Control device and control method

Abstract

The present invention provides a control device for controlling a fan of a server system. The control device includes a temperature predictor and a fan controller. The temperature predictor is configured to use a first deep learning model and generate a predicted temperature of the chip according to a rotation speed of a fan, a heat generation power of an adapter card, an inlet temperature of the adapter card, and a temperature of a processor. The fan controller is coupled to the temperature predictor and is configured to use a second deep learning model and generate a target rotation speed of the fan according to the predicted temperature of the chip, the rotation speed of the fan, and the heat generation power of the adapter card, and generate a fan control signal according to the target rotation speed of the fan to drive the fan.

Description

Technical Field

[0001] The present invention relates to a control device and a control method, and in particular to a control device and a control method capable of selecting a target speed of a fan based on a deep learning model. Background Art

[0002] When a server system is in operation, it will generate waste heat, so a cooling fan is usually configured to transfer the heat source and discharge the internal heat energy. Currently, the control of the cooling fan of the server is usually to read the temperature value of the component and then calculate the appropriate speed according to the temperature reading. For example, it is common to use a proportional integral derivative (PID) controller to control the operation state of the cooling fan. However, in the server system, there are often components or chips that cannot send back temperature readings, but the fan must still be controlled to maintain the temperature of the chip from overheating. For example, the server system contains a high-speed peripheral component interconnect express (PCIe) adapter card (also known as a high-speed serial computer bus adapter card). When the chip set on the PCIe adapter card cannot send back the temperature value, the current method uses the inlet temperature and inlet flow of the PCIe adapter card as a reference to control the fan to avoid overheating of the chip component. However, if the configuration in the server system is slightly different, such as a change in the amount of memory, a different type of hard disk device, and the addition of additional adapter cards or hardware, the actual inlet flow rate will change. In this case, since the parameters of the inlet flow rate are related to the flow rate inside the overall server system and cannot correctly reflect the temperature requirements of the chip on the adapter card, it is not suitable for fan control. However, if the controller is designed based solely on the inlet temperature, in order to avoid overheating caused by insufficient flow, the usual practice is to adopt an overcooling design, which will cause the fan to run continuously and cause unnecessary power consumption. Therefore, the conventional technology really needs to be improved. Summary of the invention

[0003] In order to solve the above problems, the present invention provides a control device and a control method for selecting a target speed of a fan based on a deep learning model to solve the above problems.

[0004] The present invention provides a control device for controlling a fan of a server system, comprising: a temperature predictor, configured to use a first deep learning model and generate a predicted temperature of a chip based on a rotational speed of the fan, a heat generation power of an adapter card, an inlet temperature of the adapter card and a temperature of a processor; and a fan controller, coupled to the temperature predictor, configured to use a second deep learning model and generate a target rotational speed of the fan based on the predicted temperature of the chip, the rotational speed of the fan and the heat generation power of the adapter card, and generate a fan control signal according to the target rotational speed of the fan to drive the fan.

[0005] The temperature predictor includes a plurality of first neural network layers to form the first deep learning model, and the fan controller includes a plurality of second neural network layers to form the second deep learning model.

[0006] The rotation speed of the fan includes a current rotation speed of the fan and a plurality of previous rotation speeds corresponding to a plurality of previous cycles, the heating power of the adapter card includes a current heating power of the adapter card and a plurality of previous heating powers corresponding to the plurality of previous cycles, the inlet temperature of the adapter card includes a current inlet temperature of the adapter card and a plurality of previous inlet temperatures corresponding to the plurality of previous cycles, and the temperature of the processor includes a current processor temperature of the processor and a plurality of previous processor temperatures corresponding to the plurality of previous cycles.

