Server Thermal CFD Adaptation Using Machine Learning Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Thermal emulation or simulation of servers is computationally intensive and not performed in real-time, leading to high time and labor costs, and there is a need for efficient thermal management solutions to ensure stable operation and extended service life.
Innovation Solution
A method using a trained machine learning model to determine benchmark CFD for a benchmark server, compare it with a target server configuration, and adapt target CFD based on the difference, enabling real-time thermal simulations and simulations with different server components.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If thermal emulation is performed using traditional CFD methods, then thermal distribution accuracy is improved, but simulation time increases to several hours
Solution Approach 1:
The patent creates a digital twin model that copies the physical server's thermal characteristics and uses a pre-trained machine learning model to simulate thermal distribution. This virtual copy allows rapid simulation without performing computationally intensive CFD calculations for each scenario, reducing simulation time from several hours to near-real-time while maintaining accuracy.
Solution Approach 2:
The patent transforms the simulation approach by changing from traditional CFD parameter calculations to machine learning-based predictions. The trained model uses input parameters (server configuration, workload, environmental conditions) to directly output thermal distribution results, bypassing the time-consuming iterative CFD solving process and achieving real-time simulation capability.
2Adaptability or versatility
If thermal emulation is performed for servers with different configurations, then adaptability is improved, but computational resources and time increase
Solution Approach 1:
The patent develops a universal thermal emulation system that can handle multiple server configurations through a single pre-trained machine learning model. The model is trained on diverse server configurations and can predict thermal distributions for new configurations without requiring separate CFD simulations, enabling the system to adapt to different server types while maintaining consistent simulation time.
Solution Approach 2:
The patent performs preliminary training of the machine learning model on a comprehensive dataset of various server configurations before actual use. This preliminary action allows the model to capture thermal patterns across different configurations in advance, so that during actual simulation, the model can quickly predict results for new configurations without performing time-consuming CFD calculations.
3Measurement precision
If traditional CFD simulation is used, then thermal model accuracy is improved, but productivity decreases due to long simulation time
Solution Approach 1:
The patent replaces the mechanical CFD calculation system with a machine learning-based prediction system. Instead of performing iterative numerical solving of fluid dynamics equations, the system uses a trained neural network to directly map input parameters to thermal distribution outputs, achieving both accuracy and high productivity through the substitution of computational methodology.
Data Source
AI summary
Embodiments of the present disclosure provide a method, an electronic device, and a computer program product for determining computational fluid dynamics (CFD) for a server. The method includes determining, by a trained machine learning model, benchmark CFD for a benchmark server. The method further includes determining a difference between a target server configuration for a target server and a benchmark server configuration for the benchmark server, and determining, by the trained machine learning model, target CFD for the target server based on the difference and the benchmark CFD. Embodiments of the present disclosure can realize real-time thermal simulations and enable real-time thermal models with specific server configurations to be adapted to servers with different components.