Multi-GPU (Graphics Processing Unit) cascade buckle card, mainboard and liquid cooling heat dissipation case

Abstract

The utility model discloses a multi-GPU cascade buckle card, a mainboard and a liquid cooling heat dissipation case, and belongs to the field of server board card design. According to the technical scheme, the buckle comprises a buckle body; the buckle body is provided with a first connector, a power interface, a power module and a plurality of GPUs. The first connector is connected with each GPU, and the GPUs are connected through an interconnection bus; the power supply module is connected with the power supply interface and takes power from the power supply interface; and the power supply module is also connected with the GPU and supplies power to the GPU. The beneficial effects of the utility model are that the buckle breaks through the limitation of the number of integrated GPUs, avoids the external connection of cables through the interconnection of signal lines between the GPUs, reduces the loss of high-speed signals through the connector, and improves the reliability of the GPUs. The liquid cooling heat dissipation case based on the buckle solves the problem that a traditional air cooling case is difficult to solve due to large power consumption of the GPU through cold plate type liquid cooling or immersed liquid cooling.

Description

Technical Field

[0001] This utility model relates to the field of server board design technology, and in particular to a multi-GPU cascade card, a motherboard, and a liquid-cooled heat dissipation chassis. Background Technology

[0002] With the increasing demand for data center computing power in the industrial internet environment, the design requirements for high-performance servers are also getting higher and higher. High computing power is no longer a simple stacking of CPUs in the traditional sense. GPUs, as a powerful supplement to CPUs, are being used more and more in servers.

[0003] Taking a standard rack-mount 2U dual-socket server as an example, this server typically uses air cooling. When performing calculations, the server usually requires the insertion of a GPU card. Since the front window of the server chassis needs to house the hard drive, a fan is placed between the hard drive and the CPU. The power supply PSU and the standard configuration card are located behind the CPU. The GPU Riser card stands vertically on the motherboard, while the GPU card is parallel to the motherboard and inserted vertically into the Riser card.

[0004] Currently, the way GPU cards are inserted into server motherboards is limited by the space in the rear window of the chassis, as well as the size of the GPU card and GPU Riser card. This type of 2U onboard server chassis can only accommodate a maximum of four GPU cards. Furthermore, the motherboard must have reserved power connectors for the GPU Riser and GPU cards, which is not conducive to large-scale GPU integration and makes server motherboard design more complex and difficult. Moreover, interconnection between GPU cards requires high-speed cables external to the rear window of the chassis, increasing the design cost and complexity of high-speed lines. In addition, regarding heat dissipation, air is drawn in by the fan from the front window and blown towards the high-power CPU and memory modules, but the fan speed slows down at the GPU card in the rear window, resulting in increased heat, which no longer solves the GPU cooling problem. Summary of the Invention

[0005] To address the problem that current server chassis use GPU Riser cards and GPU cards in combination to supplement CPU computing power, which limits the number of GPU cards and is not conducive to large-scale GPU integration, this utility model provides a multi-GPU cascading card.

[0006] To solve the above problems, firstly, the technical solution adopted by this utility model is a multi-GPU cascaded clip card, including a clip card body;

[0007] The card body is provided with a first connector, a power interface, a power module and several GPUs;

[0008] The first connector is connected to each GPU, and the GPUs are connected to each other via an interconnect bus;

[0009] The power module is connected with the power interface to take power from the power interface;

[0010] The power module is also connected with the GPU to supply power for the GPU. In this scheme, the original use of GPU Riser card and GPU card combination for CPU is changed to use the card, and the GPU chip is directly integrated on the card. It breaks through the limit of the number of integrated GPUs, and each GPU chip on the card is directly interconnected through a high-speed bus without using external cables for GPU interconnection. A first connector is provided on the card to replace the original GPU Riser card gold finger, and the first connector uses a high-speed signal connector to improve signal transmission performance.

[0011] As preferred, the number of power modules is the same as the number of GPUs, and each power module supplies power for one GPU;

[0012] The power interface includes a first power interface and a second power interface;

[0013] The first power interface is connected with a PSU module, and the PSU module is connected with a 220V external AC power supply;

[0014] The second power interface is connected with an HVDC high-voltage electronic card, and the HVDC high-voltage electronic card is connected with a 380V external AC power supply;

[0015] The first power interface and the second power interface are connected with each power module. In this preferred scheme, the power module converts 12V DC voltage into 1V to supply power for the GPU, and each power module supplies power for one GPU. At the same time, the power module is arranged in the vicinity of the GPU that needs to be powered, so as to shorten the distance of 1V power supply, thereby improving the power supply efficiency, reducing the damage of the PCB to the power supply, and directly supplying power to the GPU through the power module. The integration and efficiency are better than the general buck power supply, the size of the card can be made smaller, and the heat loss is also smaller;

[0016] The power interface is provided with two power supply interfaces, which can be flexibly selected for power supply of the GPU. The PSU module connected with the first power interface is the PSU module for power supply of the mainboard, and the second power interface is connected with the 380V external AC power supply through the HVDC high-voltage electronic card, so as to be separated from the power supply of the mainboard. By directly inserting the external power on the card, the GPU can be set to be independent of the server whole machine power supply system, so that the power supply of the card does not affect the system power supply, and the power consumption is reduced.

