Shenzhen system surpasses “El Capitan” in TOP500 rankings, shows CPU-only path to exascale computing, distinct from AI race

The world ranking of supercomputers returns to being a technical, industrial and geopolitical indicator. With LineShine, installed at National Supercomputing Center in Shenzhen, the China has regained first place in the rankings TOP500, bypassing the US system El CapitanThe news, also reported by Nature, concerns a machine capable of over two quintillion operations per second on the traditional benchmark used to measure the power of high-performance systems.

The central result is the one obtained on the test High Performance Linpack, HPL: 2,198 exaflops per second, equal to 2.198,40 petaflops per second in the official system sheet. TOP500 also indicates a theoretical peak of 2.735,82 petaflops per second, 13.789.440 core and an electricity consumption of 42.220 kilowattsThese numbers place LineShine ahead of El Capitan and confirm that exascale computing is entering a more distributed phase, with top-of-the-line machines in Asia, North America, and Europe.

The most observed peculiarity is not only the primacy. LineShine is based on a solution CPU-only, that is, without dedicated graphics accelerators. At a time when high-performance computing andartificial intelligence and big data are often associated with GPUs, tensor cores, and specialized chips for training models, the Chinese machine follows a different design choice. The system uses the base Ling Kun, processors LX2 a 304 core, proprietary interconnection Ling Qi and operating system Kylin OS.

An exascale machine based on general purpose processors

The Shenzhen supercomputer wasn't just a collection of computing nodes. According to the official specifications, the machine was built by Shenzhen Cloud Computing Center and installed in the 2025The declared configuration includes processors LX2 304C at 1,55 gigahertz, LingQi network, compiler Lclang 1.0.0, mathematical library OpenBLAS-local and a version of OpenMPI adapted to the interconnection of the system.

The choice to focus on very high density CPUs, rather than accelerators, is relevant for the research and development Because it shifts attention to a different balance between parallelism, memory, networking, and software. In accelerated systems, much of the performance depends on the ability to rapidly transfer data to specialized units. In a CPU-only machine, efficiency instead derives from the scale of the overall design: the number of cores, memory hierarchy, communication between nodes, and the ability of the scientific software to leverage millions of processing units.

The TOP500 release highlights that LineShine is the first system on the list to surpass the two sustained double-precision exaflops Using only CPUs. This is an important technical fact, because double precision remains essential in many scientific simulations, from engineering to computational physics. However, it should not be confused with the metrics that best describe the training of large AI models, where reduced-precision calculations, dedicated accelerators, and rapid communication between clusters are the most important factors.

Even the reading proposed by Xinhua It emphasizes the convergence of supercomputing and intelligent computing. According to the Chinese agency, LineShine incorporates matrix acceleration units directly into its Chinese-designed processors, aiming to reduce the data transfer bottleneck typical of CPU-GPU architectures. This explanation is consistent with the stated technical strategy, although true competitiveness on AI workloads must be assessed with specific benchmarks and not just by HPL primacy.

Why the TOP500 ranking alone doesn't measure AI power

The distinction between high performance computing and AI-oriented systems is crucial. The TOP500 primarily measures a machine's ability to perform double-precision numerical computations. This remains a central metric for scientific simulations, climate modeling, computational chemistry, fluid dynamics, and engineering design. However, it does not match the ability to train large language models or generative systems, which depend on architectures, memory, networks, and software optimized for mixed precision and tensor loading.

This is confirmed by another piece of data from the same ranking. On the benchmark HPL-MxP, closest to mixed precision loads, LineShine ranks fourth with 7,92 exaflops per second and an acceleration of 3,6 times higher compared to traditional HPL. El Capitan, however, remains first in this test, followed by Aurora and Frontier. The reading is clear: the Chinese machine dominates CPU-only numerical calculations, but it is not the strongest system in terms of meeting some of the needs of contemporary AI.

The same caution emerges in the analysis of Reuters, which reminds us that many large systems developed by Microsoft, Amazon, Google, or xAI do not participate in the TOP500 ranking. These are often private infrastructures designed for model training and inference, not to compete publicly on the HPL benchmark. Therefore, LineShine's result measures real-world performance, but within a defined technical framework.

“If the hyperscalers were to present their systems, this 'world's fastest' would not make the top five,”

said Jimmy Goodrich, senior fellow at theInstitute for Global Conflict and Cooperationn of the University of California, to Reuters

The sentence of Jimmy Goodrich It helps define the significance of this primacy. LineShine is currently the most powerful system among those presented in the TOP500 list on the HPL test. This does not automatically mean it is the largest global AI infrastructure. The comparison with hyperscalers is relevant precisely because a growing portion of the world's computational capacity is no longer concentrated in just national or university centers, but in private clusters optimized for generative models, mixed precision, and tensor loading.

For businesses, research centers, and public policymakers, the difference isn't academic. Purchasing, designing, or using computing power requires connecting the architecture to the problem being solved. Scientific computing requires numerical accuracy, stability, memory, scalability, and the ability to run many parallel simulations. deep learning industrial instead weighs low-precision throughput, accelerator availability, software stack maturity and data management.

