GPU server

Abstract

The utility model relates to the technical field of servers, in particular to a GPU (Graphic Processing Unit) server, which comprises a first box body and a second box body which are detachable, the two box bodies are respectively provided with a power supply module and are independently powered, and at most 10 GPU board cards are arranged in each box body. When the computing power of the first box body can meet the requirement, the first box body serves as an independent 4U height GPU server, and the computing power of 10 GPU cards can be provided. When the computing power of the first box body does not meet the requirement, the first box body 4 and the second box body 10 are assembled and connected, and the computing power of 20 GPU cards is provided at most.

Description

TECHNICAL FIELD

[0001] The utility model relates to server technical field, especially GPU server technical field. BACKGROUND

[0002] GPU server is a kind of computer server and its supporting equipment for computer science field, because of its computing ability, energy efficiency ratio is high, the advantage has been widely used in artificial intelligence, big data etc.

[0003] The number of GPU board cards in single GPU server is one of important indexes to measure the computing power of GPU server.At present, with the continuous improvement of CPU performance, high-performance CPU can provide more PCIe channel quantity, and create conditions for deploying more GPU board cards for GPU server expansion more PCIe card slot.

[0004] However, the existing GPU server is equipped with high-performance, high-power CPU and GPU, and the traditional power module design cannot meet the power supply demand of a large number of GPU board cards.Meanwhile, a large amount of heat is generated when the server chip works;There are other components in the case, such as high-capacity, high-power memory, a large number of high-speed solid state or mechanical hard disk, high-power power supply, these server components also generate a large amount of heat;The existing server cooling system is difficult to cool the chip and board card due to the limitation of case layout, PCB layout and other aspects;If GPU board card is forcibly increased, it will inevitably cause poor heat dissipation, heat accumulation and temperature rise in the case;Lead to the performance of each component of the server is reduced, and even cannot work.

[0005] Based on the above reasons, the existing GPU server cannot fully utilize CPU resources, fully expand the number of PCIe card slots and further improve the performance of the server.At present, in the development of computing-intensive work, such as in large-scale data rendering scene, because single GPU server cannot meet the demand, only multiple GPU servers can work simultaneously, and because of the cooperation of multiple servers, the cost is increased. INVENTION CONTENTS

[0006] The utility model provides a kind of GPU server to solve the problem of insufficient computing power of single GPU server in prior art.

[0007] The GPU server provided by the utility model comprises:

[0008] The first box body 4 and the second box body 10 fixedly connected with the first box body 4; the first box body 4 is 4U high, and a mainboard 6 is arranged in the first box body 4; a hard disk group back plate 2 is arranged in front of the mainboard 6.

[0009] A first power module 7 is arranged at the rear end of the first box body 4, and the first power module 7 provides power for components arranged in the first box body 4; the first power module 7 is directly connected with a power connector on the mainboard 6.

[0010] The second box body 10 is 4U high and located above the rear end of the first box body 4.

[0011] A power adapter plate 11 and a second power module 12 are arranged at the rear end of the second box body 10; the second power module 12 provides power for components arranged in the second box body 10; the second power module 12 is connected with a power connector on the power adapter plate 11; the power adapter plate 11 is connected with the mainboard 6 through a cable, so that management signal transmission is realized.

[0012] The second box body 10 is arranged, so that the case space is expanded; the first power module 7 provides power for components in the first box body 4, and the second power module 12 provides power for components in the second box body 10, so that the problem of too many board cards and insufficient power supply in the prior art is solved.

[0013] Meanwhile, by setting the second box 10, the second PCIe adapter board 13 is set in the second box 10, and the PCIe slot is expanded, which can accommodate more GPU cards. In order to expand more GPU cards:

[0014] Further, 11 PCIe slots are set on the first PCIe adapter board 8, and 10 first GPU cards 9 are inserted into the PCIe slots on the first PCIe adapter board 8.

[0015] As a preferred, the first GPU card 9 is a double-wide GPU card.

[0016] Further, 11 PCIe slots are set on the second PCIe adapter board 13, and 10 second GPU cards 14 are inserted into the PCIe slots on the second PCIe adapter board 13.

[0017] As a preferred, the second GPU card 13 is a double-wide GPU card.

[0018] In addition, by setting the second box 10, the high-power chips and boards are respectively set in the two cases, which creates conditions for solving the heat dissipation problem.

[0019] For the heat dissipation of the first box 4, a fan frame 5 is arranged between the backboard 2 and the mainboard 6 in the middle of the first box 4, which is used to install fans; a plurality of first fan modules 20 can be installed in the fan frame 5, and after the first fan modules 20 are installed into the fan frame 5, the power supply connector on the fan will directly interface with the mainboard 6, thereby realizing the power supply of the fan.