[0007] The fan controller uses the second deep learning model and generates a plurality of candidate future temperatures corresponding to a plurality of candidate rotational speeds according to the predicted temperature of the chip, the rotational speed of the fan, and the heat generation power of the adapter card. The fan controller compares the plurality of candidate future temperatures with a set point temperature respectively, and selects the candidate future temperature with the smallest absolute difference between the plurality of candidate future temperatures and the set point temperature as a future target temperature, and determines the candidate rotational speed corresponding to the selected future target temperature as the target rotational speed of the fan, and generates the fan control signal according to the target rotational speed of the fan to drive the fan.

[0008] The first deep learning model and the second deep learning model are the same deep learning model.

[0009] The present invention further provides a control method for controlling a fan of a server system, comprising: using a first deep learning model to generate a predicted temperature of a chip based on a rotational speed of the fan, a heat generation power of an adapter card, an inlet temperature of the adapter card, and a temperature of a processor; and using a second deep learning model to generate a target rotational speed of the fan based on the predicted temperature of the chip, the rotational speed of the fan, and the heat generation power of the adapter card, and generating a fan control signal based on the target rotational speed of the fan to drive the fan.

[0010] The first deep learning model is formed by a plurality of first neural network layers, and the second deep learning model is formed by a plurality of second neural network layers.

[0011] The rotation speed of the fan includes a current rotation speed of the fan and a plurality of previous rotation speeds corresponding to a plurality of previous cycles, the heating power of the adapter card includes a current heating power of the adapter card and a plurality of previous heating powers corresponding to the plurality of previous cycles, the inlet temperature of the adapter card includes a current inlet temperature of the adapter card and a plurality of previous inlet temperatures corresponding to the plurality of previous cycles, and the temperature of the processor includes a current processor temperature of the processor and a plurality of previous processor temperatures corresponding to the plurality of previous cycles.

[0012] The steps of using the second deep learning model and generating the target speed of the fan according to the predicted temperature of the chip, the speed of the fan and the heating power of the adapter card, and generating the fan control signal according to the target speed of the fan to drive the fan include: using the second deep learning model and generating a plurality of candidate future temperatures corresponding to a plurality of candidate speeds according to the predicted temperature of the chip, the speed of the fan and the heating power of the adapter card; comparing the plurality of candidate future temperatures with a set point temperature respectively and selecting the candidate future temperature with the smallest absolute difference between the plurality of candidate future temperatures and the set point temperature as a future target temperature; determining the candidate speed corresponding to the selected future target temperature as the target speed of the fan; and generating the fan control signal according to the target speed of the fan to drive the fan.

[0013] The first deep learning model and the second deep learning model are the same deep learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 FIG. 4 is a schematic diagram of a server system according to an embodiment of the present invention.

[0015] Figure 2 It is a schematic diagram of a process of an embodiment of the present invention.

[0016] Figure 3 for Figure 1 Schematic diagram of an embodiment of a temperature sensor for detecting inlet temperature.

[0017] Figure 4 for Figure 1 Schematic diagram of an embodiment of a temperature predictor in FIG.

[0018] Figure 5 for Figure 1 Schematic diagram of an embodiment of a fan controller in FIG.

[0019] Explanation of symbols:

