Today, compared to all industrial robotic arms, the total output value of industrial robotic arms globally is only about 100 billion RMB, which is not high. A leading car manufacturer can sell vehicles worth 100 billion RMB in a year. Therefore, if humanoid robots are currently priced at 100,000 RMB each, they are not reaching a level that is perceived as high; they will exceed the total output value of all industrial robotic arms. I expect that the output value of humanoid robots will multiply by ten every three years. So if we sell 1,000 units now, in three years it will be 10,000, and in another three years it will be 100,000. If each unit sells for several hundred thousand, it will reach 100 billion, exceeding the total output value of all industrial robotic arms.
In the next ten years, we will see a robot market that can surpass the current volume of all industrial robots. In another ten years, it may exceed the trillion-dollar market of automobiles and smartphones, so we cannot underestimate it, but it won’t happen as quickly as everyone thinks; it is impossible to reach the market size of automobiles overnight.
5. How does NVIDIA Jetson Thor differ from previous Jetson platforms, and how is it particularly beneficial for robotic applications?
Rev: With each generation of Jetson products, we strive to maximize their computational capabilities because the problems of intelligence are inherently complex computational challenges. In the field of robotics, this challenge is even greater: computation must be extremely fast and performed in very demanding environments.
Robots often operate in real-time on-site, requiring computations to be completed within tight loops, with limited power, thus needing to minimize power consumption to extend battery life while also considering heat management issues. All these factors significantly increase the difficulty. The biggest difference with Jetson Thor compared to previous versions is that it now has sufficient computational power to run larger and more powerful neural networks and models, supporting more complex reasoning tasks, which previous products could not achieve. Additionally, Jetson Thor has higher bandwidth, allowing it to process large amounts of information from various sensors more quickly, enabling robots to react rapidly and operate at high speeds in dynamically changing environments.
6. In which scenarios do you think robots will see large-scale adoption in the future?
Wang Xingxing: The future will definitely move towards more practical directions, but the specific speed will take time. Whether in industrial, service, or household fields, the entire development cycle is still quite long. Looking back, for example, with new energy vehicles, many people thought development would be rapid over a decade ago, but the overall maturity also took a lot of time.
Every industry requires a long maturation cycle. The new generation of human-robot or general robot technology has actually only developed for about two to three years. Because the new technologies used now are completely different from those of ten or twenty years ago, both hardware and software have undergone tremendous changes. However, many people like to refer to robots from ten or twenty years ago, thinking that this industry has been developing for a long time, but the technology back then is not at the same level as today.
Given that the current development time is only two to three years, overall progress will still require more time. However, based on the current development speed, I personally feel that the industry is still growing rapidly, and in the coming years, personnel and shipment volumes are likely to double each year, which is very achievable for the entire industry.
Based on this, if a more powerful and general AI large model emerges, allowing robots to perform better in factories, homes, and other more general scenarios, the more general they are, the easier it will be to popularize. Conversely, if they are not general, promotion will be more difficult. Therefore, I believe the overall time cycle will be longer, especially in the household field.
The biggest challenge for household robots is not technology, but the high requirements for ethics and safety, which makes the threshold for household robots much higher.
7. This year, many humanoid robots have begun training in automotive factories. How long will it take for robots to truly be deployed in factory workshops? What key challenges still need to be addressed?
Wang He: This year, many humanoid robots have already started training in automotive factories. We see that the vast majority of companies promoting humanoid robots in factories are mainly focused on two aspects: one is handling, and the other is sorting.
In terms of handling, the speed of the robots demonstrated in recent videos by Galaxy General is already close to human levels, with the number of items handled per hour comparable to that of humans. This stage is very close to actual deployment in factories, and I expect that by the end of this year, several dozen Galaxy General robots may enter factory workshops for actual application.
However, handling is just the first step. In addition to handling, it is necessary to achieve closed-loop capabilities for palletizing; only when both handling and palletizing are completed in a closed loop can robots truly fulfill the entire workflow; otherwise, doing half the task is not ideal.
