GPU power supply system and GPU power supply control method

Abstract

The invention discloses a GPU power supply system and a GPU power supply control method. The GPU power supply system comprises a photovoltaic energy storage inverter which comprises a direct current input end connected with a photovoltaic panel, an alternating current input end connected with a power grid and a power supply output end connected with a GPU; the monitoring scheduling module is provided with a plurality of data acquisition ports; the GPU power supply system has a first power supply mode, a second power supply mode and a third power supply mode, and the monitoring scheduling module is further provided with an instruction output port. The instruction output port is connected with the photovoltaic energy storage inverter so as to send a first instruction for switching to a first power supply mode or a second instruction for switching to a second power supply mode or a third instruction for switching to a third power supply mode to the photovoltaic energy storage inverter; wherein the monitoring scheduling module is used for controlling to send a first instruction, a second instruction or a third instruction to the photovoltaic energy storage inverter according to power supply parameters of a power grid, battery information and load information. The GPU is ensured to obtain stable and adaptive power input, and meanwhile power waste is avoided.

Description

Technical Field

[0001] The invention relates to a GPU (Graphic Process Unit) power supply system and a GPU power supply control method. Background Art

[0002] With the rapid development of computing-intensive technologies such as artificial intelligence, big data analysis, and deep learning, GPUs (Graphics Processing Units) are widely used in various data centers and computing clusters due to their powerful parallel computing capabilities. Traditional GPU power supply methods often rely solely on power grid power supply. However, the power demand of GPUs fluctuates greatly under different computing tasks, ranging from tens of watts at low load to hundreds of watts or even kilowatts at full load. Among them, when the GPU is running at high load, especially when multiple GPUs are concentrated in data centers and other scenarios, the GPU consumes a lot of electricity during operation. The sudden changes in its power demand and high power peaks put tremendous pressure on the power grid, which will cause huge instantaneous power demand on the power grid, easily causing problems such as grid voltage fluctuations, harmonic pollution, and insufficient power supply, affecting the stability of the power grid and interfering with the normal operation of other electrical equipment. At the same time, the traditional power supply model is difficult to accurately match the complex and changeable power demand of the GPU, resulting in low energy utilization efficiency.

[0003] Although there are some solutions that use capacitor-inductor filtering and basic voltage-grading adjustment devices. This simple power quality optimization measure can only roughly adjust the voltage, such as switching between several larger voltage ranges (such as ±10%), and cannot achieve accurate and stable voltage output. The harmonic suppression effect is also very limited, and the harmonic distortion rate may be as high as 10% or more. In such a power supply environment, the GPU is prone to calculation errors, performance degradation, and even hardware damage due to voltage fluctuations and electromagnetic interference, which reduces the reliability and operating efficiency of the entire computing power network, and there are obvious deficiencies in power quality assurance.

[0004] The above information disclosed in the Background section is only for enhancement of understanding of the background of the present application and therefore it may contain information that does not constitute the prior art that is already known to a person of ordinary skill in the art. Summary of the invention

[0005] The present invention provides a GPU power supply system and a GPU power supply control method, which ensure that the GPU obtains stable and adaptive power input while avoiding power waste; improves power grid stability, optimizes power grid burden, improves GPU operation stability and performance, and improves energy utilization efficiency and cost-effectiveness.

[0006] In order to achieve the above object, the present invention adopts the following technical scheme:

[0007] A GPU power supply system, comprising:

[0008] A photovoltaic energy storage inverter, comprising a DC input terminal configured to connect to a photovoltaic panel, an AC input terminal configured to connect to a power grid, and a power supply output terminal configured to connect to a GPU, wherein the photovoltaic energy storage inverter comprises a battery;

[0009] A monitoring and scheduling module having a plurality of data acquisition ports, wherein a first data acquisition port among the plurality of data acquisition ports is configured to be connected to a power grid management system to obtain power grid power supply parameters, a second data acquisition port is connected to the photovoltaic energy storage inverter to obtain battery information of the battery, and a third data acquisition port is configured to be connected to the GPU to obtain load information of the GPU;

[0010] The GPU power supply system has a first power supply mode, a second power supply mode and a third power supply mode, and the monitoring and scheduling module further has an instruction output port, and the instruction output port is connected to the photovoltaic energy storage inverter to send a first instruction to switch to the first power supply mode or a second instruction to switch to the second power supply mode or a third instruction to switch to the third power supply mode to the photovoltaic energy storage inverter;

[0011] Among them, the monitoring and scheduling module is used to control the sending of the first instruction, the second instruction or the third instruction to the photovoltaic energy storage inverter according to the power supply parameters of the power grid, the battery information and the load information.