[0017] As preferred, each GPU is provided with three interconnection interfaces;

[0018] The four GPUs on the card body form a GPU group;

[0019] At least one GPU group is arranged on the card clamping body. In the preferred embodiment, since four GPU Riser cards are used in combination with four GPU cards in the original scheme of using GPU Riser cards and GPU cards in combination with CPU for algorithmic power supplement, at least four GPUs are used in the card clamping scheme. Since each GPU is provided with three interconnection interfaces, the four GPU cards can be interconnected two by two, so that the four GPU cards are set as a group.

[0020] As a preference, at least two GPU groups are arranged on the card clamping body.

[0021] Each GPU group is sequentially cascaded.

[0022] Two GPUs in the frontmost GPU group connected with the rear GPU group are set as first interconnection GPUs.

[0023] Two GPUs in the last GPU group connected with the front CPU group are set as second interconnection GPUs.

[0024] Each first interconnection GPU is connected with one GPU in the rear GPU group through one interconnection interface, and each second interconnection GPU is connected with one GPU in the front GPU group through one interconnection interface. In the preferred embodiment, the card clamping can realize the cascade of the GPU group. Limited by the GPU power consumption and the space limitation in the case, the required GPU group can be set according to the need. Since each GPU has three interconnection interfaces, when the GPU group is cascaded, the interconnection interface connected with the GPU in the front GPU group needs to be changed to be connected with the rear GPU group, and the interconnection interface connected with the GPU in the rear GPU group needs to be changed to be connected with the rear GPU group.

[0025] As a preference, an electronic fuse, a first temperature sensor, a CPLD, an FRU and an ADC chip are further arranged on the card clamping body.

[0026] The first temperature sensor is arranged at the GPU.

[0027] The first connector is provided with a bypass voltage pin connected with the first temperature sensor, the CPLD, the FRU and the ADC chip.

[0028] The electronic fuse is arranged between the power interface and the power module. In the preferred embodiment, the bypass voltage pin supplies power to the first temperature sensor, the CPLD, the FRU and the ADC chip after taking power from the first connector; the first temperature sensor collects temperature information of the GPU. Since the single GPU has high power consumption, the CUP is controlled by the CUP to stagger the power-on time of each GPU to prevent voltage drop caused by simultaneous power-on of all GPUs from causing the power supply to be dead; after the card is working, the first connector provides 3.3V power to the card to supply power to the CPLD, the FRU and the temperature sensor on the card, and the first connector is also provided with a card-in-place signal pin; when the card-in-place signal is detected through the first connector, the FRU is read to obtain the card information.

[0029] Preferably, the card body is further provided with an I2C expansion chip;

[0030] The I2C expansion chip is provided with a summary side and an expansion side;

[0031] The expansion side of the I2C expansion chip is connected with the first temperature sensor, the FRU and the CPLD through a first I2C signal channel, connected with the ADC chip and the electronic fuse through a second I2C signal channel, connected with two GPUs in the GPU group through a third I2C signal channel, and connected with the other two GPUs in the GPU group through a fourth I2C signal channel;

[0032] The first connector is further provided with a BMC signal pin;

[0033] The BMC signal pin is connected with the summary side of the I2C expansion chip through an I2C signal line. In the preferred embodiment, the BMC outside the card obtains the information on the card body through the I2C signal line, and since there are many information to be transmitted on the card, the I2C signal channel is expanded through the I2C expansion information, and the specific expanded I2C signal channel is determined according to the number of GPUs; the external BMC can read the electronic fuse on the card body through the I2C signal line to monitor the power consumption of the GPU, and read the internal register of the ADC chip to monitor the voltage of each power on the card.

