GPU Cluster Allocation Using Model-Specific Power Profiles

The patent applies local quality by assigning different GPU types to different ML models based on their specific energy consumption characteristics. Instead of using newer GPUs for all tasks, the system measures and analyzes power consumption for each GPU-model combination, then optimally assigns GPUs to models where they provide the best energy efficiency. This resolves the contradiction by making GPU allocation heterogeneous and model-specific rather than uniform.

The patent changes the allocation parameter from performance-based to energy consumption-based. Instead of allocating GPUs according to raw performance metrics, the system measures actual power consumption for each GPU-model combination and uses this energy parameter as the primary criterion for allocation. This resolves the contradiction by shifting the optimization objective from productivity to energy efficiency while maintaining adequate performance through the three conditions.

The patent applies preliminary action by measuring and storing power consumption characteristics for each GPU-model combination before actual task allocation. The system pre-establishes an energy consumption profile database that captures the specific energy requirements of different ML models on different GPUs. This resolves the contradiction by performing the complex measurement and analysis work in advance, making the actual allocation process simple while achieving optimal energy efficiency.