This research addresses the critical challenge of power attribution in shared GPU environments using NVIDIA's Multi-Instance GPU (MIG) technology for cloud data centers.

• Develops novel methods to accurately apportion power consumption among multiple GPU partitions
• Enables fair billing and resource optimization for cloud providers and tenants
• Improves energy efficiency while maintaining performance isolation
• Tackles both engineering and security aspects of GPU resource sharing

For cloud providers and ML infrastructure teams, this research offers practical approaches to reduce operational costs and environmental impact while maximizing GPU utilization in multi-tenant environments.

On the Partitioning of GPU Power among Multi-Instances