[0020] 1: Server system

[0021] 2: Process

[0022] 10: Processor

[0023] 102,50: Temperature sensor

[0024] 20: Adapter card

[0025] 202: Chip

[0026] 30: Control device

[0027] 302: Temperature Predictor

[0028] 304: Fan controller

[0029] 40: Fan

[0030] 60: Speed ​​sensor

[0031] DL1: The first deep learning model

[0032] DL2: Second deep learning model

[0033] FS i :Candidate speed

[0034] FS t ,FS t-18 :Speed

[0035] FT t ,FT t-1 ,FT t-N+1 : Input characteristic parameter data

[0036] NR1,NR2,NR3,NR4: Neural network layers

[0037] P t ,P t-1,P t-19 :Heating power

[0038] S200, S202, S204, S206, S208: Steps

[0039] T next (FS i ):Candidate future temperature

[0040] T t ,Tt-1,T t-19 :Predicted Temperature DETAILED DESCRIPTION

[0041] Please refer to Figure 1 , Figure 1 Schematic diagram of a server system 1 according to an embodiment of the present invention. The server system 1 includes a processor 10, an adapter card 20, a control device 30, a fan 40, a temperature sensor 50 and a speed sensor 60. In the server system 1, the control device 30 can be used to control the fan 40 to dissipate heat, so as to prevent the related components inside the server system 1 from failing or even being damaged due to excessive temperature. The processor 10 includes a temperature sensor 102. The temperature sensor 102 is used to detect the temperature of the processor 10. The adapter card 20 includes a chip 202. The chip 202 is disposed on the adapter card 20. The adapter card 20 can be an adapter card with a high-speed peripheral component interconnect (PCIe) transmission interface, a serial high technology configuration (Serial Advanced Technology Attachment, SATA) transmission interface, a serial small computer system (Serial Attached SCSI, SAS) transmission interface, a universal serial bus (Universal Serial Bus, USB) transmission interface or other data transmission interface.

[0042] The control device 30 is used to generate a fan control signal to drive the fan 40. The control device 30 can be a baseboard management controller (BMC), a microprocessor, a central processing unit or a graphics processor, but is not limited thereto. The control device 30 includes a temperature predictor 302 and a fan controller 304. The temperature predictor 302 is configured to use a first deep learning model and generate a predicted temperature of the chip 202 according to the rotation speed of the fan 40, the heat power of the adapter card 20, the inlet temperature of the adapter card 20 and the temperature of the processor 10. The fan controller 304 is coupled to the temperature predictor 302. The fan controller 304 is configured to use a second deep learning model and generate a target rotation speed of the fan 40 according to the predicted temperature of the chip 202, the rotation speed of the fan 40, and the heat power of the adapter card 20, and generate a fan control signal according to the target rotation speed of the fan 40 to drive the fan 40. The temperature sensor 50 may be disposed near the adapter card 20, disposed on the adapter card 20, or disposed at any location inside the housing of the server system 1. The temperature sensor 50 is used to detect an inlet temperature of the adapter card 20. The temperature sensors 102 and 50 may be thermocouples, thermistors, resistance temperature detectors (RTDs), or integrated circuit temperature sensors, but are not limited thereto. The speed sensor 60 is used to detect the speed of the fan 40.

[0043] For detailed operation of server system 1, please refer to the following instructions. Figure 2 , Figure 2 FIG. 2 is a schematic diagram of a process 2 according to an embodiment of the present invention. Figure 2 The process 2 is used to implement the above-mentioned server system 1 control operation process, which includes the following steps:

[0044] S200: Start.

[0045] S202: Use a first deep learning model to generate a predicted temperature of the chip according to a rotation speed of a fan, a heating power of an adapter card, an inlet temperature of the adapter card, and a temperature of a processor.

[0046] S204: Using a second deep learning model and generating a target rotation speed of the fan according to the predicted temperature of the chip, the rotation speed of the fan, and the heating power of the adapter card.

[0047] S206: Generate a fan control signal according to the target rotation speed of the fan to drive the fan.

[0048] S208: End.

[0049] In order to prevent the temperature of the internal components of the server system 1 (such as the chip 202 on the adapter card 20) from being too high, the embodiment of the present invention provides the function of forced convection and heat dissipation inside the server system 1 by controlling the operation of the fan 40. When the server system 1 is operating, the processor 10 includes a corresponding temperature sensor 102 and can report the corresponding temperature value to control the operation of the fan 40. The chip 202 is configured on the adapter card 20, but the adapter card 20 is not equipped with a corresponding temperature sensor to report the real-time temperature value of the chip 202. In order to solve the problem of the chip 202 of the adapter card 20 being too high, the embodiment of the present invention controls the operation of the fan 40 through the predicted temperature and the required target speed provided by the control device 30.

