A dynamic adjustment method applied to power supply of GPU chip

Abstract

The application is suitable for the technical field of GPU power supply, and provides a dynamic adjustment method applied to GPU chip power supply, which comprises the following steps: obtaining the peak power consumption of a current model GPU; dividing the value of the peak power consumption into multiple power consumption ranges according to the size, setting corresponding register values for each group, recording the register addresses and associating them; setting the voltage transient change amount corresponding to each group of power consumption ranges, and corresponding to the register values one by one; after the GPU is powered on, the power consumption value of the GPU chip is monitored in real time, the corresponding power consumption range is obtained, the register value is obtained through the register address addressing, and finally the voltage transient change amount is automatically matched to realize the dynamic adjustment of the power supply. The application realizes real-time monitoring and dynamic adjustment, improves the degree of automation, can solve the problems caused by the power supply ripple, surge and excessive power supply voltage drop on the GPU, and improves the operation reliability of the GPU.

Description

Technical Field

[0001] This invention belongs to the field of GPU power supply technology, and in particular relates to a dynamic adjustment method for power supply of GPU chips. Background Technology

[0002] With the rapid rise of domestically produced GPUs (Graphics Processing Units), the performance of self-developed GPU chips by various domestic companies is becoming increasingly powerful, and consequently, the peak power consumption of these chips is also constantly increasing.

[0003] The power supply plays a crucial role as the power source for the GPU. During GPU operation, sudden changes in the pressure of 3D and 2D applications lead to sudden changes in power consumption. When the total power cannot change rapidly, surges can occur in the GPU power supply circuit. Moreover, when the GPU chip's power consumption is high, the power supply ripple increases, the voltage drop in the GPU power supply circuit also increases, and the frequency of surges during operation also increases. Larger ripples not only reduce power supply efficiency and interfere with the logic relationships in digital circuits, but can also generate overvoltages and even burn out components. Surges can damage the metallized surface of components, damage printed circuit board traces or contacts, and may also cause data transmission errors and failures. Reducing power supply ripple and surges generated during GPU chip operation while maximizing GPU chip performance presents a significant challenge for designers.

[0004] To address the various impacts of power supply on GPU operation, a common approach is to connect an external computer system to the GPU's communication interface. The computer system then sends commands to control the power supply chip's output parameters. While this method can modify target power parameters, it has a low degree of automation and can only modify the same parameter at a time, making it unsuitable for applications with significant load variations.

[0005] Therefore, it is necessary to design a power supply monitoring scheme, especially when the GPU enters the application verification stage, to monitor power consumption parameters in real time and dynamically adjust power parameters in order to improve power supply efficiency and GPU reliability, as well as effectively suppress the impact of power supply ripple, surges and other factors on the GPU. Summary of the Invention

[0006] In view of the above problems, the purpose of this invention is to provide a dynamic adjustment method for power supply of GPU chips, which aims to solve the problems caused to GPUs by power ripple, surge and excessive power voltage drop.

[0007] The present invention adopts the following technical solution:

[0008] The dynamic adjustment method for power supply of GPU chips includes the following steps:

[0009] Step S1: Obtain the peak power consumption of the current GPU model;

[0010] Step S2: Divide the peak power consumption values ​​into multiple power consumption ranges according to their magnitude, and set corresponding register values ​​for each group, record the register addresses and associate them;

[0011] Step S3: Set the voltage transient change corresponding to each power consumption range and match it with the register value one by one;

[0012] Step S4: After the GPU is powered on, monitor the power consumption of the GPU chip in real time, obtain the corresponding power consumption range based on the currently obtained power consumption value, perform communication bus addressing according to the associated register address to obtain the register value, then automatically match the voltage transient change, and finally realize dynamic adjustment of power supply.

[0013] Furthermore, in step S4, if the power consumption of the GPU chip is detected to be greater than the peak power consumption or less than or equal to 0, the protection mode is activated, an error interrupt command is sent, and the device is forced to shut down.

