Superknoten werden zum Schwerpunkt der WAIC. In Zukunft könnte die Ersetzungsrate heimischer GPUs über 80 % liegen.
In front of the booths of MuXi Technology, a series of supernode computing power cabinets with a dark blue sheen stand out particularly. The cabinets stand there like iron giants, and the label indicating 128 GPUs per cabinet underlines their enormous computing power. The pure liquid - cooling concept allows them to get rid of the complicated air - cooling devices of the traditional solution, which makes visitors stop.
"What you see here is actually a cluster with thousands of cards!", said an employee at the MuXi Technology booth to a reporter from IT Times. At the exhibition, it's about MuXi's latest computing power technology - the supernode computing power cabinets. Each cabinet has 128 GPUs, and ten cabinets together have a total of 1280 GPUs. The supernode computing power cabinets abandon the traditional air - cooling devices and are only equipped with liquid - cooling devices. The energy consumption of the entire cabinet is more energy - efficient compared to the previous single - cabinet groups.
MuXi is not the only one. In the exhibition hall of WAIC 2025, there are a plethora of computing power exhibits. Super Fusion has presented the world's first plug - and - play supercluster system for diverse intelligent computing, which has 128 AI - acceleration cards per cabinet. In front of the booth of Suyuan Technology for the DeepSeek All - in - One System, there are many visitors. This product was only introduced at the beginning of 2025 and supports the platform of domestic CPUs as well as the ability to optimize in different scenarios.
The concentrated release of supernode computing power by many companies is similar to a "collective assault" in the innovation of domestic computing power. With the upsurge of DeepSeek at the beginning of the year and the start of the era of large models, the demand for computing power has skyrocketed, forcing the domestic computing power chip industry to accelerate innovation. The supernode is the key tool in this battle. However, despite the wave of innovation, the industry remains highly alert to potential risks.
Innovation: The supernode computing power kicks off the breakthrough battle
"A supernode can be understood as a large cluster. In the past, a server was regarded as a node. Now, multiple nodes are 'intertwined into a strand' in a supernode. In our high - density supernode cabinets, one cabinet integrates multiple servers to achieve a computing power concentration of 128 cards or even higher density.", said an employee at the MuXi booth. The supernode computing power is essentially a cluster - based solution for the extreme computing power demand in the era of large models. "In the past, multiple cabinets were installed separately. Now, the computing power is concentrated in the supernode. This not only saves the rent for the cabinets and electricity costs but also does not require air - cooling technology. Liquid cooling can handle the cooling, which meets the requirements of cost - reduction and efficiency - improvement in the era of large models."
Put simply, it integrates multiple servers and multiple computing power chips into a larger and more powerful overall computing power unit to break through the limits of traditional computing power facilities. This concentration is not simply a stacking of hardware but improves the overall efficiency, communication bandwidth, and space - utilization efficiency of the cluster through the optimization of the internal networking technology of the nodes and liquid cooling.
At a forum titled "Core of computing power, opening a new phase, source of AI, shared future", Peng Li, co - founder, CTO, and leading hardware architect of MuXi, introduced the flagship GPU Xiyun C600. The Xiyun C600 integrates large storage capacity and mixed computing power with multiple accuracies, supports the MetaXLink supernode expansion technology, and has an integrated ECC/RAS multiple - security protection module to meet the training and inference requirements of the next - generation generative AI systems. "MuXi has once again made an important leap from 'R & D advantage' to 'product advantage'.", said Peng Li.
At the WAIC venue, the technical features and innovation value of the supernode are clearly visible.
The Cloud Suis ESL supernode system of Suyuan Technology can achieve a full - bandwidth connection of up to 64 cards at a single node and uses a liquid - cooling concept to achieve "high cost - efficiency, high density, and high energy - efficiency". An employee at the booth explained to a reporter from IT Times: "The number of parameters in large models has jumped from hundreds of billions to trillions. The traditional computing power facilities cannot handle the node communication, space requirements, and energy consumption. The supernode optimizes the internal networking to make the computing power cluster 'run faster and use resources more economically'."
Super Fusion has presented the world's first plug - and - play supercluster system for diverse intelligent computing, which has 128 AI - acceleration cards per cabinet, a high - speed connection of 112G/224G, and a super power supply of 240 kW per cabinet. One cabinet replaces 32 cabinets and uses the fifth - generation 100% native liquid cooling, which enables an energy saving of over 20%. The software and hardware infrastructure technology and the computing power ecosystem created by Super Fusion can be compatible with over 10 AI - acceleration cards and provide a secure heterogeneous computing power in a dual north - south ecosystem.
The concentrated release of supernodes is due to the computing power transformation driven by the development of large models. In traditional computing power facilities, there are problems such as communication bottlenecks between nodes, high space requirements, and high energy consumption, which cannot meet the training and inference requirements of large models. As is generally recognized in the industry, the larger the number of parameters in a model, the more cards are needed. The supernode "bundles" the distributed computing power, reduces the space requirements for the cabinets, and improves the communication efficiency. It is the inevitable choice to support the development of large models.
Bottlenecks: Where does domestic computing power still lag?