[0050] according to Figure 2 In step S202 of process 2, the temperature predictor 302 is configured to use a first deep learning model and generate a predicted temperature of the chip 202 according to the rotation speed of the fan 40, the heat generation power of the adapter card 20, the inlet temperature of the adapter card 20, and the temperature of the processor 10. The predicted temperature represents the temperature of the chip 202 predicted by the temperature predictor 30 for each cycle based on the aforementioned parameters. For example, the rotation speed sensor 60 continuously detects the rotation speed of the fan 40 and provides the detected rotation speed to the temperature predictor 302. For example, the heat generation power recorded in the factory specifications of the adapter card 20 can be used as the heat generation power parameter applied to process 2. For example, Figure 3 As shown, the temperature sensor 50 can be arranged between the adapter card 20 and the fan 40 and the temperature value detected by the temperature sensor 50 is used as the inlet temperature of the adapter card 20. The inlet temperature of the adapter card 20 can be used to represent the temperature of the environment in which the adapter card 20 is located. For example, the temperature sensor 102 detects the temperature of the processor 10 and provides the temperature of the processor 10 to the temperature predictor 302. The speed of the fan 40 may include a current speed and a plurality of previous speeds corresponding to a plurality of previous cycles. The heating power of the adapter card 20 includes a current heating power and a plurality of previous heating powers corresponding to a plurality of previous cycles. The inlet temperature of the adapter card 20 includes a current inlet temperature and a plurality of previous inlet temperatures corresponding to a plurality of previous cycles. The temperature of the processor 10 includes a current processor temperature and a plurality of previous processor temperatures corresponding to a plurality of previous cycles.

[0051] The temperature predictor 302 includes a plurality of first neural network layers to form a first deep learning model DL1. The first neural network layer may include but is not limited to a recurrent neural network (RNN), a convolutional neural network (CNN), a feed-forward neural network (FNN), a long short-term memory (LSTM) network, a gated recurrent unit (GRU), an attention mechanism, an activation function, a fully-connected layer or a pooling layer. For example, please refer to Figure 4 , Figure 4 for Figure 1 FIG. 3 is a schematic diagram of an embodiment of the temperature predictor 302 in FIG. Figure 4 As shown, the temperature predictor 302 includes a plurality of neural network layers NR1, NR2, a neural network layer NR3 and a neural network layer NR4 to form a first deep learning model DL1. For example, the neural network layer NR1 and the neural network layer NR2 may be gated recurrent units, and the neural network layer NR3 and the neural network layer NR4 may be fully connected layer neural networks, but are not limited thereto. Each neural network layer NR1 receives input feature parameter data, and each neural network layer NR1 outputs data to the corresponding neural network layer NR2. The neural network layer NR2 outputs data to the neural network layer NR3. The neural network layer NR4 receives the output data of the neural network layer NR3 and outputs the predicted temperature of the chip 202.

[0052] like Figure 4 As shown, the input characteristic parameter data FT includes the speed of the fan 40, the heat generation power of the adapter card 20, the inlet temperature of the adapter card 20 and the temperature of the processor 10 to generate the predicted temperature of the chip 202. t Indicates the input characteristic parameter data of the current period. For example, the current period is period t, and the input characteristic parameter data FT t Including the speed of the fan 40 at cycle t, the heat generation power of the adapter card 20 at cycle t, the inlet temperature of the adapter card 20 at cycle t, and the temperature of the processor 10 at cycle t. Input characteristic parameter data FT t-1 ~FT t-N+1 Represents the input characteristic parameter data of the previous cycle, where N is a positive integer. For example, the input characteristic parameter data FT t-1This includes the speed of the fan 40 at cycle (t-1), the heat generation power and inlet temperature of the adapter card 20 at cycle (t-1), the temperature of the processor 10 at cycle (t-1), and so on. Figure 4 As shown, according to the input characteristic parameter data FT t ~FT t-N+1 The temperature predictor 302 generates and outputs a predicted temperature T of the chip 202 at period t. t In this way, when the system is operating, the temperature predictor 302 continuously provides the predicted temperature of the chip 202 in each cycle to the fan controller 304 .