[0034] In a second aspect, the utility model also provides a kind of mainboard, and the mainboard is provided with second connector and CPU;

[0035] The second connector is matched with the first connector of the multi-GPU cascade card of the first aspect;

[0036] The first connector is realized by the plug-in of second connector, and the multi-cascade GPU card is installed to the mainboard;

[0037] The first connector and the second connector are provided with PCIE signal lines;

[0038] The PCIE signal lines of the first connector are connected with the GPU, and the PCIE signal lines of the second connector are connected with the CPU;

[0039] The second connector is arranged at a position between the CPU and a front window of the case or at a position between the CPU and a rear window of the case. In this scheme, the mainboard is connected with the second connector of the card of the first aspect through the second connector, and through the cooperation of the mainboard and the card, the original limitation of the number of GPUs that can be matched with the mainboard is broken, and the original four golden finger slots directly inserted with the GPU Riser card are changed into the second connector adopting a high-speed connector to improve the signal transmission performance. In addition, since the power supply interface can be connected with the PSU module and the HVDC high-voltage electronic card, the flexible setting of the position of the second connector can ensure that the position of the card on the mainboard is not affected by the power flow, specifically, the card supports a one-plus-one redundancy design of double PSU modules. When the card is placed at the rear window, it is close to the PSU module and can be directly powered by the PSU module. When the card is placed at the front window, it is far away from the PSU module. In order to prevent the problem of power flow caused by too far 12V plane, the 12V voltage output by the PSU is introduced to the front window by using a busbar.

[0040] As a preferred, the number of the CPUs is several;

[0041] The first connector and the second connector adopt a board-to-board mirror image sandwich connector. In this preferred scheme, the mainboard can support a multi-path server. In specific applications, each CPU needs to be connected with all GPUs through PCIE signal lines, that is, the number of PCIE signal lines in the first connector and the second connector is equal to the product of the number of CPUs and the number of GPUs. Specifically, the number of PCIE signal lines can be designed according to the product of the maximum supportable number of GPUs and the number of CPUs, and in actual applications, based on the principle of peripheral balance, the same number of GPUs needs to be mounted under each CPU. The board-to-board mirror image sandwich connector realizes the direct clamping of the card to the CPU. Compared with the GPU Riser card mode, the loss of high-speed signals will be smaller, and the board-to-board mirror image sandwich connector can support a signal rate transmission of up to 56Gpbs and support PCIE Gen5 signal transmission. The structures of the two connectors are completely the same, for example, 2.5mm, 5.5mm and 7.5mm can be matched with high, and the cooperation of the two connectors can meet different matching height requirements.

[0042] For the current server chassis using GPU Riser card and GPU card combination for CPU to supplement the computing power, in the heat dissipation, the wind from the front window is sucked by the fan and blown to the CPU and the memory bar with large power, and the fan slows down at the GPU card at the rear window, the heat increases, which cannot solve the heat dissipation problem of the GPU, the utility model provides a liquid cooling heat dissipation chassis.

[0043] To solve the above problems, the third aspect, the utility model adopts the scheme, a liquid cooling heat dissipation chassis, including chassis shell,

[0044] The chassis shell is provided with a mainboard, and the mainboard adopts the mainboard of the second aspect;

[0045] The CPU of the mainboard is provided with a first cold plate, and the multi-GPU cascade buckle card is provided with a second cold plate;

[0046] The first adjusting valve is arranged at the liquid inlet and the liquid outlet of the first cold plate, and the second adjusting valve is arranged at the liquid inlet and the liquid outlet of the second cold plate;

[0047] Each CPU is provided with a second temperature sensor;

[0048] The liquid cooling controller is connected with the first adjusting valve, the second adjusting valve, the first temperature sensor and the second temperature sensor.In this scheme, since the GPU power consumption is very high, the buckle card cascades multiple GPUs, and the buckle card power consumption is very large, the conventional air-cooled chassis is difficult to solve the heat dissipation problem, and the cold plate heat dissipation is used to solve the heat dissipation problem of the chassis, the GPU and the CPU are attached to the designed cold plate, the cold plate adopts the existing cold plate heat dissipation technology, and the flow rate of the liquid at the inlet and outlet of the cold plate is controlled according to the collected GPU and CPU junction temperature.

[0049] The fourth aspect, the utility model provides another liquid cooling heat dissipation chassis, including sealed chassis shell,

[0050] The mainboard is arranged in the sealed chassis shell, and the mainboard adopts the mainboard of the second aspect;

[0051] The lower part of the sealed chassis shell is provided with a liquid inlet, and the upper part of the sealed chassis shell is provided with a liquid outlet;

[0052] A first water distributor is arranged at a set distance close to the CPU, and a second water distributor is arranged at a set distance close to the GPU;

[0053] The first water distributor and the second water distributor are communicated with the liquid inlet;

[0054] The inner top of the sealed chassis shell is provided with a liquid level detection sensor and an air pressure sensor;

[0055] A second temperature sensor is arranged at each CPU;