Sorting is a greater challenge. Whether picking from a conveyor belt or retrieving items from shelves, the biggest difficulty currently is speed. Skilled workers can pick items very quickly, but robots still struggle to achieve this efficiency at the model and hardware levels.
When we do retail robots, picking items from shelves or tables is technically similar to industrial sorting, but retail has lower rhythm requirements, and the consequences of picking the wrong item are lighter. However, in industrial scenarios, such as automotive manufacturing, a production line stopping for a minute could mean a loss of tens of thousands of dollars, so the precision and speed requirements for sorting are extremely high.
In summary, while sorting technology has made significant progress, it has not yet reached the level of human workers and still requires some time for technological iteration and breakthroughs.
8. NVIDIA emphasizes a simulation-first strategy in robot training and has launched a series of supporting technologies. However, challenges still exist in bridging the gap between simulation and reality (Sim2Real). How does NVIDIA work with partners to address this issue? Looking ahead, what are the key directions for enhancing the physical realism of simulations and improving the efficiency of transferring to the real world?
Rev: This is a very good question. If we rely on simulation to build and test AI, we must ensure that the simulation is as close to reality as possible; otherwise, we cannot trust it. The AI we build cannot truly understand the real world if it is trained in a “cartoon world”; therefore, testing must also ensure that the simulated scenarios match reality. So how do we bridge the gap between simulation and reality? There are actually multiple ways, and we are fully promoting these methods.
First, we need to improve the accuracy of the simulator itself. We have been building physical simulation algorithms for decades, and these algorithms have been validated to reflect the physical laws of the real world well. For example, we use simulations to design airplane wings and cars, ensuring aerodynamic performance and verifying that simulation results match the real world. The problem is that these high-precision simulations are extremely costly, often requiring hours of computation on large computers. The challenge lies in how to increase simulation speed to a level that can be embedded in AI training processes, enabling large-scale and efficient data generation and testing.
To this end, we are leveraging AI itself as a tool to enhance simulation speed and accuracy. AI can approximate any mathematical function, and we can convert physical simulation functions into AI functions, building AI simulators to complete simulations. As long as we provide enough example data, AI can learn the simulation functions. This is precisely what we are developing with the “Cosmos” project. These “world foundation models” are AI models that understand the physical laws of the world, and we can input real-world data and credible simulation data into these models for training. Once we have such AI foundation models that understand the world, we can combine them with traditional simulations to build more accurate and efficient simulators.
Secondly, even with high-quality simulators, constructing data that represents the real world is very challenging. For example, in this room, while the simulator can simulate physical phenomena, we also need to create tablecloths and tables with the correct physical parameters (such as friction coefficients and material properties), and this information collection is very complex. Currently, only a few professionals globally—usually artists from the gaming or film industry—have this capability. However, as we build AI with physical understanding capabilities, these AIs can assist in generating these virtual environments, becoming “robot artists” that help us efficiently create realistic virtual worlds.
The third method is to directly capture the real world. We also use physical AI technology to digitize real environments (such as the room we are in) and import them into simulation environments, ensuring that virtual scenes are highly consistent with reality.
NVIDIA is comprehensively advancing in these three directions, developing related technologies, but this work far exceeds the capabilities of any single company. We are collaborating with partners across the entire ecosystem to push forward on these three paths. In fact, we have already achieved significant results; our existing simulators can generate sufficiently high-quality data to help us enhance AI performance.
9. Some experts currently believe that the large model architecture in the robotics field has not yet been unified. What directions is Yuzhu Technology focusing on for the foundational model of embodied intelligent brains? Can you share some specific content?
Wang Xingxing: I have always felt that the current model architecture is indeed very ununified, which leads to the overall progress not being as fast. If the model architecture could be more unified and the direction clearer, combined with the current industry heat, breakthroughs could be achieved more quickly. But the reality is that progress is still relatively slow.
Our company has explored many directions. For example, this morning we also showcased a project we attempted last year using video generation models as “world models” to drive and align robotic arms, which achieved some results. However, due to the massive scale of training required for video generation models, considering our company’s computing power and investment, it is challenging to conduct large-scale training.