[0053] In step S204, the fan controller 304 is configured to use a second deep learning model and generate a target speed of the fan 40 according to the predicted temperature of the chip 202 in step S202, the speed of the fan 40, and the heating power of the adapter card 20. Similar to step S202, the speed sensor 60 continuously detects the speed of the fan 40 and provides the detected speed to the fan controller 304. The heating power recorded in the factory specification of the adapter card 20 can be used as the heating power parameter applied to process 2. The temperature sensor 50 detects the inlet temperature of the adapter card 20 and provides it to the fan controller 304. The temperature sensor 102 detects the temperature of the processor 10 and provides it to the fan controller 304. The speed of the fan 40 includes the current speed and a plurality of previous speeds corresponding to a plurality of previous cycles. The heating power of the adapter card 20 includes the current heating power and a plurality of previous heating powers corresponding to a plurality of previous cycles. The temperature of the processor 10 includes the current processor temperature and a plurality of previous processor temperatures corresponding to a plurality of previous cycles.

[0054] In more detail, the fan controller 304 includes a plurality of second neural network layers to form a second deep learning model. The second neural network layer may include but is not limited to a recurrent neural network, a convolutional neural network, a feedforward neural network, a long short-term memory network, a gated recurrent unit, an attention mechanism, an activation function, a fully connected layer or a pooling layer. In step S204, the fan controller 304 may use the second deep learning model and generate a plurality of candidate future temperatures corresponding to a plurality of candidate speeds according to the predicted temperature of the chip 202 in step S202, the speed of the fan 40, the heating power of the adapter 20, and a plurality of candidate speeds. The candidate future temperature corresponding to each candidate speed may represent the temperature of the chip 202 predicted by the fan controller 304 in the next cycle when each candidate speed is used as the target speed to drive the fan 40 so that the fan 40 reaches the target speed.

[0055] For example, see Figure 5 , Figure 5 for Figure 1FIG. 3 is a schematic diagram of an embodiment of the fan controller 304. Figure 5 As shown, the fan controller 304 includes a plurality of neural network layers NR1, NR2, a neural network layer NR3 and a neural network layer NR4 to form a second deep learning model DL2. For example, the neural network layer NR1 and the neural network layer NR2 may be gated loop units, and the neural network layer NR3 and the neural network layer NR4 may be fully connected layer neural networks, but are not limited thereto. Each neural network layer NR1 receives corresponding input parameter data, and each neural network layer NR1 outputs data to the corresponding neural network layer NR2. The neural network layer NR2 outputs data to the neural network layer NR3. The neural network layer NR4 receives the output data of the neural network layer NR3 and outputs the candidate future temperature corresponding to each candidate rotation speed of the chip 202. In one embodiment, as Figure 4 and Figure 5 As shown, the first deep learning model DL1 of the temperature predictor 302 may be the same as the second deep learning model DL2 of the fan controller 304. In another embodiment, the first deep learning model DL1 of the temperature predictor 302 may be different from the second deep learning model DL2 of the fan controller 304. Figure 5 As shown, period t represents the current period, and the predicted temperature T t The predicted temperature of the chip 202 at the period t generated by the temperature predictor 302 is shown. t represents the heat generation power of the adapter card 20 at cycle t, and the rotation speed FS t represents the speed of the fan 40 at cycle t. Predicted temperature T t-1 ~T t-19 The predicted temperature of the chip 202 at the period (t-1) to (t-19) generated by the temperature predictor 302 is shown. t-1 ~P t-19 represents the heat generation power of the adapter card 20 during the period (t-1) to (t-19), and the rotation speed FS t-1 ~FS t-18 Indicates the speed of the fan 40 in the period (t-1) to (t-18). FS i represents the i-th candidate speed of the fan 40 at cycle (t+1), where i is a positive integer. next (FS i ) represents the candidate future temperature of the chip 202 when the fan 40 operates at the i-th candidate speed in period (t+1). In other words, the fan controller 304 can generate a plurality of candidate future temperatures corresponding to a plurality of candidate speeds.