[0056] The mainboard is further provided with a liquid cooling circulation controller, and the liquid cooling circulation controller is connected with the liquid level detection sensor, the air pressure sensor, the first temperature sensor, the second temperature sensor, the first water distributor and the second water distributor. In this scheme, because the GPU power consumption is very high, the card cascade multiple GPUs, and the card power consumption is very large, and the conventional air-cooled case is difficult to solve the heat dissipation problem, the immersion liquid cooling heat dissipation is used to solve the case heat dissipation problem, and the immersion liquid cooling case is in a closed state. According to the CPU and GPU temperature, the air pressure in the case, the liquid level height, the liquid flow rate at the water distributor is controlled. In addition, in order to better dissipate heat, the water distributor is arranged near the CPU and the GPU to ensure that the liquid flows through the high-power devices after being discharged. The immersion liquid cooling heat dissipation adopts the existing mode, and the water distributor is actually a water distribution pipeline. Each water distribution pipeline is connected with a liquid inlet, and a valve body is arranged at the outlet of the water distribution pipeline, so that the flow rate of the discharged liquid can be adjusted.

[0057] It can be seen from the above technical scheme that the multi-GPU cascade card provided by the scheme directly integrates a GPU chip, breaks through the limitation of the number of integrated GPUs, and directly interconnects each GPU chip through a high-speed interconnection bus on the card without using an external cable for GPU interconnection. In addition, a connector is arranged on the card to replace the original GPU Riser card gold finger, thereby improving the signal transmission performance. Through the design of two power supply interfaces, the PSU module of the mainboard can be used, and an external power supply can also be used, thereby improving the power supply flexibility. When the external power supply is used, the overall energy consumption of the mainboard can be reduced. Furthermore, the mainboard of the utility model is connected with the second connector of the first aspect card through the second connector, and the mainboard and the card are matched, thereby breaking through the limitation of the number of GPUs that can be matched with the original mainboard. The two connectors adopt a board-to-board mirror image sandwich connector, which can reduce the loss of high-speed signals. Furthermore, the utility model provides a liquid cooling heat dissipation case, which solves the problem of high GPU power consumption through cold plate liquid cooling or immersion liquid cooling. The problem that cannot be solved by a conventional air-cooled case is solved. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical scheme of the utility model, the following will briefly introduce the drawings needed to be used in the description. Obviously, the drawings in the following description are only some embodiments of the utility model, and other drawings can also be obtained by those skilled in the art without creative labor.

[0059] Figure 1 It is a structural schematic view of the multi-GPU cascade card embodiment one of the utility model.

[0060] Figure 2 Structure schematic view of the second embodiment of the multi-GPU cascade buckle card of the utility model.

[0061] Figure 3 Structure schematic view of the multi-GPU group cascade of the utility model.

[0062] Figure 4 Structure schematic view of the second connector between the CPU and the front window in the utility model.

[0063] Figure 5 Structure schematic view of the second connector between the CPU and the rear window in the utility model.

[0064] Figure 6 Control schematic view of the cold plate type liquid cooling of the liquid cooling heat dissipation case in the utility model.

[0065] Figure 7 Control schematic view of the immersion type liquid cooling of the liquid cooling heat dissipation case in the utility model.

[0066] Main figure mark explanation

[0067] 1, buckle card body, 2, first connector, 3, power interface, 3.1, first power interface, 3.2, second power interface, 4, power module, 5, GPU, 6, PSU module, 7, 220V external AC power supply, 8, HVDC high-voltage electronic card, 9, 380V external AC power supply, 10, electronic fuse, 11, 12, CPLD, 13, FRU, 14, ADC chip, 15, I2C expansion chip, 16, mainboard, 17, second connector, 18, CPU, 19, first regulating valve, 20, second regulating valve, 21, second temperature sensor, 22, liquid cooling controller, 23, first water distributor, 24, second water distributor, 25, liquid level detection sensor, 26, air pressure sensor, 27, liquid cooling circulation controller. Specific implementation

[0068] In order to make the purpose, characteristics and advantages of the utility model more obvious and easy to understand, the technical solutions in the utility model will be described clearly and completely in the following combined with the drawings in the specific embodiments. Obviously, the following described embodiments are only a part of the embodiments of the utility model, not all the embodiments. Based on the embodiments in the patent, all other embodiments obtained by the person skilled in the art without creative labor belong to the scope of protection of the patent.

[0069] In the following implementation, the involved terms are explained as follows:

[0070] GPU is the abbreviation of Graphics Processing Unit, which means graphics processor.

[0071] CPU is short for Central Processing Unit.

[0072] PCIe stands for Peripheral Component Interconnect Express, a high-speed connection standard for peripheral devices. PCIE Gen5 is the fifth generation of PCIe technology, also known as PCI Express 5.0.