In the era of the booming artificial intelligence sector, the importance of computing power as the core driving force is undeniable. However, domestic computing power faces numerous obstacles on its way, and the current development situation is worrying.
Looking at the hardware foundation, the process - generation gap is a painful point for domestic computing power chips. Meng Fanchi, general manager of the digital infrastructure business group of Taiji Corporation, told IT Times: "The main processes of domestic unicorn GPU companies are concentrated on 7/12/14 nanometers. Compared with the internationally leading 3/4 - nanometer processes, there is a significant generation gap, which can lead to a performance gap of up to 10 times. This means that when processing the same complex tasks, domestic computing power chips lag far behind the international technological level in terms of key indicators such as speed and efficiency, which severely restricts their use in high - end application areas."
In the field of computing power networks, domestic computing power is also in the initial stage. Several employees at the GPU company booths also admitted to a reporter from IT Times that the development of domestic GPUs is still restricted by the complicated processes of lithography machines and further breakthroughs are required.
The standards for data volume, interface closure, and other aspects are still in the design stage. The interconnectivity between different devices is low, and problems such as latency, bandwidth bottlenecks, and network congestion often occur, which seriously affects the efficient use of computing power resources. Even if there is an advanced intelligent algorithm system, due to the insufficient allocation of computing power, when dealing with massive business resources and complex tasks, there will be low efficiency and uneven resource distribution, so that the potential of domestic computing power cannot be fully exploited.
The cooling problem also hinders the development of domestic computing power. With the continuous increase in the power consumption of GPU servers, the traditional air - cooling method is increasingly unable to meet the requirements. Although liquid - cooling technology is regarded as a solution, its maturity still needs to be improved and the costs need to be reduced.
In view of the difficulties in procuring general - purpose GPUs, Yang Gongyifan, founder and CEO of Zhonghao Xinying, believes that "multiple chips and multiple technological paths" are the inevitable choice for the Chinese AI chip industry. "If the market volume is large enough, the cost - efficiency of customized chip development will far exceed that of general - purpose components, at least 2 - 3 times, even 5 - 10 times. This will significantly reduce the production costs of applications such as large models."
He cited the development of large models as an example: "Last year, the general view in the industry was that a 7B - parameter model would be sufficient. This year, it has already been fully updated to 70B. In view of the future trend of increasing parameters, only a special architecture that supports efficient interconnectivity of multiple chips can bear the load."
"The current computing power market is in a state of 'card - hunger'.", said Meng Fanchi to IT Times. Large Internet companies are currently in a race for computing power: Alibaba plans to invest 380 billion yuan in two years, and Tencent and ByteDance are also spending large sums to gear up. However, Chinese companies are hindered in procuring high - end NVIDIA GPUs due to the US chip ban, and the supply chain of domestic chips is not yet fully smooth. "Currently, large companies lack a supply of computing power, while the real demand continues to rise."
According to Meng Fanchi's observations, the data industry has been in a continuous decline since 2021 and reached its lowest point in 2023. However, the demand for intelligent computing power centers has increased rapidly. Taiji Corporation, as the builder of domestic projects such as the Beijing administrative cloud, transfers its experience in the administrative cloud to the field of computing power services. "In the past, we used CPU servers to build a network for SaaS services. Now, we use computing power servers to form a computing power system." He admitted that since last year, many orders for computing power have been continuously reduced. It was not that the customers' demand decreased, but the availability of computing power cards. Among the end - customers we are currently serving, whether they are domestic or imported chips, there is mainly a shortage of chips. However, the demand is still very high.
Future: The replacement rate of domestic GPUs could exceed 80%
In the era of the rapid expansion of global AI computing power demand with a 750 - fold increase every two years, the development trend of domestic GPUs is particularly observed.
Guo Yike, a member of the British Royal Academy of Engineering Sciences, said at a MuXi forum that the market volume of Chinese cloud - AI chips will exceed 4.8 billion US dollars by 2027, and the replacement rate of domestic GPUs is expected to exceed 80%. This paints a promising picture for the future of the domestic GPU industry.
Technological innovation is the core driving force for the development of domestic GPUs. Although there is currently a generation gap between the main processes of domestic unicorn GPU companies and the international leading level, the companies have never stopped chasing. Many companies continuously invest in the research and development of process technologies and chip architectures, and some companies have already achieved preliminary results and gradually improved the product performance. At the same time, customized chip development emerges and has the chance to secure a place in the market with higher cost - efficiency and open up a new path for the technological development of domestic GPUs.
In the field of ecosystem building, the cooperation between companies is getting closer. For example, Jieyue Xingchen, together with nearly 10 chip and infrastructure manufacturers, has launched the "Model - Chip Ecosystem Innovation Alliance" to promote the penetration of the entire technology chain from "Chip - Model - Platform".
In addition, the standardization work is continuously promoted. At the forum, the National Institute of Electronic Technology Standardization has established the national standard test - benchmark system for artificial intelligence "Qiushuo 2.0", which provides a national technical guideline for the development of domestic chips and other aspects and promotes the creation of a domestic computing power system with "security, cost - reduction..."