Power/Thermal-Aware System Operation

5.37.1Connecting Device, Rack, and Facility Power Budgets#

5.37.2Power Capping, Frequency Scaling, and Job Progress#

5.37.3Synchronized Training Phases, Power Swings, and Short-Term Energy Buffering#

5.37.4Cooling Capacity, Thermal Margin, and Concurrently Operable Device Count#

5.37.5Jointly Considering Placement, Concurrency, Batch Size, and Power Limits#

5.37.6Average Power, Peak Power, Tokens/Joule, and SLOs#

5.37.7The Boundary Between Hardware Design Constraints and Runtime Scheduling Policy#