[0073] BMC is short for Baseboard Management Controller.

[0074] CPLD is short for Complex Programmable Logic Device.

[0075] A Riser card is a type of server expansion card.

[0076] HVDC is short for High Voltage Direct Current, a high-voltage DC power supply module.

[0077] PSU is short for Power Supply Unit.

[0078] FRU stands for Field Replaceable Unit.

[0079] ADC is short for "Analog-to-digital converter".

[0080] I2C is short for Inter-Integrated Circuit, an integrated circuit bus.

[0081] Example 1

[0082] like Figure 1 As shown, a multi-GPU cascaded clip card includes a clip card body 1;

[0083] The card holder body 1 is provided with a first connector 2, a power interface 3, a power module 4 and several GPUs 5;

[0084] The first connector 2 is connected to each GPU 4, and the GPUs 5 are connected to each other via an interconnect bus;

[0085] The power module 4 is connected to the power interface 3 and draws power from the power interface 3;

[0086] The power module 4 is also connected with the GPU 5 to supply power for the GPU 5;

[0087] It should be noted that the original GPU Riser card combined with the GPU card is changed to use a card, and the GPU chip is directly integrated on the card. The limitation of the number of GPU integrated is broken, and each GPU chip on the card is directly interconnected through the interconnection bus as a high-speed PCIE signal line without using external cables for GPU interconnection. In addition, the connector is arranged on the card to replace the original gold finger of the GPU Riser card, and the signal transmission performance is improved.

[0088] Embodiment two

[0089] As shown in Figure 2 A multi-GPU cascading card, comprising a card body 1;

[0090] The card body 1 is provided with a first connector 2, a power interface 3, a power module 4 and a plurality of GPUs 5;

[0091] The first connector 2 is connected with each GPU 4, and each GPU 5 is connected through an interconnection bus;

[0092] The power module 4 is connected with the power interface 3 to take power from the power interface 3;

[0093] The power module 4 is also connected with the GPU 5 to supply power for the GPU 5;

[0094] The number of power modules 4 is the same as the number of GPUs 5, and each power module 4 supplies power for one GPU 5; as Figure 2 shown, taking four GPUs 5 as an example, four power modules 4 supply power for each GPU 5 respectively;

[0095] The power interface 3 comprises a first power interface 3.1 and a second power interface 3.2;

[0096] The first power interface 3.1 is connected with a PSU module 6, and the PSU module 6 is connected with a 220V external AC power supply 7;

[0097] The second power interface 3.2 is connected with an HVDC high-voltage electronic card 8, and the HVDC high-voltage electronic card 8 is connected with a 380V external AC power supply 9;

[0098] The first power interface 3.1 and the second power interface 3.2 are connected with each power module 4;

[0099] It should be noted that the power module converts 12V DC voltage into 1V to power the GPU, each power module powers a GPU alone, and the power module is arranged close to the adjacent area of the GPU to be powered, shortens the distance of 1V power, thereby improving the power supply efficiency, reducing the damage of the PCB to the power supply, and directly supplying power to the GPU through the power module, the integration and efficiency are more changed than the general buck power supply, the size of the buckle can be made smaller, and the heat loss is also smaller;

[0100] The power interface is arranged as two, which can flexibly select the power supply form of the GPU, wherein the PSU module connected with the first power interface is the PSU module for powering the mainboard, and the second power interface is connected with 380V external AC power through the HVDC high-voltage electronic card, so as to be separated from the power supply of the mainboard, and the external power is directly inserted on the buckle, so that the GPU can be set independently of the server whole machine power supply system, so that the buckle power supply does not affect the system power supply, and the power consumption is reduced;

[0101] Each GPU 5 is provided with three interconnection interfaces;

[0102] The four GPUs 5 on the buckle body 1 form a GPU group;

[0103] At least one GPU group is arranged on the buckle body 1;

[0104] It should be noted that since the original GPU Riser card and GPU card combination scheme for CPU power supplement uses four GPU Riser cards to match four GPU cards, at least four GPUs are used in the buckle scheme, and each GPU is provided with three interconnection interfaces, so that the four GPU cards can be interconnected two by two, so that the four GPU cards are set as a group;

[0105] The buckle body 1 is further provided with an electronic fuse 10, a first temperature sensor 11, a CPLD 12, an FRU 13 and an ADC chip 14;

[0106] The first temperature sensor 11 is arranged at the GPU 5;

[0107] The first connector 2 is provided with a bypass voltage pin, and the bypass voltage pin is connected with the first temperature sensor 11, the CPLD 12, the FRU 13 and the ADC chip 14;

[0108] The electronic fuse 10 is arranged between the power interface 3 and the power module 4;