[0014] Furthermore, the specific process of step S2 is as follows:

[0015] S21. Set N-1 dividing points to control the peak power consumption P. max The power consumption is divided into N groups, and the N-1 dividing points are respectively

[0016] S22. Set the hexadecimal register value for each power consumption range;

[0017] S23. Set the monitoring period T and the register address;

[0018] S24. The power consumption range grouping, register value, and register address are associated and written to the GPU chip power management module.

[0019] Furthermore, the specific process of step S3 is as follows:

[0020] Set the voltage transient change of the power supply chip output corresponding to each power consumption range of the group, and match it one by one with the register value corresponding to each power consumption range of the group. Write each voltage transient change into the read-only memory in the power management module of the GPU chip. For the group with a smaller power consumption range, the voltage transient change is set to be larger. When the power consumption range group is the largest, the corresponding voltage transient change is adjusted to be the smallest.

[0021] Furthermore, the specific process of step S4 is as follows:

[0022] S41. When the GPU is powered on, the GPU chip power management module monitors the power consumption of the GPU chip in real time according to the monitoring cycle T and initializes the registers.

[0023] S42. Based on the currently acquired power consumption value, automatically classify it into the corresponding power consumption range group, perform communication bus addressing according to the associated register address, and send the corresponding register value to the power supply chip after finding it.

[0024] S43. The power supply chip automatically matches the voltage transient change according to the register value, sends the result to the output controller and applies it;

[0025] S44. If the power consumption of the GPU chip is detected to be >120% P max If the value is ≤0, activate protection mode, send an error interrupt command, and force shutdown.

[0026] The beneficial effects of this invention are as follows: This invention designs a dynamic adjustment method for GPU chip power supply. The GPU chip power management module monitors the power consumption value in real time, with a monitoring period typically at the millisecond level. Based on the power consumption range group that the power consumption value falls into, the register value of the power supply chip is adjusted in real time. When the power supply chip receives the register value of the preset parameters, it will change the corresponding power parameters, that is, find the corresponding voltage transient change. This adjustment method achieves the purpose of real-time monitoring and dynamic adjustment, improves the degree of automation, and realizes the "interaction" between the GPU and the power supply. It can solve the problems caused to the GPU by power ripple, surge and excessive power voltage drop, and improve the reliability of GPU operation. Attached Figure Description

[0027] Figure 1 This is a flowchart of a dynamic adjustment method for power supply of a GPU chip provided in an embodiment of the present invention;

[0028] Figure 2 This is a schematic diagram of the GPU power supply system provided in an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0030] To illustrate the technical solution described in this invention, specific embodiments are described below.

[0031] like Figure 1 , 2 As shown, this embodiment provides a dynamic adjustment method for the power supply of a GPU chip, which is based on... Figure 2 The GPU power supply system shown includes a GPU chip power management module and a power supply chip, which are connected via a transmission bus. The transmission bus can be configured as an IIC bus, UART, SPI, JTAG, etc., depending on the GPU model and the power supply chip. Furthermore, the register types can be converted according to the protocol, enabling flexible configuration of the data transmission carrier and transmission type.

[0032] Based on the aforementioned GPU power supply system, the dynamic adjustment method for GPU chip power supply provided in this embodiment includes the following steps:

[0033] Step S1: Obtain the peak power consumption of the current GPU model.

[0034] The peak power consumption P of different GPU models during normal operation max The peak power consumption of the current GPU model can be obtained from the parameters, but the load varies greatly in real-world applications.

[0035] Step S2: Divide the peak power consumption values ​​into multiple power consumption ranges according to their magnitude, and set corresponding register values ​​for each group, record the register addresses and associate them.

[0036] This step first divides the peak power consumption into N groups based on its magnitude. For example, N-1 dividing points are set, namely... The N power consumption ranges are divided into groups as follows:

[0037] Each group of power consumption ranges is set with a corresponding hexadecimal register value (R1, R2, R3, ..., R...). N Then, the monitoring period T = T0 (ms) and the register address (REG Addr) are set. The power consumption range grouping, register value, register address (i.e., addressing address) and other information are associated and written into the GPU chip power management module. The specific grouping and assignment settings are shown in Table 1 below.