[0012] In a preferred embodiment, the monitoring and scheduling module pre-stores a first set load value for the load of the GPU and a second set load value greater than the first set load value, and the monitoring and scheduling module pre-stores a first set power value for the power of the battery and a second set power value greater than the first set power value;

[0013] The monitoring and scheduling module is configured to send the first instruction to the photovoltaic energy storage inverter when the load of the GPU is less than the first set load value and the power of the battery is greater than the second set power value, the GPU power supply system is in the first power supply mode, and the photovoltaic energy storage inverter is configured to supply power to the GPU through the power of the battery in response to the first instruction;

[0014] The monitoring and scheduling module is configured to send the second instruction to the photovoltaic energy storage inverter when the load of the GPU is greater than the second set load value and the power of the battery is less than the first set power value, the GPU power supply system is in the second power supply mode, and the photovoltaic energy storage inverter is configured to respond to the second instruction and integrate the power of the battery and the power taken from the power grid to power the GPU;

[0015] The monitoring and scheduling module is configured to send the third instruction to the photovoltaic energy storage inverter when the power grid is in a low-power consumption period and / or the electricity price is low, the GPU power supply system is in the third power supply mode, and the photovoltaic energy storage inverter is configured to draw electricity from the power grid in response to the third instruction to power the GPU.

[0016] In a more preferred embodiment, the monitoring and scheduling module is configured to send the third instruction to the photovoltaic energy storage inverter when the power grid is in a low-power period, the electricity price is low, and the battery is in a rechargeable state, and the photovoltaic energy storage inverter is configured to draw electricity from the power grid in response to the third instruction to power the GPU and charge the battery.

[0017] In a more preferred embodiment, when the GPU power supply system is in the first power supply mode, the photovoltaic energy storage inverter is configured to convert the DC power of the battery into AC power to power the GPU, and the AC power of the power grid is supplemented or used as a backup.

[0018] In a more preferred embodiment, when the GPU power supply system is in the second power supply mode, the monitoring and scheduling module is configured to analyze the power supply capacity and stability of the power grid according to the received power grid power supply parameters, and calculate the power supply ratio of the power grid to the battery according to the analysis results, the second instruction includes the power supply ratio, and the photovoltaic energy storage inverter is configured to determine and integrate the battery power and the power grid power from the power grid according to the power supply ratio.

[0019] In a more preferred embodiment, when the GPU power supply system is in the third power supply mode, the third instruction is a charging instruction, the photovoltaic energy storage inverter is configured to draw power from the power grid, convert the AC power taken from the power grid into DC power to charge the battery, and use the AC power taken from the power grid to power the GPU.

[0020] In a preferred embodiment, the photovoltaic energy storage inverter further includes:

[0021] A DC / DC conversion circuit, which is used to convert the direct current input by the photovoltaic panel to charge the battery;

[0022] An AC / DC conversion circuit, which is used to convert the electricity from the power grid into direct current and then charge the battery;

[0023] The DC / AC conversion circuit is used to convert the battery power into alternating current to supply power to the GPU.

[0024] The present invention also adopts the following technical solution:

[0025] A power supply control method for a GPU power supply system comprises the following steps:

[0026] S100, making power supply decisions based on power grid power supply parameters, battery information and load information;

[0027] If the load of the GPU is less than the first set load value and the power of the battery is greater than the second set power value, jump to step S110: put the GPU power supply system in the first power supply mode, and supply power to the GPU through the power of the battery;

[0028] If the load of the GPU is greater than the second set load value and the power of the battery is less than the first set power value, jump to step S120: put the GPU power supply system into the second power supply mode, and integrate the power of the battery and the power taken from the power grid to power the GPU;

[0029] If the power grid is in a low-power period and / or the electricity price is low, jump to step S130: put the GPU power supply system in the third power supply mode to obtain power from the power grid to power the GPU.

[0030] In a preferred embodiment, step S110 specifically includes:

[0031] Step S111, the monitoring and scheduling module sends a first instruction to the photovoltaic energy storage inverter to draw power from the battery;

[0032] Step S112: The photovoltaic energy storage inverter converts the direct current of the battery into alternating current to power the GPU.

[0033] In a preferred embodiment, step S120 specifically includes:

[0034] Step S121, the monitoring and dispatching module analyzes the power supply capacity and stability of the power grid according to the power supply parameters of the power grid;

[0035] Step S122: the monitoring and scheduling module calculates the power supply ratio of the power grid and the battery according to the analysis result, and sends a second instruction to the photovoltaic energy storage inverter;

[0036] Step S123: The photovoltaic energy storage inverter integrates the AC power of the grid and the DC power of the battery according to the power supply ratio, and supplies power to the GPU after conversion.

[0037] In a preferred embodiment, if the power grid is in a low-power period and the electricity price is low and the battery can be charged, the process jumps to step S130; step S130 specifically includes:

[0038] Step S131, the monitoring and dispatching module enters the energy storage charging process during the low-peak period of the power grid;

[0039] Step S132: the monitoring and scheduling module sends a third charging instruction to the photovoltaic energy storage inverter;

[0040] Step S133, the photovoltaic energy storage inverter draws power from the power grid, converts the AC power from the power grid into DC power to charge the battery, and converts the AC power from the power grid to power the GPU.

[0041] In a preferred embodiment, the first set load value is 30% of the maximum load of the GPU, and the second set load value is 70% of the maximum load of the GPU; the first set power value is 20% of the maximum energy storage capacity of the battery, and the second set power value is 80% of the maximum energy storage capacity of the battery.