[0109] It should be noted that the bypass voltage pin supplies power to the first temperature sensor, the CPLD, the FRU and the ADC chip after taking power from the first connector; the first temperature sensor collects temperature information of the GPU. Since the single GPU has high power consumption, the card integrates multiple GPUs, and the power-on time of each GPU is staggered and sequentially powered on to prevent voltage drop when all GPUs are powered on at the same time; after the card is working, the first connector provides 3.3V power to the card, which is used to power the CPLD, FRU and temperature sensor on the card; the first connector also has a card in place signal pin, which reads the FRU and obtains the card information when the card in place signal is detected through the first connector;

[0110] The card body 1 is also provided with an I2C expansion chip 15;

[0111] The I2C expansion chip 15 is provided with a summary side and an expansion side;

[0112] The expansion side of the I2C expansion chip 15 is connected with the first temperature sensor 11, the FRU 13 and the CPLD 12 through a first I2C signal channel, connected with the ADC chip 14 and the electronic fuse 10 through a second I2C signal channel, connected with two GPUs 5 in the GPU group through a third I2C signal channel, and connected with another two GPUs 5 in the GPU group through a fourth I2C signal channel;

[0113] The first connector 2 is also provided with a BMC signal pin;

[0114] The BMC signal pin is connected with the summary side of the I2C expansion chip 15 through an I2C signal line;

[0115] It should be noted that the BMC outside the card acquires information on the card body through the I2C signal line, and since there are many information to be transmitted on the card, the I2C signal channel is expanded through the I2C expansion information, and the specific expanded I2C signal channel is determined according to the number of GPUs. The external BMC can read the electronic fuse on the card body through the I2C signal line to monitor the power consumption of the GPU, and read the internal register of the ADC chip to monitor the voltage of each power on the card.

[0116] In the above embodiment, as shown in Figure 3 The card body 1 is provided with two GPU groups;

[0117] Each GPU group is sequentially cascaded;

[0118] The two GPUs 5 in the frontmost GPU group connected with the rear GPU group are set as the first interconnected GPU;

[0119] Two GPUs 5 in the last-stage GPU group connected with the CPU group in the previous stage are set as second interconnection GPUs;

[0120] Each first interconnection GPU is connected with one GPU 5 in the GPU group in the next stage through an interconnection interface, and each second interconnection GPU is connected with one GPU 5 in the GPU group in the previous stage through an interconnection interface;

[0121] It should be noted that the GPU group can be cascaded by the buckle card. Limited by the GPU power consumption and the space limitation in the case, the required GPU group can be set according to the needs. Since each GPU has three interconnection interfaces, when the GPU group is cascaded, the interconnection interface connected with the GPU in the GPU group in the previous stage needs to be changed to be connected with the GPU group in the next stage, and the interconnection interface connected with the GPU in the GPU group in the next stage needs to be changed to be connected with the GPU group in the previous stage.

[0122] Embodiment three

[0123] As shown in Figure 4 A mainboard 16 is provided with a second connector 17 and a CPU 18;

[0124] The second connector 17 matches the first connector 2 of the multi-GPU cascading buckle card in the above-mentioned embodiment one or embodiment two;

[0125] The first connector 2 realizes the installation of the multi-cascading GPU buckle card to the mainboard 16 by plugging with the second connector 17;

[0126] The first connector 2 and the second connector 17 are both provided with PCIE signal lines;

[0127] The PCIE signal line of the first connector 2 is connected with the GPU 5, and the PCIE signal line of the second connector 17 is connected with the CPU 18;

[0128] The second connector 17 is arranged at a position between the CPU 18 and the front window of the case;

[0129] The number of CPUs is several;

[0130] The first connector and the second connector adopt a board-to-board mirror image sandwich connector;

[0131] It should be noted that the mainboard can support multiple servers, and in specific applications, each CPU needs to be connected with all GPUs through PCIE signal lines, that is, the number of PCIE signal lines in the first connector and the second connector is equal to the product of the number of CPUs and the number of GPUs, specifically, the number of PCIE signal lines can be designed according to the product of the maximum supportable GPU number and CPU number, and in actual application, based on the principle of peripheral balance, the same number of GPUs needs to be mounted under each CPU; the board-to-board mirror interlayer connector realizes that the card is directly buckled to the CPU, compared with the GPU Riser card mode, the loss of high-speed signals will be smaller, and the board-to-board mirror interlayer connector can support a signal transmission rate of up to 56Gpbs, support PCIE Gen5 signal transmission, and the structures of the two connectors are completely the same, for example, 2.5mm, 5.5mm and 7.5mm can be used, and through the cooperation of the two connectors, different height requirements can be met.