Chips, technological sovereignty and China's return to the list

The technical data is intertwined with a political dimension. Reuters highlights that China had stopped submitting its rankings in 2023, after years of US controls on the export of advanced chips and high-performance computing-related equipment. The return of a top-ranked system therefore also serves as an industrial signal: Beijing demonstrates its ability to design and integrate a complex environment comprising processors, networks, operating systems, mathematical libraries, and communications tools.

“What surprises me is that they presented it and want to get recognition for it.”

said Addison Snell, CEO of Intersect360 Research, told Reuters

The statement of Addison Snell It shifts attention from its primacy itself to its public display. The achievement does not demonstrate complete autonomy over the entire semiconductor supply chain, nor does it solve the problem of access to the most advanced AI accelerators. However, it does indicate that China sought to showcase its national capacity in general-purpose supercomputing, at a time when computing infrastructure is increasingly tied to industrial policy, technological security, and scientific competition.

Goodrich, in the same Reuters analysis, also links the reading to export controls, arguing that Beijing wants to present the result as proof of their ineffectiveness. This interpretation should be handled with caution. LineShine is not based on advanced AI chips comparable to the accelerators most in demand for generative models, and its fourth-place ranking in HPL-MxP confirms that mixed-precision performance remains less competitive than accelerated architectures. The machine demonstrates strength in general-purpose supercomputing, not the end of technological dependencies in the AI ​​sector.

For the United States, the overtaking has a symbolic value. El Capitan, installed at Lawrence Livermore National Laboratory, maintains 1,809 exaflops per second on HPL, 11.340.000 core and an energy efficiency of 60,94 gigaflops per wattIt's a machine built on the HPE Cray EX255a architecture with AMD EPYC processors and AMD Instinct MI300A accelerators. The LineShine outperforms it in most respects, but consumes more absolute power and follows a very different design.

Also 'XNUMX-XNUMX business days remains in the game. In the same ranking appears JUPITER Booster, installed at Jülich Supercomputing Center in Germany, indicated by TOP500 as the first European system above the exascale threshold on HPL. In sixth place comes HPC7 by Eni, a sign of the growing presence of large industrial users in advanced computing.Italy, with scientific and industrial infrastructures such as Leonardo and HPC7, therefore remains part of the European geography of supercomputing.

From climate to simulations, where scale produces value

An application example comes from the scientific preprint linked to the Nature article, published on arXivThe work describes CAPES, a hybrid numerical-AI system for seasonal forecasting of summer rainfall in East Asia. The authors point to a computational flow with 174 numerical members e 1.600 AI-generated members, For a total of 1.774 members, used for annual reconstructions since 2016 al 2025.

According to the paper abstract, the entire LineShine machine completes ten annual hindcasts in 14,6 hours and improves the average prediction score by 71,8 a 75,9 compared to the ECMWF reference indicated by the authors. The system works at 15 kilometers resolution for decadal hindcasts and declares a capacity to 1 kilometer for more detailed typhoon simulations. As this is a preprint, the results should be viewed as scientific material under discussion and not as definitive conclusions subjected to formal peer review.

The case is nevertheless useful for understanding why exascale supercomputing retains a specific role even in the era of generative models. High-resolution climate and weather forecasting requires many parallel simulations, data assimilation, coupling between the atmosphere, ocean, and land surface, as well as statistical methods to assess uncertainty. In this context, the hybridization of physical and machine learning algorithm It can reduce calculation times and expand the number of scenarios analyzed.

The potential implications aren't limited to research. More robust forecasts of heavy rainfall, typhoons, or seasonal events can impact dam management, agriculture, insurance, logistics, civil defense, and urban planning. The connection isn't automatic: validation, data access, operational expertise, and institutional capacity to use the information produced are required. But this case demonstrates how a large computational infrastructure can become a tool for data-intensive public and industrial decisions.

Then there is the energy issue. An absorption of 42,22 megawatts This equates to an infrastructure with requirements comparable to those of a large industrial plant. This figure alone does not reveal the environmental impact, which depends on the electricity mix, cooling, utilization rate, and component life cycle. However, it does indicate that the growth of supercomputing requires data center policies, energy planning, and resource management strategies. resource efficiency increasingly integrated.

LineShine's primacy should therefore be interpreted as technical news with broad industrial implications. It marks China's return to the top of the public TOP500 ranking, confirms the expansion of exascale competition, and reminds us that computing power is now a key component of industrial policy. The next phase will be less about the title of "fastest" and more about the ability to transform these machines into verifiable results: more useful climate models, reliable simulations, portable software, sustainable energy infrastructure, and distributed expertise across research, industry, and public administration.

Here are three insights that might interest you:

JUPITER brings Europe into the era of exascale supercomputing
HPC6, the most advanced supercomputer for energy transition
Quantum Bologna: Italy's first hybrid supercomputer is born.