On October 18, at CCF HPC 2026, China's national high-performance computing conference held in Shenzhen, Sugon announced 10 research results produced with its "Sugon 8000" computing system. According to Jiemian News, the results span astronomy, earth science and life science, as well as materials simulation and engineering R&D. In July, the company had announced that it completed the Sugon 8000, which it described as China's first AI cluster of 100,000 cards built with domestically developed technology. The latest announcement draws attention to how this large pool of computing resources is actually being used in research, and how closely high-precision scientific computing and AI are being linked.
How the 10 research results differ from the July completion announcement
The Sugon 8000's completion was announced on July 10, and the physical system was first shown to the public on July 17. The new development is the announcement of 10 application results on October 18. Putting Sugon's official announcements and Jiemian News's report in chronological order shows that each announcement about the same system means something different.
| Date (2026) | Event | What can be confirmed and source |
|---|---|---|
| July 10 | Completion announced | Sugon announced the system's completion and its connection to the national supercomputing internet |
| July 17 | System shown publicly for the first time | The company said it debuted the system at WAIC in Shanghai |
| October 18 | 10 application results announced | Jiemian News reported on the presentation at CCF HPC 2026 |
Separating completion, public debut and research results clarifies where this announcement stands. The Sugon 8000 has moved from showing that the facility itself is finished to showing concrete research done with it. However, this announcement alone does not substantiate the technical superiority of each of the 10 results, nor has the company's "world first" claim been confirmed through independent verification.
The July 10 completion announcement also said that optimization of more than 300 application programs had been completed, and that more than 70 of them could scale to 10,000 compute cards. One caveat: the number of supported programs and the number of compute cards any individual program can use are different metrics. It does not mean all 300 programs used all 100,000 cards at once, and for the 10 results announced now, it is necessary to check how much computing resource each one used.
Sina Finance, which republished a China Central Television (CCTV) article, and IT Home, which summarized it, described the Sugon 8000's computing power with wording to the effect that it could "hold the universe." This expression refers to the ability to model complex phenomena, such as the evolution of the universe or the workings of the brain, on a computer; it does not mean the entire universe has been completely reproduced. Both articles are based on the same CCTV report, so they cannot be treated as two independent verifications of performance.
How high-precision scientific computing and AI are combined
The Sugon 8000 is distinguished by a design that handles high-precision scientific computing and AI training and inference in a single system. According to Sugon, it supports numerical formats from FP64 to INT8. FP64 is 64-bit floating point and INT8 is 8-bit integer, so calculations of different precision can be used as needed. The configuration supports scientific simulations that demand high precision and AI workloads that can run at relatively low precision on the same computing platform.
Numerical simulations based on physical laws can generate data that is difficult to collect in real experiments. AI learns trends from that data and narrows down candidates worth examining in detail. Running the selected candidates through numerical simulation again allows researchers to verify predictions and refine models. In an official announcement in August, Sugon also described a research approach that repeats data generation, model training and verification of applications.
For example, simulating an enormous number of candidates one by one at high precision would take a great deal of computing time. In such research, using AI to pick promising candidates and concentrating computing resources on them may improve the efficiency of the research as a whole. If scientific computing and AI can run on the same equipment, it also becomes easier to combine model training with checks of physical validity. However, this is the research method the integrated system aims to support, and it has not been demonstrated that the same benefit was achieved in all 10 announced results. Even when AI predictions are verified by numerical simulation, that does not guarantee that the assumptions and computing conditions of the simulation itself are correct.
Running a cluster of 100,000 compute cards efficiently depends not only on computing performance but also on communication, storage and cooling. Sugon combines its high-speed network "scaleFabric," which handles data transfer between compute cards, with its distributed storage system "ParaStor," which reads and writes large volumes of data. For cooling, it adopted an immersion cooling method that uses the phase change of a coolant. In large-scale scientific computing, communication between cards and access to storage affect processing speed, and removing the heat generated efficiently is essential for stable long-term operation.
What the rocket plume simulation shows, and what remains open
One concrete use of the Sugon 8000 is the rocket plume simulation that Sugon published on October 9. The company and a research team at Xi'an Jiaotong University ran a three-dimensional, high-precision numerical simulation of a structure with 33 nozzles arranged in a ring. The study analyzes the complex pressure fluctuations and heat distribution that arise when the high-temperature, high-speed flows from multiple engines interfere with one another.
The computational scale reported by the company reaches 6.774 trillion grid cells. This is a conversion into English units of the Chinese announcement's "6.774万億," and it indicates a scale at which the space being computed is divided finely and fluid behavior is calculated in each region. By examining in detail how multiple plumes collide and mix, the aim is to inform the design of rocket structures and thermal protection and to lead to tests on actual vehicles. However, this is a result of numerical simulation, not an announcement that a rocket recovery flight succeeded.
On computing performance, the company reports achieving weak scaling efficiency of 99.01% and strong scaling efficiency of 73%. Weak scaling efficiency indicates how well processing efficiency is maintained when the problem size grows along with computing resources. Strong scaling efficiency, by contrast, measures how efficiently processing time shrinks when more computing resources are applied to a problem of the same size. Distinguishing the two makes it possible to evaluate separately the ability to handle larger problems and the ability to solve the same problem faster.
These figures, however, are results Sugon reported for a specific rocket plume simulation. The announcement does not give the baseline number of compute cards used for comparison or detailed measurement conditions. The 99.01% figure therefore cannot be read as the utilization rate of all 100,000 compute cards, nor can it be assumed that the same level of performance applies to every use. The full paper and third-party replication results have not been released either, so comparing performance with other supercomputers would require further verification under matched computing conditions.
How far AI was used in this rocket plume study is also an important point. Sugon lists the construction of "surrogate models" that combine simulation data with AI to approximate complex physical calculations, and the development of fast prediction methods, as future research topics. In other words, while concrete results have been shown for running the high-precision simulation, how much AI integration can shorten research time is left to future verification.
For the 10 announced results as well, it matters to check not only the size of the computation but also each study's problem setting, required precision and actual computing time. In particular, if researchers can compare how required time and prediction error change with and without AI, the benefit of integrating scientific computing and AI could be assessed more concretely. If power consumption data were also shown, research institutions could judge not only whether they can run a computation but also how efficient it would be to use the system continuously. The true measure of the Sugon 8000 lies not only in the number of compute cards it carries, but in how efficiently those cards allow research trials and verification to proceed.