UPS DC Link Voltage Control for Cyclic AI Server Loads
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Server systems with variable loads, such as those used for AI-based platforms, experience reduced battery life due to repeated discharge cycles from high-load and low-load elements, leading to inefficient power management.
Innovation Solution
A power component utilizing AI/ML models to predict load cycles and adjust DC link voltage accordingly, anticipating high-load elements to reduce battery discharge frequency and severity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If the UPS battery provides continuous power to servers with variable AI loads, then the server power requirements are met, but the battery life is reduced due to repeated discharge cycles
Solution Approach 1:
The system performs preliminary actions by detecting the cyclic pattern of AI workloads (data collection followed by model training) and proactively adjusting the DC link voltage before the high-load phase begins. The controller identifies when the system is in a low-load data collection phase and prepares the power component by increasing DC link voltage, so that when the high-load training phase starts, the battery is not forced into a deep discharge cycle, thereby extending battery life while maintaining power delivery capability
Solution Approach 2:
The system implements dynamic adjustment of the DC link voltage based on the detected workload phase. Rather than maintaining a static voltage level, the controller dynamically modulates the DC link voltage in response to the cyclic nature of AI workloads, increasing voltage during data collection phases and maintaining appropriate levels during training phases, thereby optimizing both power delivery and battery preservation
2Duration of action of stationary object
If the DC link voltage is adjusted in response to predicted load cycles, then battery discharge cycles are reduced, but the system complexity increases due to AI/ML model integration
Solution Approach 1:
The system implements feedback by continuously monitoring the actual power draw and workload phase, comparing it against the AI/ML model's predictions, and adjusting the DC link voltage accordingly. The controller receives feedback from the server's power consumption patterns and refines its voltage adjustment strategy, allowing the system to adapt to varying workload characteristics while managing complexity through learned patterns rather than complex rule-based logic
Solution Approach 2:
The AI/ML model enables the system to serve itself by automatically learning and adapting to the specific workload patterns of the AI server without requiring manual configuration or complex control algorithms. The system self-adjusts the DC link voltage based on patterns it has learned from historical data, reducing the need for complex external control systems while extending battery life
Data Source
Figure 1A
Figure 1A
Figure 1B
AI summary
A system may include a power component, the power component including at least one device processor configured to: obtain load data, wherein the load data comprises a power characteristic associated with at least one of a high-load element or a low-load element, obtain a trained power management artificial intelligence (AI) and/or machine learning (ML) model, based at least on the load data and the trained power management Al and/or ML model, infer a direct current (DC) link voltage adjustment, wherein the DC link voltage adjustment is correlated with a predicted load cycle parameter; and cause the power component to alter a DC link voltage from an initial level to an adjusted level based on the DC link voltage adjustment.