Multi-GPU server cluster liquid cooling flow distribution method based on reinforcement learning
Through the reinforcement learning-based graph neural network and Actor-Critic architecture, combined with the GPU thermal dynamic characteristics and liquid cooling topology status, adaptive regulation of liquid cooling flow in multi-GPU server clusters is achieved, solving the problems of delayed response and insufficient model adaptability of liquid cooling systems in existing technologies, and improving the system's stability and energy efficiency.
Patent Information
- Application Number
- CN202510802664.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-16
- Publication Date
- 2025-09-26
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The liquid cooling system in the existing multi-GPU server cluster cannot reflect GPU thermal fluctuations in real time, resulting in delayed cooling response or excessive flow regulation, affecting system stability and energy efficiency. In addition, the reinforcement learning model lacks the ability to adapt to dynamic changes in hardware topology and fluid paths, increasing deployment costs and difficulty.
A reinforcement learning-based approach is adopted, through graph neural networks and actor-critic architecture, combined with GPU thermal dynamic characteristics and liquid cooling topology status, to achieve adaptive control of liquid cooling flow, and introduce a policy migration mechanism to support seamless adaptation when the topology structure changes.
It improves the heat dissipation efficiency and energy consumption balance control capabilities of multi-GPU server clusters, improves the response speed and robustness of the liquid cooling system under dynamic load changes, and reduces deployment costs and training time.