Moreover, we found that the versatility of these models does not fully meet expectations, so we basically did not continue using them. However, recently Google released a new video generation model that has excellent physical alignment effects, and they publicly attempted to use video generation models as world models directly for robotic arms and general intelligence. This makes me feel that this direction is worth exploring again.
Due to the limitations of our company’s scale and computing talent, we are only exploring this direction preliminarily and have not deeply advanced it. However, Google’s results prove that this direction has great potential. Video generation models have reached good expectations in terms of temporal content, data sources, and effects. For example, if we control a video generation model to generate a video of a robot cleaning an entire house, and the effect is good, theoretically, as long as we align the video with the robot’s actions, we can achieve similar results.
However, the alignment work is still very complex and challenging. This direction is mainstream and worth investing in, both for robotic applications and for pure video generation technology itself, which will continue to receive increased investment and optimization from major companies.
In addition, there are other solutions. As the capabilities of foundational models rapidly improve, many potentials have yet to be fully explored. We have found that if we incorporate robotic instruction control and spatial understanding training into the training of foundational models, the effects can be significantly enhanced. For example, some of the robot control effects demonstrated by Wang He’s team based on foundational models are quite impressive.
Our company’s strategy is very simple: to continuously try various new models and new ideas. Today there may be one idea, and tomorrow it may be adjusted; this is normal. For emerging technologies, I believe everyone should boldly try. The AI field is full of possibilities, and often a flash of inspiration can lead to breakthroughs. I hope to encourage more people to explore; perhaps the next innovation will come from you.
10. We have seen that OpenAI recently released GPT-5, which gives the impression that the technological breakthroughs are not significant, and it seems more like a system rather than a single model. Can we understand this as the scaling law of large models encountering some challenges?
Wang He: Currently, there are many types of large models, including pure text large models, multimodal large models, and multimodal large models that are further divided into visual understanding and video generation types, including our embodied intelligence VLA, which is also a type of large model. Therefore, whether we can say that the expansion of large models has encountered bottlenecks cannot be simply summarized with a unified conclusion.
I understand that at the current pure text stage, our main data source is publicly available data on the internet, but much private knowledge is not online, which leads to a discrepancy between the data needed for the model’s reasoning capabilities and the publicly available data. Unless we can obtain additional data enhancement through some controllable means, relying solely on public data to enhance model capabilities is limited.
This part of capability growth will not naturally occur simply by expanding the model scale. However, we cannot underestimate the progress of reasoning models; for example, in the IMO International Mathematical Olympiad, the text model won a gold medal, demonstrating excellent performance on previously unseen problems, indicating that the capabilities of text large models are continuously improving.
Regarding multimodal large models (such as VLM and VLA), if we say that text models are now quite strong, then multimodal models are still slightly weaker than language models. The core reason is the lack of data: text data is very rich, while text-image paired data is relatively scarce, and action data is even less, so there is still a significant gap in visual understanding capabilities and action operation capabilities based on vision.
This is why synthetic data and simulation technology are very important. As Rev mentioned, simulation can reproduce real-world scenes and actions in virtual environments, generating a large amount of data with actions, images, and semantic pairs, which will greatly promote the development of text-image, multimodal large models, and embodied intelligence large models.
If we rely entirely on real data, progress will be significantly limited. Overall, I believe that fully utilizing simulation technology will be the most effective way for multimodal large models and embodied large models to address data bottlenecks.
11. What are the key technical bottlenecks currently limiting the large-scale deployment of humanoid robots?
Wang He: The core issue is actually quite simple—the ability of robots to perform tasks is still not strong enough, and the types of tasks they can complete are relatively limited. However, if we can achieve a very general level of performance within these limited skill ranges, it will empower many scenarios.
Galaxy General’s main breakthroughs now are in “grasping” and “mobility.” As long as robots can grasp any object, achieve lower limb movement and upper limb extension in a scene, and finally accurately place objects,
these three capabilities will enable many application scenarios. Behind this requires a truly precise target recognition and positioning system, which we are currently promoting through synthetic data.
