BMC Power Ratio Table for Processor GPU Throttling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional power capping algorithms in information handling systems primarily focus on processors and memory subsystems, neglecting the increasing power consumption of GPUs, which results in insufficient power savings and performance losses when GPUs are not adequately capped.
Innovation Solution
Incorporating a power ratio table within the baseboard management controller (BMC) to proportionally throttle the processor and GPU power levels based on workload-specific power ratios, ensuring balanced power management across both components.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If traditional power capping algorithms focus only on processors and memory subsystems, then processor power consumption is reduced, but overall power savings are insufficient due to neglected GPU power consumption
Solution Approach 1:
The power management system segments the information handling system into distinct components (processor, GPU, memory subsystems) and applies targeted power capping algorithms to each segment. The BMC independently manages power for each component based on workload requirements, allowing comprehensive power savings across all segments rather than treating the system as a monolithic unit.
Solution Approach 2:
The power capping algorithms dynamically adjust power allocation based on real-time workload conditions. The BMC continuously monitors workload characteristics and modifies power cap settings for processors, GPUs, and memory subsystems accordingly, enabling the system to adapt power management strategies to varying operational demands and maximize power savings under different conditions.
2Loss of energy
If power capping is applied without considering workload-specific power ratios, then power consumption is reduced, but performance losses increase due to inadequate GPU power management
Solution Approach 1:
The system applies local quality by implementing workload-specific power ratios that tailor power allocation to the particular characteristics of each workload type. Different workloads receive customized power cap configurations based on their computational patterns, ensuring that power is distributed optimally between processors and GPUs for each specific workload, thereby maintaining performance while reducing power consumption.
Solution Approach 2:
The BMC dynamically changes power management parameters (power cap levels, power ratios) based on detected workload characteristics. By monitoring workload behavior and adjusting power parameters in response to changing conditions, the system optimizes the balance between power consumption and performance for each specific workload scenario.
3Loss of energy
If uniform power capping is applied to all components, then overall power consumption is reduced, but performance losses are maximized due to lack of proportional adjustment
Solution Approach 1:
Instead of uniform power capping, the system applies local quality by implementing component-specific power management with unique power cap settings for each processor and GPU based on their individual workload requirements. This allows each component to operate at the optimal power level for its specific function, preventing unnecessary performance degradation while achieving overall power reduction.
Solution Approach 2:
The power capping system dynamically adjusts power allocation ratios between different components based on real-time workload conditions. The BMC continuously modifies power cap settings to maintain optimal performance-power balance, ensuring that power is distributed proportionally to match actual computational demands of each component rather than applying static uniform restrictions.
Data Source
AI summary
An information handling system includes a processor, a graphics processing unit (GPU), and a baseboard management controller (BMC). The BMC includes a power ratio table with a plurality of entries, each correlating a workload with a power ratio of a power level of the processor when the particular workload is instantiated on the information handing system to a power level of the GPU when the particular workload is instantiated on the information handling system. The BMC determines that the power ratio table includes an entry associated with a workload instantiated on the information handling system, determines that a total power level of the information handling system is greater than a power level threshold, and throttles a power level of the processor and a power level of the GPU based upon a power ratio of the entry in response to determining that the total power level is greater than the power level threshold.