GPU Rack Power Balancing With Fused Frequency Configuration

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Data centers face challenges in efficiently assigning processing jobs to racks due to variable operating conditions and differing constraints among sub-systems, leading to inefficient power consumption and performance variations.

Innovation Solution

Implementing a system that utilizes fused power coefficients for computing units to calculate and adjust operating frequencies, balancing power consumption across baseboards and racks, and optimizing rack configurations based on task parameters.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If processing jobs are assigned to data center racks based on variable operating conditions, then job processing efficiency is improved, but rack configuration complexity increases

Engineering Contradiction:

Improvejob processing efficiencyVSAvoidrack configuration complexity

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The system changes operating parameters (frequency, voltage, power limits) of computing units based on their measured characteristics and job requirements. By dynamically adjusting these parameters, the system optimizes job processing efficiency while adapting to variable rack configurations without requiring fixed complex hardware arrangements.

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The patent implements dynamic rack configuration where computing units can be selectively enabled or disabled based on job requirements and power constraints. The system dynamically determines which computing units to activate and at what power levels, transforming static rack configurations into adaptive, job-optimized arrangements.

Inventive Principle:

Principle #15Dynamics

2Use of energy by moving object

If power consumption is reduced across data center racks, then energy efficiency is improved, but performance variability increases

Engineering Contradiction:

Improvepower consumptionVSAvoidperformance consistency

Core Design Contradiction:

Use of energy by moving objectVSReliability

Solution Approach 1:

The system applies different power management strategies to different computing units within racks based on their individual characteristics and job requirements. Each computing unit can operate at optimized power levels tailored to its specific performance capabilities and the demands of assigned tasks, achieving overall energy efficiency while maintaining consistent performance through localized optimization.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The system measures actual operating conditions and performance outcomes of computing units, then uses this feedback to adjust power allocation and frequency settings. This closed-loop control ensures that power consumption is minimized while maintaining reliable, consistent performance by continuously adapting to actual system behavior.

Inventive Principle:

Principle #23Feedback

3Speed

If computing units operate at higher frequencies, then processing speed is improved, but power consumption increases

Engineering Contradiction:

Improveprocessing speedVSAvoidpower consumption

Core Design Contradiction:

SpeedVSUse of energy by moving object

Solution Approach 1:

The system applies partial frequency scaling where computing units operate at elevated frequencies only when and where needed for specific job requirements, rather than uniformly across all units. By selectively boosting frequencies for critical tasks while maintaining lower frequencies for less demanding operations, the system achieves improved processing speed for essential workloads while controlling overall power consumption.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS20260047031A1Static data center power balancing and configuration

Publication Date: 2026.02.12 NVIDIA CORP

AI summary

A system to select graphics processing units (GPUs) to execute a task is disclosed. In at least one embodiment, GPUs are selected based on one or more task parameters and one or more fused parameters.