[0056] Further, in step S204, after generating a plurality of candidate future temperatures corresponding to a plurality of candidate rotational speeds, the fan controller 304 may compare the plurality of candidate future temperatures with a set point temperature respectively, and then select the candidate future temperature with the smallest absolute difference with the set point temperature among the plurality of candidate future temperatures as a future target temperature, and determine the candidate rotational speed corresponding to the selected future target temperature as the target rotational speed of the fan 40. The set point temperature may be pre-set according to the chip 202. Therefore, for each candidate future temperature, the fan controller 304 may calculate an absolute difference between each candidate future temperature and the set point temperature. The fan controller 304 then selects the candidate future temperature with the smallest absolute difference with the set point temperature from the plurality of candidate future temperatures as the future target temperature, and determines the candidate rotational speed corresponding to the selected future target temperature as the target rotational speed of the fan 40. In short, the fan controller 304 may determine the target rotational speed of the fan 40 according to formula (1):

[0057]

[0058] Among them FS next Indicates the target speed at cycle (t+1) (future cycle); T set Indicates the set point temperature; T next (FSi) represents the candidate future temperature of the chip 202 when the fan 40 operates at the i-th candidate speed in cycle (t+1); abs() represents the absolute value of the value in the brackets; Min() represents a function that obtains a minimum value of the value in the brackets; and m is a positive integer.

[0059] For example, assuming the set point temperature T set is 70 degrees (℃), m is equal to 5, the candidate speed FS 1 is 2000 revolutions per minute (RPM), the candidate speed FS 1 The corresponding candidate future temperature T next (FS 1 ) is 100°C, which means that if the fan 40 operates at the candidate speed FS in the period (t+1) 1 When the fan controller 304 predicts the temperature of the chip 202 (the candidate future temperature T next (FS 1 )) is 100°C. Candidate speed FS 2 The candidate future temperature T corresponding to the candidate speed FS2 is 4000RPM. next (FS 2 ) is 85°C. Candidate speed FS 3 The candidate speed is 6000RPM, FS 3 The corresponding candidate future temperature T next(FS 3 ) is 80°C. Candidate speed FS 4 is 8000RPM, candidate speed FS 4 The corresponding candidate future temperature T next (FS 4 ) is 65°C. Candidate speed FS 5 is 10000RPM, candidate speed FS 5 The corresponding candidate future temperature T next (FS 5 ) is 60° C. In step S204 , the fan controller 304 calculates the candidate future temperature T next (FS 1 ) and the set point temperature T set The absolute difference is 30℃, and the candidate future temperature T next (FS 2 ) and the set point temperature T set The absolute difference is 15℃, and the candidate future temperature T next (FS 3 ) and the set point temperature T set The absolute difference is 10℃, and the candidate future temperature T next (FS 4 ) and the set point temperature T set The absolute difference is 5℃, and the candidate future temperature T next (FS 5 ) and the set point temperature T set The absolute difference is 10°C. In this case, the candidate future temperature T next (FS 4 ) and the set point temperature T set The fan controller 304 selects the candidate future temperature T next (FS 4 ) as the future target temperature and the candidate future temperature T next (FS 4 ) is determined as a target speed of the fan 40.