[0042] The above solution adopted by the present invention has the following advantages:

[0043] In the GPU power supply system and GPU power supply control method of the present invention, when the GPU computing demand is at a peak, such as the large-scale artificial intelligence model training stage, the photovoltaic energy storage inverter can release electricity from the battery to power the GPU, reduce the instantaneous power demand on the power grid, avoid peak loads on the power grid, and play a role in peak shaving; and during the low load period of the power grid, such as late at night, low-priced electricity can be used to charge the battery to achieve valley filling, optimize the load curve of the power grid, and improve the stability and economy of the power grid operation. Based on the optimization effect of the energy storage inverter on power quality, such as reducing voltage sags and suppressing harmonics, the risk of GPU calculation errors or hardware damage caused by power quality problems is reduced, the reliability and accuracy of GPU calculations are improved, and it helps to improve the success rate and accuracy of the entire computing network in complex task processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solution of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0045] Figure 1 A structural block diagram of a GPU power supply system according to an embodiment of the present invention is shown.

[0046] Figure 2 A structural block diagram of a photovoltaic energy storage inverter according to an embodiment of the present invention is shown.

[0047] Figure 3 A flow chart of a power supply control method according to an embodiment of the present invention is shown.

[0048] in:

[0049] 100, GPU; 200, photovoltaic energy storage inverter; 300, power grid; 400, photovoltaic panel; 500, monitoring and scheduling module;

[0050] 201. Battery; 202. DC / DC conversion circuit; 203. AC / DC conversion circuit; 204. DC / AC conversion circuit; 301. Power grid management system. DETAILED DESCRIPTION

[0051] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0052] The power demand of GPU fluctuates greatly under different computing tasks, ranging from tens of watts at low load to hundreds of watts or even kilowatts at full load. On the other hand, energy storage technology has made great progress in recent years. Photovoltaic energy storage inverters, as key interface devices between energy storage systems and power grids or power-consuming equipment, have multiple functions such as power conversion, energy storage and release control. However, the application of energy storage inverters in computing power networks has not yet been fully explored, and there is huge room for innovation.

[0053] Based on this, this embodiment proposes a computing network construction scheme that is conducive to photovoltaic energy storage inverter driving GPU to reduce the load on the power grid. However, the photovoltaic energy storage inverter needs to have an extremely wide power adjustment range and an extremely fast response speed in order to accurately match the power changes of the GPU. Otherwise, insufficient power supply may cause the GPU to reduce the frequency of operation, affecting the computing efficiency, or excessive power supply may cause energy waste and equipment overheating. In this regard, in this embodiment, when the GPU computing demand is at a peak, such as the large-scale artificial intelligence model training stage, the photovoltaic energy storage inverter can release electricity from the energy storage system (battery) to power the GPU, reduce the instantaneous power demand on the power grid, avoid peak loads on the power grid, and play a role in peak shaving. During the low load period of the power grid, such as late at night, low-priced electricity can be used to charge the battery to fill the valley, optimize the load curve of the power grid, and improve the stability and economy of the power grid operation.

[0054] At the same time, due to the optimization effect of photovoltaic energy storage inverters on power quality, such as reducing voltage sags and suppressing harmonics, the risk of GPU calculation errors or hardware damage caused by power quality problems is reduced, the reliability and accuracy of GPU calculations are improved, and it helps to improve the success rate and accuracy of the entire computing network in complex task processing.

[0055] Reference Figure 1 and Figure 2As shown, the GPU power supply system of this embodiment includes a photovoltaic energy storage inverter 200 and a monitoring and scheduling module 500. The photovoltaic energy storage inverter 200 has a DC input terminal, an AC input terminal and a power supply output terminal. Figure 1 In the figure, the thick line with an arrow indicates the direction of power transmission, and the thin line with an arrow indicates the direction of signal transmission. The DC input terminal is connected to the photovoltaic panel 400 (PV panel), and the photovoltaic panel 400 includes a solar cell 201 that can convert solar energy into electrical energy. The DC power generated by the photoelectric conversion of the PV panel is transmitted to the photovoltaic energy storage inverter 200 for storage or conversion. The AC input terminal is connected to the power grid 300, and can extract AC power from the power grid 300. The power supply output terminal is connected to the GPU 100 to supply power to the GPU 100; in this embodiment, the power supply output terminal outputs AC power and provides AC power to the GPU 100. The power grid 300 is used to transmit city power. The power grid 300 is monitored by the power grid management system 301, and parameters such as the power supply voltage, frequency, and current power supply load can be obtained.

[0056] As a key energy storage unit of the photovoltaic energy storage inverter 200, the battery 201 is composed of a plurality of battery 201 groups. It can convert the AC power of the power grid 300 into DC power and store it through the energy storage inverter when the power grid 300 has sufficient power or the power price is low (such as the low power consumption period at night). The energy storage inverter plays the role of a bridge for bidirectional power conversion. It can not only rectify the AC power of the power grid 300 into DC power suitable for charging the battery 201 group, but also invert the DC power of the battery 201 into AC power that meets the requirements of the GPU 100 when it is necessary to power the GPU 100.