[0132] In this embodiment, as shown in Figure 5 The second connector 17 is arranged between the CPU 18 and the rear window of the case.

[0133] It should be noted that the mainboard is connected with the second connector of the first aspect card through the second connector, and through the cooperation of the mainboard and the card, the original limitation of the number of GPUs that can be matched with the mainboard is broken, and the original four golden finger slots directly inserted with the GPU Riser card are changed to the second connector, improving the signal transmission performance; in addition, since the power supply interface can be connected with the PSU module and the HVDC high-voltage electronic card, the flexible setting of the position of the second connector can ensure that the position of the card on the mainboard is not affected by the power flow, specifically, the card supports a one-plus-one redundancy design of double PSU modules, when the card is placed at the rear window, it is close to the PSU module, and can be directly powered by the PSU module, and when the card is placed at the front window, it is far away from the PSU module, in order to prevent the problem of power flow caused by too far 12V plane, the 12V voltage output by the PSU is introduced to the front window by using the busbar.

[0134] Embodiment four

[0135] As shown in Figure 6 A liquid cooling heat dissipation case, comprising a case shell;

[0136] A mainboard 16 is arranged in the case shell, and the mainboard 16 adopts the mainboard 16 of embodiment three;

[0137] A first cold plate is arranged at the CPU 18 of the mainboard 16, and a second cold plate is arranged at the multi-GPU cascade card;

[0138] The first cold plate is provided with a first adjusting valve 19 at the liquid inlet and the liquid outlet, and the second cold plate is provided with a second adjusting valve 20 at the liquid inlet and the liquid outlet;

[0139] The second temperature sensor 21 is arranged at each CPU 18;

[0140] The liquid cooling controller 22 is further arranged on the mainboard 16, and the liquid cooling controller 22 is connected with the first adjusting valve 19, the second adjusting valve 20, the first temperature sensor 11 and the second temperature sensor 21;

[0141] It should be noted that, since the GPU has high power consumption, a plurality of GPUs are cascaded on a card, and the power consumption of the card is large, it is difficult for a conventional air-cooled case to solve the heat dissipation problem, and the cold plate heat dissipation is used to solve the case heat dissipation problem, and the GPU and the CPU are attached to the designed cold plate, and the cold plate adopts the existing cold plate heat dissipation technology, and the flow rate of the liquid at the inlet and outlet of the cold plate is controlled according to the collected GPU and CPU junction temperature.

[0142] Example five

[0143] As shown in Figure 7 A liquid cooling heat dissipation case, comprising a sealed case shell;

[0144] The mainboard 16 is arranged in the sealed case shell, and the mainboard 16 is the mainboard 16 of the third embodiment;

[0145] The sealed case shell is provided with a liquid inlet at the lower part and a liquid outlet at the upper part;

[0146] The first water distributor 23 is arranged at a set distance close to the CPU, and the second water distributor 24 is arranged at a set distance close to the GPU;

[0147] The first water distributor 23 and the second water distributor 24 are in communication with the liquid inlet;

[0148] The liquid level detection sensor 25 and the air pressure sensor 26 are arranged at the inner top of the sealed case shell;

[0149] The second temperature sensor 21 is arranged at each CPU 18;

[0150] The liquid cooling circulation controller 27 is further arranged on the mainboard 16, and the liquid cooling circulation controller 27 is connected with the liquid level detection sensor 25, the air pressure sensor 26, the first temperature sensor 11, the second temperature sensor 21, the first water distributor 23 and the second water distributor 24;

[0151] It should be noted that, since the GPU power consumption is very high, the card is cascaded with multiple GPUs, and the card power consumption is very large, and the conventional air-cooled case is difficult to solve the heat dissipation problem, the immersion liquid cooling heat dissipation is used to solve the case heat dissipation problem, the immersion liquid cooling case is in a closed state, according to the CPU and GPU temperature, the case internal air pressure, the liquid level height, the liquid flow rate at the water distributor is controlled, in addition, in order to better dissipate heat, the water distributor is placed near the CPU and GPU, and the liquid outlet is ensured to flow through the high-power device first. The existing way is used for the immersion liquid cooling heat dissipation, the water distributor is actually a water distribution pipeline, each water distribution pipeline is connected with the liquid inlet, and a valve body is arranged at the outlet of the water distribution pipeline, so that the flow rate of the outflowing liquid can be adjusted.

[0152] The above description of disclosed embodiments enables those skilled in the art to carry out or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-GPU level daisy chain card, comprising: The card body comprises a first connector, a power supply interface, a power supply module and a plurality of GPUs. The first connector is connected with each GPU, and each GPU is connected through an interconnection bus. The power supply module is connected with the power supply interface and obtains power from the power supply interface. The power supply module is also connected with the GPUs and supplies power to the GPUs. The number of the power supply modules is the same as the number of the GPUs, and each power supply module supplies power to one GPU.

