On August 15, the offline finals and award ceremony of the 2026 National College Students Computer System Capability Competition - Intelligent Computing Innovation Design Competition ("Xiandao Cup") concluded at Tongji University.
A total of 160 universities, over 500 teams, and nearly 1,900 students participated in the competition. The two major competition topics focused on the inference optimization of the Qwen Large Model and the deployment of the Meteorological Large Model, respectively. During the competition, participants conducted deployment, tuning, and real-machine testing on the National Supercomputing Internet Platform opened by Sugon.
As the intelligent computing track of the "National College Students Computer System Capability Competition," an event certified by the China Higher Education Society, the "Xiandao Cup" has been held for its seventh edition. This year's competition lasted for nearly three months. After fierce competition, and through expert review, real-machine testing, and on-site defense, the topic "Optimization of Qwen Large Model Inference Services Based on Domestic Accelerator Cards" finally awarded 2 first prizes, 5 second prizes, and 16 third prizes; the topic "Optimization of Meteorological Large Model Inference Deployment Based on Domestic Scientific Intelligent Computing Ecosystem" finally awarded 2 first prizes, 5 second prizes, and 14 third prizes. The total prize pool reached up to 800,000 RMB, and the winning teams can also receive direct offers for campus recruitment.
It is understood that throughout the competition, the Qwen topic had its scores updated a cumulative 7,294 times, and the meteorological topic had its scores updated a cumulative 10,726 times, reaching a total of 18,020 times. On the final competition day, there were still 90 teams in the Qwen topic and 61 teams in the meteorological topic updating their scores.
This year's competition integrated the domestic computing power environment throughout the entire schedule and introduced responsible experts, university teachers, and enterprise technical teams to jointly participate in topic design and technical support. Gong Xiaoliang, a guiding teacher from Tongji University, mentioned in an interview that the models and hardware used in the current competition are the latest, and what students need to think about is how to truly put these software and hardware platforms to use. In this process, teachers provide comprehensive guidance around the problems encountered by students, allowing them to boldly explore various possibilities in scientific research. From last year's operator competition to this year's end-to-end optimization of large models, the evolution of the topics closely follows the technological trends. The optimization solutions produced by students can not only feed back into scientific research but also contribute youthful strength to the construction of the domestic software and hardware ecosystem.
For the domestic computing ecosystem, the optimization solutions and usage feedback formed by the participating teams will further promote the adaptation and iteration of models, frameworks, toolchains, and computing power platforms.
Ying Zhiwei, Vice President of Hygon Information Technology, accepted an interview with a reporter from EEFocus after the competition, engaging in in-depth exchanges on the domestic computing power security system, ecosystem construction, the value of the competition, and the future development direction of the industry.
The Underlying Logic of Security and Controllability
As the wave of generative Artificial Intelligence (AI) sweeps across the globe, large models are penetrating key industries related to national economy and people's livelihood, such as government affairs, finance, energy, and scientific research, at an astonishing speed. However, the superpower of large models is also a double-edged sword. With frequent real-world cases of "large models being able to breach hundreds of firewalls overnight" or discovering unknown vulnerabilities in a short time, how to ensure that large models operate efficiently while taking into account the "security and controllability" of the underlying architecture has become a milestone that the intelligent computing industry must cross.
In response, Ying Zhiwei pointed out that in the face of increasingly complex cybersecurity threats, the traditional boundary defense model is no longer adequate in the AI era. The key to ensuring the security of intelligent computing platforms lies in deeply integrating security mechanisms into the underlying architecture of chips.
Taking the layout of Hygon CPUs and DCUs as an example, the construction of hardware-level security chains has become the core defense line. By providing secure boot capabilities, the computing platform can conduct a full-process verification chain review of the BIOS and operating system at the initial stage of host startup, eliminating the implantation of malicious code from the source.
