NVIDIA MIGs + RL cut AI energy use by 68% in new scheduling study

A new paper accepted at IEEE CCGrid 2025 tackles the growing energy cost of AI/ML workloads by optimizing job scheduling on NVIDIA Multi Instance GPUs (MIGs). MIGs allow a single GPU to be partitioned into up to seven isolated instances, but choosing the right partition size for fluctuating workloads is an open challenge. The authors—from Columbia University and IBM Research—propose a two-part framework: first, they compare four scheduling algorithms (e.g., priority-based, shortest processing time) to pick the best one for a fixed MIG configuration. Then, they apply reinforcement learning (RL) to dynamically repartition the GPU throughout the day based on queue conditions and diurnal patterns.

Using real-world data center traces, the dynamic RL policy outperformed all baselines: 26% better than twice-daily manual repartitioning, 31% better than static partitioning, and 68% better than no partitioning at all (i.e., a monolithic GPU). The multi-objective function balanced energy consumption and job tardiness, critical for both cost and service-level agreements. The RL agent learned to switch between specific MIG configurations (e.g., 1g.5gb vs 3g.40gb slices) at different times, demonstrating that predictive, automatic reconfiguration is feasible without human intervention. This work points toward self-optimizing data centers that can adapt to AI workload surges while keeping power costs in check.