[0057] The photovoltaic energy storage inverter 200 is connected to the GPU 100 cluster through a specific power transmission line and a control interface. This connection method can realize the precise control of the photovoltaic energy storage inverter 200 on the power supply of the GPU 100, dynamically adjust the output voltage, current and power according to the real-time computing load of the GPU 100, ensure that the GPU 100 obtains a stable and adaptive power input, and avoid power waste. The photovoltaic energy storage inverter 200 integrates a power quality management module, which can filter and stabilize the input and output power, reduce the interference of harmonics and voltage fluctuations generated during power conversion and transmission on the operation of the GPU 100, and improve the accuracy and reliability of the GPU 100 calculation. The photovoltaic energy storage inverter 200 has a bidirectional energy flow capability, which can not only supply the power in the battery 201 to the GPU 100, but also under certain conditions (such as when the GPU 100 is under low load and the energy storage system is insufficient), the excess power of the GPU 100 is fed back to the power grid 300, realizing the bidirectional optimization of energy utilization.

[0058] The monitoring and scheduling module 500 monitors the multi-dimensional data such as the computing task volume of GPU 100, power consumption trend and real-time load status of power grid 300 in real time. Through in-depth analysis and prediction of these data, photovoltaic energy storage inverter 200 can plan the power distribution strategy in advance, and use battery 201 to store excess power when the load of power grid 300 is low; when the computing demand of GPU 100 increases and the load of power grid 300 is high, GPU 100 is powered from battery 201 first, thereby effectively reducing the instantaneous power impact of GPU 100 on power grid 300 and achieving the balance of the burden of power grid 300. At the same time, it has self-learning and optimization functions, and can continuously adjust the power scheduling parameters according to long-term operation data to adapt to the power supply needs of GPU 100 in different application scenarios and power environments, and further improve energy utilization efficiency and the stability of power grid 300.

[0059] The monitoring and scheduling module 500 has multiple data acquisition ports, the first data acquisition port of which is configured to connect to the power grid management system 301 to obtain the power supply parameters of the power grid 300, the second data acquisition port is connected to the photovoltaic energy storage inverter 200 to obtain the battery information of the battery 201, and the third data acquisition port is configured to connect to the GPU 100 to obtain the load information of the GPU 100. The monitoring and scheduling module 500 is electrically connected to the power grid management system 301, the photovoltaic energy storage inverter 200 and the GPU 100. The monitoring and scheduling module 500 is the intelligent brain of the entire system. It communicates with the GPU 100, the photovoltaic energy storage inverter 200 and the power grid management system 301 in real time to obtain key data from various aspects. For the GPU 100, its load condition is monitored, including information such as the amount of computing tasks and real-time power consumption; for the photovoltaic energy storage inverter 200, the power state of the battery 201 and the battery health status are mastered; for the power grid 300, the power supply voltage, frequency, current load and other parameters are understood. Based on these large amounts of real-time data, the monitoring and scheduling module 500 uses specific algorithms and logics for analysis and processing.

[0060] The GPU power supply system has a first power supply mode, a second power supply mode and a third power supply mode. The monitoring and scheduling module 500 also has an instruction output port, which is connected to the photovoltaic energy storage inverter 200 to send a first instruction to switch to the first power supply mode, a second instruction to switch to the second power supply mode, or a third instruction to switch to the third power supply mode to the photovoltaic energy storage inverter 200. The monitoring and scheduling module 500 is used to control the sending of the first instruction, the second instruction or the third instruction to the photovoltaic energy storage inverter 200 according to the power supply parameters of the power grid, the battery information and the load information.

[0061] The monitoring and scheduling module 500 pre-stores a first set load value for the load of the GPU 100 and a second set load value greater than the first set load value, and the monitoring and scheduling module 500 pre-stores a first set power value for the power of the battery 201 and a second set power value greater than the first set power value. These set values ​​can be set by the user and input into the system. Among them, the first set load value is 30% of the maximum load (full load) of the GPU 100, and the second set load value is 70% of the maximum load of the GPU 100. The first set power value is 20% of the maximum energy storage power of the battery 201, and the second set power value is 80% of the maximum energy storage power of the battery 201.

[0062] The monitoring and scheduling module 500 is configured to send a first instruction to the photovoltaic energy storage inverter 200 when the load of the GPU 100 is less than a first set load value and the power of the battery 201 is greater than a second set power value. The GPU power supply system is in a first power supply mode, and the photovoltaic energy storage inverter 200 is configured to supply power to the GPU 100 through the power of the battery 201 in response to the first instruction.

[0063] The monitoring and scheduling module 500 is configured to send a second instruction to the photovoltaic energy storage inverter 200 when the load of the GPU 100 is greater than the second set load value and the power of the battery 201 is less than the first set power value. The GPU power supply system is in the second power supply mode, and the photovoltaic energy storage inverter 200 is configured to respond to the second instruction and integrate the power of the battery 201 and the power taken from the power grid 300 to power the GPU 100.

