NHN Cloud Unveils 'GPU Live', an Integrated Solution for AI Training and GPU Operations

NHN Cloud has unveiled GPU Live, an installable multi-tenant all-in-one solution that handles AI training execution and GPU resource operations within a single console. This solution can be installed in both on-premises and cloud environments, designed to simultaneously meet the network isolation requirements of public and financial institutions and the need for GPU sharing among multiple teams within enterprises.

Dynamic Allocation Technology to Enhance Resource Efficiency

GPU Live manages GPUs as a single pool based on the Kubernetes standard interface, adopting a dynamic allocation method that assigns resources on-demand when workloads are generated and immediately reclaims them upon termination. In particular, it applies a Bin Packing policy to prioritize placing new workloads on nodes already in use, thereby securing empty nodes required for distributed training. Additionally, through Gang Scheduling, it executes distributed training instances in bulk only when all are ready, preventing resource waste caused by partial starts. Workloads with GPU utilization below a threshold are automatically reclaimed and reassigned to other tasks.

End-to-End Automation of Training Workloads

Users only need to define a workload that bundles container images, GPU resources, and execution methods, and the entire lifecycle—from infrastructure allocation to result storage and resource return—is automatically handled. It allows combining single or distributed node configurations with interactive (Jupyter/SSH) or batch execution methods. It provides a feature that automatically retries up to 3 times in case of training failure, and allows easy uploading and downloading of large-scale training data via Web SFTP.

Control and Visibility in Multi-Tenant Environments

It isolates resource quotas and permissions through a three-tier structure of system, organization, and project, and supports 6-level Role-Based Access Control (RBAC) from SuperAdmin to Project Viewer. All changes are automatically recorded in immutable audit logs, and personal information of withdrawn users is automatically masked. It provides real-time monitoring based on NVIDIA DCGM and metering data at 5-minute intervals, which can be utilized for billing basis and internal settlement materials.