[0038] Table 1: Correspondence between Power Consumption Value, Register Value, and Voltage Transient Change

[0039]

[0040] Step S3: Set the voltage transient change corresponding to each power consumption range and match it with the register value.

[0041] Voltage transient change is the amplitude of voltage fluctuation per unit time. The power management module outputs the absolute value of the voltage rise or fall per unit time. The higher the value, the larger the voltage change per unit time, but the higher the MOSFET's conduction stability; the lower the value, the lower the voltage change per unit time, and the lower the MOSFET's conduction stability.

[0042] This step sets the voltage transient level (Volt_Level) output by the power supply chip corresponding to each power consumption range group. According to the power consumption range of the groups from largest to smallest, the corresponding Volt_Level values ​​are assigned as LEVEL_1, LEVEL_2, LEVEL_3, ..., LEVEL_N. These values ​​are then mapped one-to-one with the register values ​​corresponding to each power consumption range group, and the voltage transient level is written to the read-only memory in the GPU chip's power management module. For groups with smaller power consumption ranges, the corresponding voltage transient level is set to be larger; when the power consumption range group is the largest, the corresponding voltage transient level is adjusted to be the smallest. That is, the voltage transient level corresponding to register value R1 is the largest, LEVEL_N; register value R... N The corresponding voltage transient change is the smallest, LEVEL_1.

[0043] Step S4: After the GPU is powered on, monitor the power consumption of the GPU chip in real time, obtain the corresponding power consumption range based on the currently obtained power consumption value, perform communication bus addressing according to the associated register address to obtain the register value, then automatically match the voltage transient change, and finally realize dynamic adjustment of power supply.

[0044] When the GPU powers on, the GPU chip's power management module monitors the chip's power consumption in real time according to a monitoring cycle T and initializes its registers. For the currently acquired power consumption value P, it automatically categorizes it into the corresponding power consumption range group, performs communication bus addressing based on the associated register address, finds the corresponding register value, and sends it to the power supply chip. The power supply chip automatically matches the voltage transient change Volt_Level based on the register value, sends the result to the output controller, and applies it.

[0045] Specifically, if the GPU chip's power consumption is greater than 120% of its peak power... max If the value is ≤0, the protection mode is activated, and an error interrupt command FULL is sent to indicate that a malfunction has occurred, forcing a shutdown.

[0046] After each automatic matching of voltage transient changes is completed, the monitoring is repeated every T0 milliseconds, and the voltage transient changes are automatically adjusted and applied according to the information.

[0047] In this embodiment, the GPU chip power management module monitors power consumption in real time. When power consumption is high, it sends corresponding register values ​​based on the power consumption range. The power supply chip automatically matches the voltage transient changes based on the register values, thereby reducing the voltage transient changes output by the GPU power supply chip and addressing the issues of large surges and ripples. However, if the voltage transient changes remain at a low level for an extended period, the conduction stability of the power chip's MOSFETs will decrease, and the stability of the GPU power supply will also decline, potentially leading to occasional computer freezes, black screens, and other abnormal situations. Therefore, in this embodiment, when power consumption is low and power ripple is low, the corresponding register values ​​are sent in real time to increase the transient changes and improve the conduction stability of the MOSFETs.

[0048] The GPU chip power management module is part of the GPU power control system. It can switch the power supply on and off within the GPU chip and monitor power consumption parameters, making it widely used in scenarios with low power consumption requirements. To address the potential for cumbersome procedures and reduced reliability associated with high-power GPU chips, this invention improves the functionality of the GPU chip power management module, as follows:

[0049] a) Real-time monitoring function: Used to monitor the power consumption of the GPU chip. The power consumption value P will be read every time T. The real-time monitoring function will be automatically turned on after the GPU is powered on.