[0064] The monitoring and scheduling module 500 is configured to send a third instruction to the photovoltaic energy storage inverter 200 when the power grid 300 is in a low power consumption period and / or the electricity price is low, the GPU power supply system is in a third power supply mode, and the photovoltaic energy storage inverter 200 is configured to respond to the third instruction and take power from the power grid 300 to power the GPU 100. Further, the monitoring and scheduling module 500 is configured to send a third instruction to the photovoltaic energy storage inverter 200 when the power grid 300 is in a low power consumption period, the electricity price is low, and the battery 201 is in a chargeable state, and the photovoltaic energy storage inverter 200 is configured to respond to the third instruction and take power from the power grid 300 to power the GPU 100 and charge the battery 201.

[0065] When the GPU power supply system is in the first power supply mode, the photovoltaic energy storage inverter 200 is configured to convert the direct current of the battery 201 into alternating current to power the GPU 100, and the alternating current of the power grid 300 is supplemented or used as a backup. When the GPU power supply system is in the second power supply mode, the monitoring and scheduling module 500 is configured to analyze the power supply capacity and stability of the power grid 300 according to the received power supply parameters of the power grid 300, and calculate the power supply ratio of the power grid 300 and the battery 201 according to the analysis results. The second instruction includes the power supply ratio, and the photovoltaic energy storage inverter 200 is configured to determine and integrate the power of the battery 201 and the power of the power grid 300 from the power grid 300 according to the power supply ratio. When the GPU power supply system is in the third power supply mode, the third instruction is a charging instruction, and the photovoltaic energy storage inverter 200 is configured to draw power from the power grid 300, convert the alternating current taken from the power grid 300 into direct current to charge the battery 201, and use the alternating current taken from the power grid 300 to power the GPU 100.

[0066] like Figure 2 As shown, the photovoltaic energy storage inverter 200 also includes a DC / DC conversion circuit 202, an AC / DC conversion circuit 203 and a DC / AC conversion circuit 204. The DC / DC conversion circuit 202 is connected to the battery 201 and is configured to be connected to the output end of the photovoltaic panel 400, so as to convert the direct current input by the photovoltaic panel 400 to charge the battery 201. The AC / DC conversion circuit 203 is connected to the battery 201 and is configured to be connected to the power grid 300, so as to convert the power from the power grid 300 to direct current and charge the battery 201. The DC / AC conversion circuit 204 is connected to the battery 201 and is connected to the GPU 100, so as to convert the power of the battery 201 to alternating current and supply power to the GPU 100. The DC / DC conversion circuit 202, the AC / DC conversion circuit 203 and the DC / AC conversion circuit 204 can use known circuit modules, and the circuit structures of these circuits are not the key points of the invention of this application and will not be described in detail.

[0067] The power supply principle of the above GPU power supply system is described as follows.

[0068] When the GPU 100 is running at low load and the battery 201 has sufficient power, the monitoring and scheduling module 500 determines that the battery 201 should be used for power supply first according to the pre-set rules and algorithms to reduce the dependence on the power grid 300. It sends instructions to the photovoltaic energy storage inverter 200, and the photovoltaic energy storage inverter 200 immediately starts the process of taking power from the battery 201 and converting it into AC power to power the GPU 100. This process is based on the fact that the electric energy stored in the energy storage unit can meet the energy consumption requirements of the low-load GPU 100, and by reducing the power taken from the power grid 300, the power supply pressure of the power grid 300 during this period is reduced, a local "off-grid" power supply mode is realized, and energy distribution is optimized.

[0069] When the GPU 100 is under high load, if the battery 201 is insufficient, the monitoring and scheduling module 500 will quickly evaluate the power supply capacity and stability of the power grid 300. For example, analyze whether the current voltage fluctuation range of the power grid 300 is within an acceptable range, whether the frequency is stable, and the overall load situation. Based on these evaluation results, the optimal ratio of taking power from the power grid 300 and replenishing power from the battery 201 is calculated according to the intelligent algorithm. Comprehensively consider multiple factors such as the bearing capacity of the power grid 300, the remaining power of the battery 201, and the stable power supply requirements required by the GPU 100. Then, this ratio instruction is sent to the photovoltaic energy storage inverter 200. The photovoltaic energy storage inverter 200 obtains the corresponding proportion of AC power from the power grid 300 according to the instruction, and obtains DC power from the battery 201 at the same time, and processes the two through its internal power conversion and integration circuit to output stable AC power to power the GPU 100. This method not only meets the large amount of power required for high-load operation of GPU 100, but also cleverly balances the burden on power grid 300 and the utilization of battery 201, avoiding the occurrence of voltage drops or other instabilities in power grid 300 due to the instantaneous high power demand of GPU 100.

[0070] During the entire power supply process, the power quality optimization function of the photovoltaic energy storage inverter 200 continues to play a role. Its internal voltage regulation module monitors the output voltage in real time through a closed-loop feedback control system and compares it with the preset standard voltage value. Once a voltage deviation is found, the internal power electronic devices (such as the on and off time of the power transistor) are immediately adjusted to accurately stabilize the output voltage within a very small fluctuation range (for example, ±3%), ensuring that the GPU 100 receives a stable and reliable voltage input, avoiding calculation errors or performance degradation of the GPU 100 caused by voltage fluctuations. At the same time, the harmonic suppression module uses advanced filtering technology to perform spectrum analysis on the input power, identify the harmonic components therein, and filter out the harmonics through filters (such as LC filters, active power filters, etc.), so that the harmonic distortion rate of the output power is kept at an extremely low level (such as less than 3%). This is because harmonics will cause electromagnetic interference to the GPU device and affect its normal operation. By reducing the harmonic distortion rate, a pure electromagnetic environment is created for the GPU 100, ensuring its stable and efficient operation and extending the service life of the GPU 100.

