China's Sugon 8000 Puts 100,000 AI Cards to Work on Brain, Universe and Weather Simulations

From four trillion cosmic particles to a digital brain

A simulation using four trillion dark matter particles, a digital brain containing 100 billion neurons and an ocean forecasting system with a resolution of one kilometre are among the applications presented alongside China's Sugon 8000 computing system at a major conference in Shenzhen.

The machine, described by its developer as China's first fully domestically developed AI supercluster containing 100,000 accelerator cards, was showcased at the 22nd CCF National Conference on High Performance Computing. The event ran from October 17 to 20, 2026, at the Shenzhen World Exhibition and Convention Center, with ten frontier application achievements released on October 18.

Sugon 8000, also called Dengfeng, was built in Zhengzhou, Henan province, and connected to the National Supercomputing Internet. Its central design combines high performance computing, which calculates from mathematical and physical rules, with artificial intelligence, which learns patterns from data.

The practical aim is to give researchers a faster way to investigate difficult problems. A conventional simulation can generate training data, AI can search through possible answers, and another detailed simulation can check the most promising results.

Sugon's completion announcement dates the system's completion and network connection to July 10, 2026. The October presentation therefore focused on applications from a system already in use, including work in astronomy, materials, life sciences, weather and agriculture.

What 100,000 connected cards actually do

An AI accelerator card is a piece of computing hardware designed to carry out large numbers of calculations in parallel. Connecting 100,000 cards allows a cluster to divide demanding tasks among many processors, provided they can exchange data quickly enough.

Sugon says the system supports numerical formats ranging from FP64 to INT8. FP64 uses 64 bits to represent a floating point number, giving scientific calculations substantial numerical precision. INT8 represents integers using eight bits and is useful for some AI operations where smaller numbers can reduce memory use and calculation costs.

Different stages of research need different levels of precision. A detailed physical simulation may require FP64, while an AI model screening possible materials may use a lower precision format. Sugon 8000 is designed to support these workloads on the same computing platform.

The move from 10,000 to 100,000 cards is a tenfold increase in card count. Useful performance also depends on how efficiently software divides the work, how quickly data moves and how much time processors spend waiting for one another.

Sugon describes its approach as native integration of high performance computing and AI. In practical terms, the infrastructure is intended to handle scientific simulation, model training and inference, the process of using a trained AI model to produce an answer.

A research loop that checks AI with physics

Scientific simulation and AI start from different places. A simulation might use equations describing how atoms interact or how air moves. It begins with an initial state and calculates what happens next, often through many small steps.

An AI model learns from examples. Given enough simulation results, it can identify patterns and estimate outcomes much faster than repeating every detailed calculation. That speed is valuable when researchers need to examine a huge number of possible designs or molecular arrangements.

Sugon 8000's combined approach follows three stages. First, scientific computing produces simulation data. Second, AI learns from that data and screens candidates. Third, detailed scientific calculations check a smaller selection of promising answers.

The final check matters because AI can produce plausible but inaccurate results, often called hallucinations. In scientific work, an attractive prediction still needs testing against the physical behaviour being studied. The combined workflow directs expensive calculations towards candidates that have already passed an initial screening.

For researchers, the benefit is a change in how computing time can be spent. Instead of running the most detailed calculation on every possibility, they can use AI to narrow the search and reserve greater precision for the cases that deserve closer examination.

The July launch and October showcase

Sugon's July announcement and the October conference presentation describe different stages of the system's rollout. The company announced completion first, followed by operational figures and then the broader scientific showcase.

• July 10, 2026: Sugon announced completion of Sugon 8000 and its connection to the National Supercomputing Internet.
• July 18: Figures released during the World Artificial Intelligence Conference described full utilisation during the first week of operation, with more than 150,000 jobs processed daily on average.
• October 17 to 20: The CCF National Conference on High Performance Computing took place in Shenzhen.
• October 18: Sugon presented ten frontier application achievements at the conference.

Sugon's official completion date is July 10; an October account also described the cluster as formally entering service in August, without explaining how that milestone related to the July launch.

The first week figures put the daily peak above 500,000 jobs, more than three times the stated daily average threshold of 150,000. A job is a submitted computing task, so the total can include work of very different sizes and durations.

The July operational update also described support for millions of concurrent user requests through scheduling that considers task type, data location and resource availability. Such scheduling is the computing equivalent of assigning work to the right teams while keeping shared routes and equipment from becoming bottlenecks.

What the universe and brain projects involve

The Qianyan cosmological simulation used four trillion dark matter particles to model cosmic evolution. Dark matter is matter whose gravitational effects can be observed even though it does not emit light. Simulated particles represent portions of that matter, allowing researchers to calculate how large structures develop.

The numerical scale gives substance to descriptions of the machine as capable of containing a simulated universe. Researchers use a mathematical representation of cosmic matter and its behaviour, then calculate how that representation changes.

A Fudan University team also presented a digital brain at the scale of 100 billion neurons, described at the conference as the world's first at that scale. A digital brain represents nerve cells and their interactions in software so researchers can investigate their collective behaviour.

Both projects demand more than storage space for large numbers of objects. The computer must repeatedly calculate interactions and move information between processors as the simulated system changes.

The astronomy and brain projects show why the October presentation focused on applications rather than card count alone. Their useful measures include how much complexity a model can handle, the detail of its calculations and how long a simulation takes to finish.