[0050] b) Control function: The monitored power consumption value is matched with the pre-grouped register value REG Value, and the register value is sent to the GPU transmission interface to realize the conversion between power consumption parameters and transmission data;

[0051] c) Transmission function: The register values ​​in the GPU transmission interface are sent to the power supply chip via the transmission bus through the preset register address to realize dynamic adjustment of power supply.

[0052] Therefore, the GPU chip power management module in this embodiment monitors power consumption in real time, obtains pre-group register values ​​based on power consumption, and then dynamically adjusts power supply parameters, i.e., voltage transient changes. This real-time monitoring and dynamic adjustment method can effectively suppress the impact of power supply on the GPU chip.

[0053] In summary, this invention is a method for dynamically adjusting power supply parameters in the field of GPU chips by monitoring power consumption parameters in real time. By monitoring power consumption values, the corresponding pre-group register values ​​are obtained, and then the corresponding voltage transient changes are obtained, thereby realizing self-adjustment of power supply parameters. This allows the power supply status of the GPU (not limited to voltage and current) to be adjusted in real time according to the load size, achieving a dynamic balance.

[0054] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic adjustment method for power supply of a GPU chip, characterized in that, The method includes the following steps: Step S1: Obtain the peak power consumption of the current GPU model; Step S2: Divide the peak power consumption values ​​into multiple power consumption ranges according to their magnitude, and set corresponding register values ​​for each group, record the register addresses and associate them; Step S3: Set the voltage transient change corresponding to each power consumption range and match it with the register value one by one; Step S4: After the GPU is powered on, monitor the power consumption of the GPU chip in real time, obtain the corresponding power consumption range based on the currently obtained power consumption value, perform communication bus addressing according to the associated register address to obtain the register value, then automatically match the voltage transient change, and finally realize dynamic adjustment of power supply. The specific process of step S3 is as follows: Set the voltage transient change of the power supply chip output corresponding to each power consumption range of the group, and match it one by one with the register value corresponding to each power consumption range of the group. Write each voltage transient change into the read-only memory in the power management module of the GPU chip. For the group with a smaller power consumption range, the voltage transient change is set to be larger. When the power consumption range group is the largest, the corresponding voltage transient change is adjusted to be the smallest.

2. The dynamic adjustment method for power supply of a GPU chip as described in claim 1, characterized in that, In step S4, if the power consumption of the GPU chip is detected to be greater than the peak power consumption or less than or equal to 0, the protection mode is activated, an error interrupt command is sent, and the device is forced to shut down.

3. The dynamic adjustment method for power supply of a GPU chip as described in claim 1, characterized in that, The specific process of step S2 is as follows: S21. Set N-1 dividing points to control peak power consumption. The power consumption is divided into N groups, and the N-1 dividing points are respectively , ... ; S22. Set the hexadecimal register value for each power consumption range; S23. Set the monitoring period T and the register address; S24. The power consumption range grouping, register value, and register address are associated and written to the GPU chip power management module.

4. The dynamic adjustment method for power supply of a GPU chip as described in claim 3, characterized in that, In step S22, each power consumption range is set with a corresponding hexadecimal register value, which are respectively , , ... In step S3, the register value The corresponding voltage transient change is the largest, and the register value is... The corresponding voltage transient change is the smallest.

5. The dynamic adjustment method for power supply of a GPU chip as described in claim 4, characterized in that, The specific process of step S4 is as follows: S41. When the GPU is powered on, the GPU chip power management module monitors the power consumption of the GPU chip in real time according to the monitoring cycle T and initializes the registers. S42. Based on the currently acquired power consumption value, automatically classify it into the corresponding power consumption range group, perform communication bus addressing according to the associated register address, and send the corresponding register value to the power supply chip after finding it. S43. The power supply chip automatically matches the voltage transient change according to the register value, sends the result to the output controller and applies it; S44. If the power consumption of the GPU chip is detected to be >120% If the value is ≤0, activate protection mode, send an error interrupt command, and force shutdown.

Images