[0071] In addition, the monitoring and scheduling module 500 is also responsible for the feedback and adjustment mechanism of the entire system. During operation, the GPU 100 will regularly feedback its own operating status information to the monitoring and scheduling module 500, such as the execution progress of the computing task, the change of computing efficiency, and the fluctuation of the internal temperature of the device. The photovoltaic energy storage inverter 200 will also feedback its power changes and the health status information of the battery 201 to the monitoring and scheduling module 500, and the power grid management system 301 will continue to provide the real-time power supply parameter changes of the power grid 300, such as voltage fluctuations, frequency drifts, load changes, etc. The monitoring and scheduling module 500 comprehensively summarizes and deeply analyzes these feedback information at regular intervals (such as 10 minutes) and re-evaluates the operating status of the entire system. If it is found that the current power supply strategy is no longer the optimal solution due to changes in the GPU 100 load, fluctuations in the power grid 300 state, or changes in the performance of the battery 201, for example, it is found that the operating efficiency of the GPU 100 may be reduced because the quality of the power supply voltage is slightly reduced or the power supply ratio is unreasonable, or the state of the power grid 300 has changed significantly and affected the power supply stability, etc., the monitoring and scheduling module 500 will immediately recalculate and adjust the intelligent power supply strategy. This includes re-determining the power supply mode of the photovoltaic energy storage inverter 200, such as changing the proportion of power drawn from the power grid 300 and the battery 201, adjusting the voltage regulation parameters to further optimize the power quality, etc., and sending the adjusted instructions to the photovoltaic energy storage inverter 200. After receiving the new instructions, the photovoltaic energy storage inverter 200 quickly adjusts its internal power conversion and control circuits, and changes the power supply mode according to the new requirements, so that the entire computing power network power supply system can be dynamically optimized and adaptively adjusted according to the actual operating conditions, and always maintain the optimal operating state to meet the ever-changing computing power requirements and maintain the stable operation of the power grid 300.

[0072] Reference Figure 3 As shown, this embodiment provides a power supply control method for a GPU power supply system. The GPU power supply system can be Figure 1 and Figure 2 The power supply control method comprises the following steps:

[0073] S100, making power supply decisions based on power grid power supply parameters, battery information and load information;

[0074] If the load of the GPU is less than the first set load value (30%) and the battery power is greater than the second set power value (80%), jump to step S110: put the GPU power supply system in the first power supply mode, and supply power to the GPU through the battery power;

[0075] If the GPU load is greater than the second set load value (70%) and the battery power is less than the first set power value (20%), jump to step S120: put the GPU power supply system into the second power supply mode, and integrate the battery power and the power taken from the power grid to power the GPU;

[0076] If the power grid is in a low-power period and / or the electricity price is low, jump to step S130: the GPU power supply system is placed in a third power supply mode to obtain power from the grid to power the GPU.

[0077] Step S110 specifically includes:

[0078] Step S111, the monitoring and scheduling module sends a first instruction to the photovoltaic energy storage inverter to draw power from the battery;

[0079] Step S112: The photovoltaic energy storage inverter converts the direct current of the battery into alternating current to power the GPU.

[0080] Step S120 specifically includes:

[0081] Step S121, the monitoring and dispatching module analyzes the power supply capacity and stability of the power grid according to the power supply parameters of the power grid;

[0082] Step S122: the monitoring and scheduling module calculates the power supply ratio of the power grid and the battery according to the analysis result, and sends a second instruction to the photovoltaic energy storage inverter;

[0083] Step S123: The photovoltaic energy storage inverter integrates the AC power of the grid and the DC power of the battery according to the power supply ratio, and supplies power to the GPU after conversion.

[0084] If the power grid is in a low-power period and the electricity price is low and the battery can be charged, then jump to step S130. Step S130 specifically includes:

[0085] Step S131, the monitoring and dispatching module enters the energy storage charging process during the low-peak period of the power grid;

[0086] Step S132: the monitoring and scheduling module sends a third charging instruction to the photovoltaic energy storage inverter;

[0087] Step S133: The photovoltaic energy storage inverter draws power from the power grid, converts the AC power from the power grid into DC power to charge the battery, and converts the AC power from the power grid into DC power to power the GPU.

[0088] The GPU power supply system and GPU power supply control method of this embodiment have the following advantages.