A more essential change lies in the precise computing capability of "dual-chip collaboration". Ying Zhiwei told the EEFocus reporter that in the past, CPUs and GPUs often operated independently, and data transmission between the two was highly susceptible to security vulnerabilities. However, by bridging the security domains of CPUs and DCUs, large models are protected throughout the entire process from the moment they are loaded onto the CPU, preventing malicious tampering. Coupled with confidential communication technology, the prompts entered by users remain encrypted from the moment they leave the client, directly entering the DCU security domain for inference processing, and finally returning the encrypted results to the client. This full-environment encryption closed loop covering communication links, inference processes, and deployment domains can build a "security wall" for large models that cannot be easily breached.
Domestic Computing Power Has Advanced from "Usable" to "Highly Usable"
The success or failure of the chip industry often depends not only on the stacking of hardware indicators but also on the prosperity of the software ecosystem. In the past, foreign chips, relying on years of deep cultivation in universities and research institutions, established an unshakeable software barrier.
In this regard, Ying Zhiwei stated, "To achieve a breakthrough, domestic computing power must adopt a dual approach focusing on compatibility and localized optimization."
- Seamless Migration and Compatibility: By being compatible with international mainstream ecosystems such as x86 and CUDA, and expanding on this basis (such as Hygon's DTK software stack), domestic platforms enable developers to "seamlessly migrate" their accumulated programming skills and algorithms, significantly lowering the development threshold and adaptation costs.
- Localized Algorithm Optimization: While being compatible with the international ecosystem, and in response to the needs of the domestic market, a large number of domestic-specific algorithms are integrated into the underlying software, operating systems, and operators, allowing domestic chips to better serve local enterprises and national strategies.
This transition from "usable" to "highly usable" has been empirically verified in the training of trillion-parameter large models. The birth and stable operation of top domestic large models such as DeepSeek and Zhipu are precisely based on the maturity of domestic computing power platforms in the software and hardware ecosystem and their performance comparable to world-advanced levels.
Synergistic Advantages of Super Nodes and Full-Precision Scientific Computing
As the scale of large models grows geometrically, the era of "single-machine operation" is in the past. Future AI competition is shifting towards the architectural game of "super nodes" and intelligent computing clusters.
Ying Zhiwei told the EEFocus reporter that since ultra-large models cannot be accommodated on a single machine, dozens or even hundreds of cards must be connected via high-speed bandwidth. Traditional network networking modes are inefficient, while the new generation of super node ecosystems requires chips to achieve memory address sharing and ultra-fast interconnection. In a cluster where thousands of cards operate collaboratively, the latency of any single card can drag down the entire task, posing an unprecedented challenge to the interconnection design of chips.
Meanwhile, in the field of AI for Science, the dual-chip collaboration of CPU+DCU has demonstrated unique technical advantages. Taking the highly complex meteorological large model forecasting as an example, the CPU and GPU operators of traditional international manufacturers often fail to collaborate deeply. In contrast, the domestic dual-chip platform can perceive the entire system state in real time, handing over large and complex operators to the DCU for efficient processing, and allocating fragmented tasks with low computing power requirements to the CPU for computation.
This deep coordination of underlying computing libraries, compilers, and operating systems, coupled with full-precision high-end chip capabilities, can not only improve the utilization rate of the whole machine's computing power but also achieve the dual implementation of scientific research and applications in major livelihood scenarios such as daily weather forecasting, typhoon prediction, and flood prevention.
Conclusion
For intelligent computing platforms to truly gain a foothold, they must simultaneously cross three hurdles: security, ecosystem, and cluster architecture.
When domestic computing power can support the training and inference of hundred-billion and trillion-parameter large models, and is implemented in key livelihood infrastructures such as telecommunications and finance, it indicates that domestic chips have completed the transformation from hardware parameters to actual productivity.
This integrated process of software and hardware, which continuously integrates and receives feedback in practical applications, will become the most pragmatic growth path for the future computing power foundation. And national competitions like the "Xiandao Cup" are precisely one of the best carriers.