A UK university has witnessed rapid growth in AI research and education, yet its existing GPU resources remain fragmented without unified scheduling or management mechanisms. This has led to high idle rates, cumbersome research environment deployment, and underutilized computing power. To address these issues, the university plans to establish an integrated AI management platform, enabling GPU resource pooling and intelligent management to create an efficient, flexible, and scalable AI research and teaching environment.

GPU servers are distributed across faculties, laboratories, and training buildings, lacking centralized monitoring and management, resulting in low resource utilization.

AI development environments are diverse, with different faculties and laboratories having varying software version and framework dependencies. Manual deployment is time-consuming and prone to errors.

As the large student population and growing GPU needs for research and coursework, traditional methods struggle to support multi-tenant concurrent usage scenarios.

Additionally, research collaboration efficiency is hampered by the dispersion of data, algorithms, and models across disparate systems, with no unified management or sharing mechanism in place.

The project ultimately adopted the AICPLIGHT Integrated AI Management Platform as its core system to achieve resource consolidation and computing power pooling.

Through containerization and Kubernetes (K8s) orchestration, the platform enables flexible allocation of GPU, CPU, and other computing resources. It supports elastic resource scaling, multi-tenant isolation, and on-demand distribution.

The platform covers end-to-end management from dataset management, algorithm development, model training to model deployment. Supports multiple mainstream frameworks.

Tailored permissions are set for faculty, students, and researchers, supporting self-service environment creation and experiment replication. The intuitive interface lowers the barrier to AI research while enhancing efficiency in both academic and research activities.

Unified scheduling and virtualization of over 100 GPUs achieve >90% utilization.

Supports concurrent AI training and experiments for thousands of users, with hundreds of parallel AI containerized environments.

Containerization cuts environment setup time by 70%, while automated management significantly reduces operational overhead.

Unified data and model management fosters cross-disciplinary collaboration and accelerates the delivery of research outcomes.