2. The multi-GPU daisy chain sled of claim 1, wherein, The power supply interface comprises a first power supply interface and a second power supply interface. The first power supply interface is connected with a PSU module, and the PSU module is connected with a 220V external AC power supply. The second power supply interface is connected with an HVDC high-voltage electronic card, and the HVDC high-voltage electronic card is connected with a 380V external AC power supply. The first power supply interface and the second power supply interface are connected with each power supply module. Each GPU is provided with three interconnection interfaces.

3. The multi-GPU daisy chain sled of claim 1, wherein, The four GPUs on the card body form a GPU group. The card body is provided with at least one GPU group. The card body is provided with at least two GPU groups.

4. The multi-GPU daisy chain sled of claim 3, wherein, Each GPU group is sequentially cascaded. Two GPUs in the GPU group at the first stage which are connected with the GPU group at the next stage are set as first interconnection GPUs. Two GPUs in the GPU group at the last stage which are connected with the GPU group at the previous stage are set as second interconnection GPUs. Each first interconnection GPU is connected with one GPU in the GPU group at the next stage through one interconnection interface, and each second interconnection GPU is connected with one GPU in the GPU group at the previous stage through one interconnection interface. The card body is also provided with an electronic fuse, a first temperature sensor, a CPLD, an FRU and an ADC chip.

5. The multi-GPU daisy chain sled of claim 3, wherein, The first temperature sensor is arranged at the GPU. The first connector is provided with a bypass voltage pin which is connected with the first temperature sensor, the CPLD, the FRU and the ADC chip. The electronic fuse is arranged between the power supply interface and the power supply module. The card body is also provided with an I2C expansion chip.

6. The multi-GPU daisy chain sled of claim 5, wherein, The I2C expansion chip is provided with a summary side and an expansion side. The expansion side of the I2C expansion chip is connected with the first temperature sensor, the FRU and the CPLD through a first I2C signal channel, connected with the ADC chip and the electronic fuse through a second I2C signal channel, connected with two GPUs in the GPU group through a third I2C signal channel, and connected with the other two GPUs in the GPU group through a fourth I2C signal channel. The first connector is also provided with a BMC signal pin. The BMC signal pin is connected with the summary side of the I2C expansion chip through an I2C signal line. The mainboard is provided with a second connector and a CPU.

7. A main board characterized by comprising: The second connector matches the first connector of the multi-GPU cascaded card according to any one of claims 1-6. The first connector is connected with the second connector to realize the installation of the multi-GPU cascaded card on the mainboard. ​ The first connector and the second connector are provided with PCIE signal lines; The PCIE signal lines of the first connector are connected with the GPU, and the PCIE signal lines of the second connector are connected with the CPU; The second connector is arranged at a position between the CPU and a front window of the case or at a position between the CPU and a rear window of the case.

8. The main plate of claim 7, wherein, The number of the CPUs is several; The first connector and the second connector adopt a board-to-board mirror sandwich connector.

9. A liquid-cooled heat-dissipating chassis, characterized in that, The case shell is sealed. The case shell is provided with a mainboard, and the mainboard adopts the mainboard according to any one of claims 7-8. The first cold plate is arranged at the CPU of the mainboard, and the second cold plate is arranged at the multi-GPU cascade buckle card. First adjusting valves are arranged at the liquid inlet and the liquid outlet of the first cold plate, and second adjusting valves are arranged at the liquid inlet and the liquid outlet of the second cold plate. Second temperature sensors are arranged at the CPUs. A liquid cooling controller is further arranged on the mainboard, and the liquid cooling controller is connected with the first adjusting valves, the second adjusting valves, the first temperature sensors and the second temperature sensors.

10. A liquid-cooled thermal chassis, characterized by, The case shell is sealed. The case shell is provided with a mainboard, and the mainboard adopts the mainboard according to any one of claims 7-8. The lower part of the sealed case shell is provided with a liquid inlet, and the upper part of the sealed case shell is provided with a liquid outlet. A first water distributor is arranged at a set distance close to the CPU, and a second water distributor is arranged at a set distance close to the GPU. The first water distributor and the second water distributor are communicated with the liquid inlet. A liquid level detection sensor and an air pressure sensor are arranged on the inner top of the sealed case shell. Second temperature sensors are arranged at the CPUs. A liquid cooling circulation controller is further arranged on the mainboard, and the liquid cooling circulation controller is connected with the liquid level detection sensor, the air pressure sensor, the first temperature sensors, the second temperature sensors, the first water distributor and the second water distributor.

Images