[0060] In step S206, the fan controller 304 generates a fan control signal to drive the fan 40 according to the target speed of the fan 40 determined in step S206. In this way, based on the control of the fan control signal, the fan 40 can be operated at the target speed, thereby preventing the chip 202 on the adapter card 20 from overheating and maintaining normal operation. In other words, the embodiment of the present invention can select an optimal fan target speed based on the predicted temperature of the chip 202 to drive the fan 40 to operate, so that the temperature of the chip 202 on the adapter card 20 can be maintained within the minimum error range of the set temperature value, thereby preventing the chip 202 from overheating, and also preventing the fan 40 from over-operating and consuming unnecessary power.

[0061] On the other hand, the training of the deep learning model can establish the corresponding deep learning model through the machine learning and deep learning methods of classification, clustering and regression. When the first deep learning model DL1 and the second deep learning model DL2 are designed for model training, the embodiment of the present invention can use the same set of collected input feature data (such as the rotation speed of the fan 40, the heat generation power of the adapter card 20, the inlet temperature of the adapter card 20, and the temperature of the processor 10) to simultaneously train the first deep learning model DL1 of the temperature predictor 302 and the second deep learning model DL2 of the fan controller 304. After the training is completed, the output data of the temperature predictor 302 can be input to the fan controller 304, and the fan controller 304 can control the fan according to the predicted temperature of the chip to achieve the temperature control of the chip. In addition, when the first deep learning model DL1 and the second deep learning model DL2 are trained, a temperature sensor can also be set on the chip 202 to measure the actual temperature of the chip 202 to verify the training data input when training the model.

[0062] A person skilled in the art may combine, modify or change the above-described embodiments according to the spirit of the present invention, but is not limited thereto. All statements, steps, and / or processes (including recommended steps) in the embodiments of the present invention may be implemented by hardware, software, firmware (i.e., a combination of hardware devices and computer instructions, where the data in the hardware devices are read-only software data), electronic systems, or a combination of the above devices. The device may be a server system 1. The hardware may include analog, digital and hybrid circuits (i.e., microcircuits, microchips or silicon chips). For example, the hardware may be an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a programmable logic element, a coupled hardware element, or a combination of the above hardware. In other embodiments, the hardware may include a general purpose processor, a microprocessor, a controller, a digital signal processor, or a combination of the above hardware. The software may be a combination of program codes, a combination of instructions and / or a combination of functions, which are stored in a storage device, such as a computer readable recording medium or a non-transitory computer-readable medium. For example. The computer-readable recording medium may include a read-only memory (ROM), a flash memory (Flash Memory), a random access memory (RAM), a subscriber identity module (SIM), a hard disk, a floppy disk or a CD-ROM (CD-ROM / DVD-ROM / BD-ROM), but is not limited thereto. The process steps and embodiments of the present invention may be compiled into a program code or instruction form and stored in a computer-readable recording medium. The processing circuit may be used to read and execute the program code or instructions stored in the computer-readable medium to implement all the aforementioned steps and functions. The server system 1 may include a processing circuit and a computer-readable recording medium coupled to the processing circuit. The computer-readable recording medium stores instructions or program codes to provide the processing circuit with access and execution. The processing circuit may read and execute the instructions or program codes stored in the computer-readable recording medium. The server system 1 may be any computer device that includes the processing circuit and the computer-readable recording medium and can execute the instructions or program codes of the aforementioned steps and processes to implement the aforementioned functions.

[0063] In summary, the embodiment of the present invention can select an optimal fan target speed based on the predicted temperature of the chip 202 to drive the fan 40 to operate, so that the temperature of the chip 202 on the adapter card 20 can be maintained within the minimum error range of the set temperature value, thereby preventing the chip 202 from overheating and maintaining normal operation, and also preventing the fan 40 from over-operating and consuming unnecessary power.

[0064] In one embodiment of the present invention, the control device and control method of the present invention are applicable to the fan control of the server to dynamically adjust the fan target speed so that the server can achieve a balance between performance and energy consumption. Therefore, the server to which the control device and control method of the present invention are applied can be suitable for artificial intelligence (AI) computing, edge computing, and can also be used as a 5G server, cloud server or Internet of Vehicles server.

