Power consumption control method and system for GPU server and related components

Abstract

The invention discloses a GPU server power consumption control method and system, electronic equipment and a computer readable storage medium, an alert signal line of a PSU is connected with a GPIO interface of a PCH, and the power consumption control method comprises: obtaining the level state of the GPIO interface according to a preset rule; judging whether the level state meets a frequency reduction condition or not; and if yes, generating a frequency reduction instruction through the operating system management software so as to perform frequency reduction processing on the CPU and the GPU. According to the invention, the dynamic change of the system power consumption of the GPU server can be monitored in real time; signals transmitted from a hardware bottom layer are processed by operating system management software and are rapidly converted into a system software instruction, and frequency reduction operation is performed on key GPU / CPU components; and compared with the prior artusing IO frequency reduction, the GPU server power consumption control method has the advantages that the system instructions can flexibly and quickly control the CPU and GPU frequency reduction of the key components, and the system performance of a GPU server is guaranteed to the maximum extent.

Description

technical field [0001] The present application relates to the field of GPU servers, and in particular to a method, system and related components for controlling power consumption of GPU servers. Background technique [0002] In the era of AI, big data, and cloud computing, all walks of life are striving to build high-efficiency, low-cost data management and operating models to obtain greater return on investment. Therefore, GPU (Graphics Processing Unit, Graphics Processing Unit) servers are rapidly emerging in the AI ​​wave, becoming the most widely used heterogeneous accelerated computing solution in the AI ​​field with the best input-output ratio. The GPU server combines the advantages of GPU parallel computing capabilities and CPU logic control, and meets the configuration of key modules such as computing units, Cache, and controllers according to the needs of actual application scenarios. [0003] In the existing technical solution of the GPU server, since a large numb...

AI Technical Summary

Problems solved by technology

If the power consumption of a single GPU is too high, when multiple GPUs are configured, the total power consumption of the system may exceed the maximum power consumption of a single PSU, causing the GPU server to be powered off; and when the GPU is performing high-speed parallel computing, the power consumption of the GPU changes dynamically It is very fast and has high requirements for the transient response capability of the PSU (Power Supply Unit, power supply unit), but the actual existing GPU / PSU is difficult to match in terms of transient response parameters; the power consumption information of the existing GPU server system can be obtained through The I2C (Inter-Integrated Circuit, two-wire serial bus) bus is read by the BMC (Baseboard Management Controller, Baseboard Management Controller) in the form of polling. The real-time performance and stability are poor, and fast power consumption control cannot be achieved.

Images