[0089] 1. Energy management and grid load reduction

[0090] Traditional GPU power supply methods often rely solely on power grid power supply. When the GPU is running at high load, especially when multiple GPUs are concentrated in data centers and other scenarios, it will cause huge instantaneous power demand on the power grid, causing problems such as grid voltage fluctuations and insufficient power supply, affecting the stability of the power grid and possibly interfering with other power-consuming equipment. This embodiment introduces a photovoltaic energy storage inverter with a battery energy storage unit, and realizes the dynamic allocation of energy storage units and power grid power supply through a monitoring and scheduling module. When the GPU is under low load and the energy storage is sufficient, energy storage power supply is used first to reduce power grid power consumption; when the load is high and the energy storage is insufficient, the proportion of power grid and energy storage power supply is reasonably allocated according to the power grid conditions, which effectively smooths the power impact of the GPU on the power grid, reduces the burden on the power grid, improves the reliability and stability of the power grid operation, and realizes efficient energy management and "peak shaving and valley filling". Compared with traditional methods, it greatly optimizes the energy utilization structure and reduces the pressure on the power grid.

[0091] 2. Power quality assurance

[0092] As a high-precision, high-performance computing device, the GPU has extremely strict requirements on power quality. Traditional power supply methods are difficult to ensure stable voltage, frequency and low harmonic interference power supply, which can easily lead to GPU calculation errors, performance degradation and even hardware damage. The photovoltaic energy storage inverter in this embodiment has a powerful power quality optimization function, which can accurately stabilize the output voltage within a very small fluctuation range (such as ±3%), strictly control the frequency deviation, and use high-performance filters to effectively filter out harmonics, so that the harmonic distortion rate is reduced to an extremely low level (such as less than 3%). This high-quality power supply ensures that the GPU always operates in a stable and pure power environment, significantly improves the calculation accuracy and operation efficiency of the GPU, and extends its service life. Compared with traditional power supply, there has been a qualitative leap in power quality assurance, which has greatly improved the reliability and stability of the entire computing network.

[0093] 3. Intelligent monitoring and dynamic adjustment

[0094] Previous GPU power supply systems lacked a comprehensive and intelligent monitoring and dynamic adjustment mechanism. The monitoring and scheduling module of this embodiment can collect key information such as GPU load conditions, energy storage unit power status, and power supply parameters of the power grid in real time and in all directions, and conduct in-depth analysis and intelligent decision-making based on these data. By predicting the trend of GPU load changes, the charging and discharging strategy of the energy storage unit can be planned in advance, and the valley charging period of the energy storage can be flexibly arranged according to the fluctuation of the power grid electricity price. At the same time, the operating status of the entire system can be continuously evaluated and the power supply strategy can be dynamically adjusted. For example, at regular intervals (such as 10 minutes), the current power supply mode is reviewed based on the latest feedback data to see whether it is optimal. If a problem is found, the power supply source and proportion of the photovoltaic energy storage inverter are immediately adjusted. This intelligent dynamic adjustment capability enables the system to adapt to various complex and changeable operating scenarios. Whether it is a sudden increase or decrease in the amount of GPU tasks, fluctuations in the power supply status of the power grid, or changes in the performance of the energy storage unit, it can respond and optimize in time to ensure that the entire computing network is always in an efficient and stable operating state. This is a major improvement in intelligent management and adaptability compared to traditional GPU power supply systems.

[0095] As shown in this specification and claims, the terms "comprises" and "include" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list, and the method or device may also include other steps or elements. The term "and / or" as used herein includes any combination of one or more of the relevant listed items.

[0096] It should be noted that, unless otherwise specified, when a feature is referred to as being "fixed" or "connected" to another feature, it may be directly fixed or connected to the other feature, or it may be indirectly fixed or connected to the other feature. In addition, the descriptions of up, down, left, right, etc. used in the present invention are only relative to the relative positional relationship of the components of the present invention in the drawings.

[0097] The above embodiment is only for illustrating the technical concept and features of the present invention, and is a preferred embodiment. Its purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and it cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A GPU power supply system, characterized in that: include: A photovoltaic energy storage inverter, comprising a DC input terminal configured to connect to a photovoltaic panel, an AC input terminal configured to connect to a power grid, and a power supply output terminal configured to connect to a GPU, wherein the photovoltaic energy storage inverter comprises a battery; A monitoring and scheduling module having a plurality of data acquisition ports, wherein a first data acquisition port among the plurality of data acquisition ports is configured to be connected to a power grid management system to obtain power grid power supply parameters, a second data acquisition port is connected to the photovoltaic energy storage inverter to obtain battery information of the battery, and a third data acquisition port is configured to be connected to the GPU to obtain load information of the GPU; The GPU power supply system has a first power supply mode, a second power supply mode and a third power supply mode, and the monitoring and scheduling module further has an instruction output port, and the instruction output port is connected to the photovoltaic energy storage inverter to send a first instruction to switch to the first power supply mode or a second instruction to switch to the second power supply mode or a third instruction to switch to the third power supply mode to the photovoltaic energy storage inverter; Among them, the monitoring and scheduling module is used to control the sending of the first instruction, the second instruction or the third instruction to the photovoltaic energy storage inverter according to the power supply parameters of the power grid, the battery information and the load information.

