Analysis of Application Scenarios and Computing Power Demand in Research Institutions, High-Performance Computing Becomes the Cornerstone of AI Research

AI model training, climate simulation, and vaccine design on campus are turning from theories into reality through the round-the-clock operation of thousands of high-performance GPUs.

In the early morning on university campuses, high-performance computing centers begin their daily operations. Thousands of GPU indicator lights blink in server rooms, carrying training tasks for various artificial intelligence models. Some of these models are analyzing cosmic images captured by astronomical telescopes, some are simulating protein folding structures, and others are learning the subtle patterns of human language.

Against the backdrop of the global AI wave, research institutions are experiencing an explosive growth in computing power demand. The University of Texas at Austin recently doubled the computing capacity of its generative AI center, now equipped with over 1,000 advanced graphics processing units.

The university's goal is clear: to take the lead in open-source AI innovation and drive breakthroughs across multiple fields from vaccine design to natural language processing.

01 Application Scenarios

The application of computing power in research institutions has long expanded beyond traditional engineering calculations and data analysis, with the rise of artificial intelligence significantly broadening the boundaries of computing applications.

AI large model research has become the primary scenario for computing power application in universities. The expansion of the University of Texas's generative AI center enables researchers to train large AI models "from scratch."

This approach offers significant advantages over using pre-built models—researchers can control the "interpretability" of models, understanding why models reach specific conclusions, which is crucial for reducing bias and ensuring accuracy.

In the field of multidisciplinary scientific computing, computing power has become a universal tool supporting cutting-edge research. From precise weather forecasting to new material development, from drug molecule screening to gene sequence analysis, computing power is driving the core engine of scientific innovation.

Tsinghua University's Earth System Model is a typical example—this model serves as an important tool for understanding the mechanisms of past climate and environmental evolution and projecting potential future global change scenarios.

Jilin University's High-Performance Computing Center has supported the Ministry of Land and Resources' deep exploration project, developing the independently copyrighted OpenProbe geoscience data 3D visualization and interpretation platform.

Campus intelligent construction equally relies on computing power support. The DeepSeek platform deployed by Jilin University serves as a supplement to public services, providing safe and stable on-campus AI solutions for all teachers and students.

The deep integration of AI technology with school management enables more comprehensive and accurate utilization of school basic data, optimizing management processes and assisting scientific decision-making, thereby improving the efficiency of school administrative work.

02 Demand Characteristics

The computing power demands of research institutions display distinct characteristics, fundamentally different from traditional commercial applications.

The requirement for computing performance is particularly prominent. University research tasks typically involve complex large-scale data processing, high-precision simulation, and large AI model training—these tasks place extremely high demands on computing capability, storage bandwidth, and network latency.

The computing platform of Tsinghua University's Department of Statistics and Data Science is equipped with various GPU resources, including NVIDIA 3090, 4090, V100, and A100 series, with 8 cards per single node, also providing NVLink and NVSwitch to meet the performance requirements of different research tasks.

Flexibility in resource scheduling is another key demand. Research project workloads fluctuate significantly, often requiring only minimal computing power during normal periods, while potentially needing concentrated substantial computing resources during critical phases.

China Agricultural University's computing power sources are diversified, including university-level platform computing power, key laboratory-specific computing power, and small-scale computing power built by teachers and students, forming a multi-level computing power supply system.

Research institutions are particularly sensitive to cost control. For research projects requiring substantial computing power over extended periods, this represents a significant expense. Yao Ge, Deputy Director of Nanjing University's Information Construction Management Service Center, points out that AI computing clusters require substantial initial investment and continuous subsequent expansion, posing considerable pressure for universities.

Data security and compliance are also particularly important considerations for research institutions. Many studies involve sensitive data, such as medical records, patented technologies, or unpublished research results, which are unsuitable for storage on public clouds.

Jilin University's choice to deploy the DeepSeek platform on-campus is precisely to provide AI capabilities while ensuring data remains within the campus.

03 Challenges

Facing growing computing power demands, research institutions nevertheless encounter multiple challenges.

Insufficient funding is the most prominent obstacle. Computing platform construction and maintenance require substantial funds, including high-performance computing equipment, storage devices, computing networks, management systems, and supporting facility rooms.

Lao Fengdan, Deputy Director of China Agricultural University's Network Technology Center, admits that due to insufficient funding, many universities build according to relatively low standards, with subsequent operation and maintenance facing numerous limitations.

Talent team building is another major pain point. HPC community surveys show that nearly 70% of university operation teams have fewer than 5 people, with some even having "one person multiple roles" situations.

The core operation team of China Agricultural University's Intelligent Computing Center mainly consists of 2 part-time teachers—such staffing makes it difficult to provide deep technical support.

Lagging infrastructure also constrains computing power development. Many university equipment rooms lack the power and cooling conditions to support intelligent computing platform expansion.

Fan Chun, Director of Peking University Computing Center's System Management Office, offers a vivid analogy: "If building computing power is like buying aircraft, it's like India—having advanced aircraft but unable to form advanced combat capability."

Security risks are equally not ignored. AI output content carries security risks, sometimes involving sensitive material. When AI is deployed locally, knowledge base systems may mount multiple different knowledge bases, with much information not allowed off-campus—if hacked, this could lead to information leakage.

04 Solutions

Facing complex and diverse computing power demands with severe challenges, research institutions are exploring various solutions.

Multi-level computing power systems have become the choice for many universities. China Agricultural University divides computing power into three levels: university-level high-performance computing platforms, key laboratory-specific computing power, and small-scale cluster computing power built by departments, teachers, and students.

University-level platforms serve all teachers and students, meeting multiple research groups' simultaneous online submission of HPC and AI computing jobs; key laboratory computing power mainly serves laboratory teachers and students; while small-scale computing power executes small computing tasks for teacher teams.

Resource sharing mechanisms are key to improving resource utilization. Northeast University opens data center computing resources, allowing staff to apply for cloud host resources as needed.

Lin Xinhua, Deputy Director of Shanghai Jiao Tong University's Network Information Center, emphasizes: "Building computing power is currently a hotspot, but managing computing power hasn't attracted sufficient attention. Like raising children, three parts birth, seven parts rearing. Computing power operation and maintenance are actually more important than building computing power."

University-enterprise cooperation models bring new possibilities for universities. Jilin University's DeepSeek platform was jointly deployed by the Big Data and Network Management Center, High-Performance Computing Center, and Huawei Company.

Such cooperation both solves universities' technical and resource bottlenecks and provides application scenarios for enterprises.

Flexible charging systems help balance resource allocation and cost recovery. Nanjing University's Yao Ge suggests referring to the evolution of campus network services—setting basic free quotas, with excess usage charged by volume; differential pricing could also be implemented based on model scale differences.

Reasonable charging models not only ensure fairness in resource allocation but also effectively regulate supply-demand relationships.

At Jilin University's High-Performance Computing Center, a mature computing power service system has formed. They provide high-performance computing services for universities, research institutions in Jilin Province, and China FAW Group Technical Center, cumulatively supporting users in publishing 903 SCI papers and supporting over 90 national and ministerial-level projects.

The Omniyq Cloud computing power rental platform (www.omniyq.com) precisely targets this market demand, providing cost-effective computing power services to support research computing in artificial intelligence fields for research institutions like Tsinghua University and Jilin University. Omniyq Cloud's support program is becoming the silent guardian of research breakthroughs.