[0065] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made according to the claims of the present invention should fall within the scope of the present invention.

Claims

1. A control device for controlling a fan of a server system, It is characterized in that Included: a temperature predictor configured to generate a predicted temperature of a chip using a first deep learning model according to a rotation speed of the fan, a heat generation power of an adapter card, an inlet temperature of the adapter card, and a temperature of a processor, the chip being disposed on the adapter card, the inlet temperature being detected by a temperature sensor disposed between the adapter card and the fan; and A fan controller is coupled to the temperature predictor and is configured to use a second deep learning model to generate a target speed of the fan according to the predicted temperature of the chip, the speed of the fan and the heat generation power of the adapter card, and to generate a plurality of candidate future temperatures corresponding to the plurality of candidate speeds. The fan controller compares the plurality of candidate future temperatures with a set point temperature respectively, and selects the candidate future temperature with the smallest absolute difference between the plurality of candidate future temperatures and the set point temperature as a future target temperature and determines the candidate speed corresponding to the selected future target temperature as the target speed of the fan, and generates a fan control signal according to the target speed of the fan to drive the fan.

2. The control device according to claim 1, It is characterized in that The temperature predictor includes a plurality of first neural network layers to form the first deep learning model, and the fan controller includes a plurality of second neural network layers to form the second deep learning model.

3. The control device according to claim 1, It is characterized in that The rotation speed of the fan includes a current rotation speed of the fan and a plurality of previous rotation speeds corresponding to a plurality of previous cycles, the heating power of the adapter card includes a current heating power of the adapter card and a plurality of previous heating powers corresponding to the plurality of previous cycles, the inlet temperature of the adapter card includes a current inlet temperature of the adapter card and a plurality of previous inlet temperatures corresponding to the plurality of previous cycles, and the temperature of the processor includes a current processor temperature of the processor and a plurality of previous processor temperatures corresponding to the plurality of previous cycles.

4. The control device according to claim 1, It is characterized in that The first deep learning model and the second deep learning model are the same deep learning model.

5. A control method for controlling a fan of a server system, It is characterized in that Included: Using a first deep learning model and generating a predicted temperature of a chip according to a rotation speed of the fan, a heat generation power of an adapter card, an inlet temperature of the adapter card, and a temperature of a processor, the chip being disposed on the adapter card, the inlet temperature being detected by a temperature sensor disposed between the adapter card and the fan; and A second deep learning model is used to generate a target speed of the fan according to the predicted temperature of the chip, the speed of the fan, and the heat generation power of the adapter card, and a plurality of candidate future temperatures corresponding to the plurality of candidate speeds are generated. The fan controller compares the plurality of candidate future temperatures with a set point temperature respectively, and selects the candidate future temperature with the smallest absolute difference between the plurality of candidate future temperatures and the set point temperature as a future target temperature, and determines the candidate speed corresponding to the selected future target temperature as the target speed of the fan, and generates a fan control signal according to the target speed of the fan to drive the fan.

6. The control method according to claim 5, It is characterized in that The first deep learning model is formed by a plurality of first neural network layers, and the second deep learning model is formed by a plurality of second neural network layers.

7. The control method according to claim 5, It is characterized in that The rotation speed of the fan includes a current rotation speed of the fan and a plurality of previous rotation speeds corresponding to a plurality of previous cycles, the heating power of the adapter card includes a current heating power of the adapter card and a plurality of previous heating powers corresponding to the plurality of previous cycles, the inlet temperature of the adapter card includes a current inlet temperature of the adapter card and a plurality of previous inlet temperatures corresponding to the plurality of previous cycles, and the temperature of the processor includes a current processor temperature of the processor and a plurality of previous processor temperatures corresponding to the plurality of previous cycles.

8. The control method according to claim 5, It is characterized in that The first deep learning model and the second deep learning model are the same deep learning model.

Images