2. The GPU power supply system according to claim 1, characterized in that: The monitoring and scheduling module pre-stores a first set load value for the load of the GPU and a second set load value greater than the first set load value, and the monitoring and scheduling module pre-stores a first set power value for the power of the battery and a second set power value greater than the first set power value; The monitoring and scheduling module is configured to send the first instruction to the photovoltaic energy storage inverter when the load of the GPU is less than the first set load value and the power of the battery is greater than the second set power value, the GPU power supply system is in the first power supply mode, and the photovoltaic energy storage inverter is configured to supply power to the GPU through the power of the battery in response to the first instruction; The monitoring and scheduling module is configured to send the second instruction to the photovoltaic energy storage inverter when the load of the GPU is greater than the second set load value and the power of the battery is less than the first set power value, the GPU power supply system is in the second power supply mode, and the photovoltaic energy storage inverter is configured to respond to the second instruction and integrate the power of the battery and the power taken from the power grid to power the GPU; The monitoring and scheduling module is configured to send the third instruction to the photovoltaic energy storage inverter when the power grid is in a low-power consumption period and / or the electricity price is low, the GPU power supply system is in the third power supply mode, and the photovoltaic energy storage inverter is configured to draw electricity from the power grid in response to the third instruction to power the GPU.

3. The GPU power supply system according to claim 2, characterized in that: The monitoring and scheduling module is configured to send the third instruction to the photovoltaic energy storage inverter when the power grid is in a low-power period, the electricity price is low, and the battery is in a chargeable state. The photovoltaic energy storage inverter is configured to draw electricity from the power grid in response to the third instruction to power the GPU and charge the battery.

4. The GPU power supply system according to claim 2, characterized in that: When the GPU power supply system is in the first power supply mode, the photovoltaic energy storage inverter is configured to convert the direct current of the battery into alternating current to power the GPU, and the alternating current of the power grid is used for supplementation or backup.

5. The GPU power supply system according to claim 2, characterized in that: When the GPU power supply system is in the second power supply mode, the monitoring and scheduling module is configured to analyze the power supply capacity and stability of the power grid according to the received power grid power supply parameters, and calculate the power supply ratio of the power grid to the battery according to the analysis results. The second instruction includes the power supply ratio, and the photovoltaic energy storage inverter is configured to determine and integrate the battery power and the power grid power from the power grid according to the power supply ratio.

6. The GPU power supply system according to claim 2, characterized in that: When the GPU power supply system is in the third power supply mode, the third instruction is a charging instruction, the photovoltaic energy storage inverter is configured to draw power from the power grid, convert the AC power taken from the power grid into DC power to charge the battery, and use the AC power taken from the power grid to power the GPU.

7. The GPU power supply system according to any one of claims 1 to 6, characterized in that: The photovoltaic energy storage inverter also includes: A DC / DC conversion circuit, which is used to convert the direct current input by the photovoltaic panel to charge the battery; An AC / DC conversion circuit, which is used to convert the electricity from the power grid into direct current and then charge the battery; The DC / AC conversion circuit is used to convert the battery power into alternating current to supply power to the GPU.

8. A GPU power supply control method, characterized in that: The steps include: S100, making power supply decisions based on power grid power supply parameters, battery information and load information; If the load of the GPU is less than the first set load value and the power of the battery is greater than the second set power value, jump to step S110: put the GPU power supply system in the first power supply mode, and supply power to the GPU through the power of the battery; If the load of the GPU is greater than the second set load value and the power of the battery is less than the first set power value, jump to step S120: put the GPU power supply system into the second power supply mode, and integrate the power of the battery and the power taken from the power grid to power the GPU; If the power grid is in a low-power period and / or the electricity price is low, jump to step S130: put the GPU power supply system in the third power supply mode to obtain power from the power grid to power the GPU.

9. The GPU power supply control method according to claim 8, characterized in that: Step S110 specifically includes: Step S111, the monitoring and scheduling module sends a first instruction to the photovoltaic energy storage inverter to draw power from the battery; Step S112: The photovoltaic energy storage inverter converts the direct current of the battery into alternating current to power the GPU.

10. The GPU power supply control method according to claim 8, characterized in that: Step S120 specifically includes: Step S121, the monitoring and dispatching module analyzes the power supply capacity and stability of the power grid according to the power supply parameters of the power grid; Step S122: the monitoring and scheduling module calculates the power supply ratio of the power grid and the battery according to the analysis result, and sends a second instruction to the photovoltaic energy storage inverter; Step S123: The photovoltaic energy storage inverter integrates the AC power of the grid and the DC power of the battery according to the power supply ratio, and supplies power to the GPU after conversion.

11. The GPU power supply control method according to claim 8, characterized in that: If the power grid is in a low-power period and the electricity price is low and the battery can be charged, jump to step S130; step S130 specifically includes: Step S131, the monitoring and dispatching module enters the energy storage charging process during the low-peak period of the power grid; Step S132: the monitoring and scheduling module sends a third charging instruction to the photovoltaic energy storage inverter; Step S133, the photovoltaic energy storage inverter draws power from the power grid, converts the AC power from the power grid into DC power to charge the battery, and converts the AC power from the power grid to power the GPU.

12. The GPU power supply control method according to claim 8, characterized in that: The first set load value is 30% of the maximum load of the GPU, and the second set load value is 70% of the maximum load of the GPU; the first set power value is 20% of the maximum energy storage capacity of the battery, and the second set power value is 80% of the maximum energy storage capacity of the battery.

Images