Materials, weather, oceans and virtual maize

MatPL, a machine learning force field software package, extended molecular dynamics simulation to a trillion atoms with accuracy described as approaching quantum mechanical calculations. Molecular dynamics follows the movement of atoms over time. A force field estimates the forces between them, which determine that movement.

Machine learning can make those force estimates less expensive to calculate than repeatedly solving a more detailed quantum problem. The conference presentation placed MatPL's achievement at the meeting point of simulation size and numerical accuracy, two demands that are often difficult to satisfy together.

The Fengqing Ensemble V1.0 system for global weather forecasts over short and medium ranges had entered operational use. An ensemble forecast runs multiple versions of a prediction to explore how weather could develop from slightly different starting conditions or assumptions.

For ocean research, swLFS was presented as the world's first global ocean circulation forecasting system with a resolution of one kilometre. Finer resolution lets a model represent smaller spatial features, while increasing the amount of calculation needed across the global ocean.

China Agricultural University's Virtual Maize platform was credited with shortening breeding cycles from more than ten years to several years. The platform brings computing into the search for useful crop traits, helping researchers narrow the combinations they pursue through breeding.

The ten achievements also covered energy exploration and advanced equipment development. Together, the named examples connect the cluster to concrete work: predicting weather, modelling ocean circulation, studying atomic behaviour and selecting crop breeding options.

The network and storage behind the calculations

Sugon says its domestic infrastructure spans chips, computing equipment, storage, networking, cooling, applications and services. Hygon and other domestically developed chips provide the computing foundation. Conference coverage also described six internally developed chips within the broader architecture.

The scaleFabric network uses native remote direct memory access, or RDMA. This allows data to move between machines with less involvement from their main processors, reducing some of the overhead of exchanging information.

Fast connections are essential because a large simulation or AI training run repeatedly shares intermediate results. If processors finish calculations but must wait for data, part of the installed computing capacity sits idle.

Sugon identifies ParaStor distributed storage as another key component. Distributed storage spreads data across multiple machines so large workloads can read and write information through many paths. The company says ParaStor placed first in the production full node and 10 node performance categories of the 2026 IO500 rankings.

Storage performance matters during model training, when large datasets must be read repeatedly, and during scientific simulation, when intermediate results may need to be saved. Saving progress also helps a long calculation resume after an interruption.

At this scale, network congestion, load balancing and fault isolation become central engineering tasks. The scheduler needs to place work where suitable processors and data are available, while the network needs to carry the resulting traffic without turning a local problem into a wider slowdown.

Cooling, density and the cost of keeping cards busy

Sugon 8000 uses immersion cooling with a phase change process. Equipment transfers heat into a cooling liquid, and the liquid's change of state helps carry that heat away. Sugon says the design can support cabinets with power density in the megawatt class.

The company also describes domestically developed coolants and cooling that takes advantage of outdoor conditions throughout the year. These features address a basic constraint: putting more computing power into a confined space creates more heat that must be removed continuously.

The July operational update claimed a twentyfold increase in computing density per unit compared with other supernodes. Sugon's July completion statement separately identified energy efficiency, memory access and network design as necessary tests of a cluster at this scale.

Electricity supply and cooling equipment must support both sustained heavy use and changes in demand. A machine running scientific simulations, AI training and large volumes of inference may face different traffic and power patterns as its workload changes.

The next practical measurements include energy consumed per completed task, time spent waiting for data, job completion times and recovery after faults. The published announcements do not provide a complete set of comparable figures for total power use, operating costs, application speed or user charges.

How widely the system is being used

Sugon's July completion announcement said more than 300 combined scientific computing and AI applications had been optimised across more than 20 fields. More than 70 applications had expanded to the scale of 10,000 cards. Named examples included protein folding, molecular dynamics involving a trillion water atoms and turbulence simulation using 100 trillion grid cells.

The July operational update described nearly 1,000 application adaptations. At the October conference, figures referred to more than 300 deeply optimised priority applications across more than 30 fields, including more than 20 running across the whole system. Another October description counted more than 300 computing tasks across 26 fields. These totals use different categories, including software adaptation, priority optimisation and computing tasks.

The National Supercomputing Internet is intended to make the cluster accessible to universities, research institutions, businesses and individual users. Its role is to connect computing demand with available resources, extending access beyond organisations that can build and operate a large cluster themselves.

Sugon 8000 was also selected as a representative application case under the HPC+AI for Science frontier observation programme launched on October 18. The conference introduced an industry collaboration action plan and an open community for joint work on computing infrastructure, software and applications.

Sugon has announced a strategic partnership with the Beijing Academy of Artificial Intelligence and development of a second fully domestic system containing 100,000 cards. For the first machine, the October results put the focus on completed scientific work and sustained service. The proposed second system would take that approach beyond a single installation.

Key Points

• Sugon 8000 connects 100,000 AI accelerator cards and operates in Zhengzhou, Henan province.
• Sugon announced its completion and connection to the National Supercomputing Internet on July 10, 2026.
• Ten frontier application achievements were presented in Shenzhen on October 18.
• Named projects include a simulation using four trillion dark matter particles, a digital brain with 100 billion neurons and molecular simulation at a trillion atoms.
• The system combines detailed scientific calculations with AI training and inference on one platform.
• Sugon has announced development of a second fully domestic cluster containing 100,000 cards.