In this session, we will describe about our first AI Factory in Vietnam and Japan regions that serves large NVIDIA GPU cloud and AI platform/application marketplace, that includes the architecture, services, operation, optimization of GPU cloud and valuable knowledge of how to avoid bottlenecks to achieve high performance AI workloads.

In our system, we utilizes OpenStack as IaaS platform and Kubernetes as AI application platform. We will share our knowledge and issues/solutions in our development and deployment the AI Factory. For multi-tenant, our resource scheduler assign CPU and GPU considering performance and power optimization and manage auto-scaling algorithms.

AI workloads have different features such as training, inference and rendering, so optimized hardware are different for each workloads. So it is necessary to schedule resources and tasks depend on them. We will share the latest tips and know-hows in the presentation.