Of course, even if this key issue is resolved, there are still many tasks that robots cannot currently complete. However, as long as the issues of target recognition and positioning can be tackled, the humanoid robot market will have at least a trillion-level scale, and visible results can be expected within five years. Once this key technical bottleneck is resolved, based on such a huge market investment, robots will inevitably unlock more skills and take steps towards a trillion-dollar market.
12. Why should robots be bipedal? Besides emotional value, what other commercial considerations are there for bipedalism?
Wang Xingxing:In a sense, I have previously mentioned that questioning why not to make bipedal robots is worth considering. Because making bipedal robots is relatively convenient, and most importantly, bipedalism provides more general capabilities. The ability to move is, to some extent, a weaker AI capability. You see, small animals, even ants and insects, walk very well, but their AI capabilities are actually quite weak. Therefore, I have always believed that a truly general and capable embodied AI model will have movement ability or leg capability as a subsidiary.
If robots can work, then controlling legs will naturally not be a problem; if they cannot even control their legs, it indicates that they have not reached the stage of a very general AI model as imagined. So this is a development direction. Additionally, because bipedalism is relatively simple, despite still having challenges, our company itself focuses on legs, so for us, this is a natural and interesting thing. People generally also like this direction. Moreover, if everyone makes wheeled chassis, it will lead to homogenized competition, which is unnecessary. Our company focuses on legs, hoping to enhance the overall movement and working capabilities of robots, which is a very good direction. I have also worked on wheeled chassis, and I believe there is a difference between wheeled and legged robots, and this will change over time. Currently, wheeled chassis are very stable and energy-efficient in open industrial scenarios and can shuttle between shelves, but they may not pass through complex environments. If the chassis is made smaller, stability will be lost. Therefore, at different points in time, the lower body solutions for robots will definitely differ. I firmly believe that legs are the future because they can reach all accessible spaces of the upper body and flexibly mobilize the waist’s flexibility. However, at different stages, there will be the most suitable forms for practical applications, and we will not limit ourselves to a single solution. We are simultaneously conducting research on both wheeled chassis and humanoid robots for lower body and even full-body control.
13. We know that physical AI, especially in the robotics field, has high requirements for energy consumption, thermal management, and size limitations. How is NVIDIA addressing these challenges? How will future computing platforms meet these needs?
Rev: Looking back in history, NVIDIA has significantly improved performance per watt and performance per dollar with each generation of products. In the past, we had Moore’s Law, which meant that computing power would grow exponentially—at its best, performance could increase tenfold every five years and a hundredfold in ten years. However, relying solely on Moore’s Law is no longer sufficient to solve many of the problems we face. We foresee that the utility of Moore’s Law in CPUs and general-purpose computers will gradually end.
Therefore, we are committed to building specialized computers for specific algorithms. This specialized computer requires not only optimization at the chip level but also at the algorithm, software, and application levels to achieve maximum performance. This cannot be achieved by a single factor, such as making chips smaller or faster, but through full-stack optimization. This is a very challenging engineering task and is precisely NVIDIA’s core competitiveness.
We initially applied this approach to computer graphics rendering (especially in the gaming field), and then expanded it to other areas. After the launch of CUDA, we began applying it to physical simulation, and later deep learning and AI emerged on GPUs, leading us to continuously specialize processors. With each generation of products, we have achieved significant performance leaps at the same power consumption and cost, and this will continue in the future because our path of innovation is far from over.
14. Compared to other countries, how do you view the demand and challenges of AI in China in practice?
Rev: China is both an important market and a production base for AI technology and products. China has a large number of intelligent, well-educated, and passionate AI researchers and developers, with nearly half of the world’s top AI talent concentrated here, along with top AI universities.
In the field of physical AI and robotics, China has unique scale advantages, combined with talent advantages, forming a unique ecosystem. China has deep professional capabilities in manufacturing the key components required for electronic hardware and robots, and such an ecosystem and manufacturing scale are difficult for other countries to match. This enables companies like Galaxy General and Yuzhu Technology to manufacture robots on a large scale and learn and iterate quickly. China’s unique comprehensive conditions provide a solid foundation for the rapid development of the physical AI and robotics industry.
Copyright Statement:Some articles were unable to contact the original author during the push.If there are copyright issues, please contact the original author.