[0020] As a preferred, 4 first fan modules 20 are installed in the fan frame 5, which can still provide stable heat dissipation capacity when part of the fans fail. Each first fan module 20 is independently controlled in start-stop and speed, which can provide more accurate heat dissipation and is more energy-saving.

[0021] Further, the first fan module 20 includes an upper fan and a lower fan, each fan has a power supply connector, the connector is fixed on the first fan module 20, and the upper fan and the lower fan are independently controlled.

[0022] Further, a wind guide cover 19 is arranged above the mainboard 6 in the middle of the first box 4; the wind guide cover 19 is located above the CPU and the memory, separates the air ducts of the upper fan and the lower fan in the first fan module 20 in the middle of the first box 4, and forms two air flow channels; the lower fan in the first fan module 20 blows directly to the CPU and the memory through the lower air flow channel, and the CPU and the memory are cooled by heat dissipation; the upper fan in the first fan module 20 blows directly to the installed first GPU card 9 through the upper air flow channel, and the plurality of first GPU cards 9 are cooled by heat dissipation.

[0023] For the heat dissipation of the second box body 10, a second box body front panel 17 is arranged at the front end of the second box body 10; a plurality of second fan modules 16 are fixed on the second box body front panel 17, the second fan modules 16 are connected with the power adapter board 11 through connectors, and the second fan modules 16 provide heat dissipation for the second GPU cards 14.

[0024] As a preferred, the number of the second fan modules 16 is 3, each second fan module 16 is independently controlled to start and stop and rotate at a speed, so that more accurate heat dissipation and more energy saving can be provided.

[0025] The GPU server of the utility model, through 8U case layering design, supports 20 double-wide GPU cards; at the same time, power layering power supply, the first power module supplies power for the mainboard, the first fan module, the hard disk and the first GPU card, and the second power module supplies power for the second GPU card; further, the first box body can support mainboards of different platforms, the equipment of the second box body is removed, the first box body and internal components can be used as a 4U server, and 10 double-wide GPU cards can be supported. Through the above design, the GPU server of the utility model can adopt but is not limited to the following two configurations: the first kind: 12 3.5-inch hard disks, 2 350W high-power CPUs, 20 350W or more high-power GPU cards and 32 64G memory bars; the second kind: 12 3.5-inch hard disks, 2 500W high-power CPUs, 20 350W or more high-power GPU cards and 24 64G memory bars. The GPU server of the utility model can satisfy the air cooling heat dissipation of the two kinds of configuration servers, and the computing power of the GPU server single machine is greatly improved. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 The GPU server structure decomposition schematic view provided by the utility model is shown in the figure;

[0027] Figure 2 The GPU server front schematic view provided by the utility model is shown in the figure;

[0028] Figure 3 The GPU server rear schematic view provided by the utility model is shown in the figure.

[0029] Label explanation:

[0030] 1. hard disk group; 2. hard disk group back plate; 3. first hanging ear; 4. first box body; 5. fan frame; 6. mainboard; 7. first power module; 8. first PCIe adapter plate; 9. first GPU card; 10. second box body; 11, power adapter plate; 12, second power module; 13, second PCIe adapter plate; 14, second GPU card; 15, second box body upper cover; 16, second fan module; 17, second box body front panel; 18, first box body middle cover; 19, air baffle; 20, first fan module; 21, first box body front cover; 22, second hanging ear DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the utility model will be clearly and completely described below with reference to the drawings in the embodiments of the utility model. Obviously, the described embodiments are only part of the embodiments of the utility model, rather than all the embodiments. Based on the embodiments in the utility model, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the utility model.

[0032] As Figure 1 The GPU server provided by the embodiments of the utility model comprises:

[0033] The first box body 4 with the height of 4U is located at the lower part of the GPU server, and the first box body 4 has the standard 4U case size in the field;

[0034] In the first box body 4, the mainboard 6 is arranged, and the hard disk group back plate 2 is arranged in front of the mainboard 6;

[0035] The hard disk group 1 is arranged at the front end of the first box body 4, the hard disk group 1 is movably connected with the hard disk group back plate 2, one or more hard disks can be plugged into the hard disk group 1 according to the need, the hard disk group 1 is connected with the hard disk group back plate 2 through the interface of the hard disk group back plate 2, and the hard disk group 1 is connected with the mainboard 6 through a cable;

[0036] The first hanging ear 3 is arranged at the right side of the front end of the first box body 4, a power switch and a signal indicator light plate are arranged in the first hanging ear 3, the power switch is used for starting or shutting down the server, and the signal indicator light plate displays the running state of the server through the color, brightness, flicker frequency and other states of the indicator light. The first hanging ear 3 is connected with the mainboard 6 through a cable, the first hanging ear 3 transmits a power switch operation signal to the server, receives an indicator light display signal from the server, and is used for indicator light display;

[0037] The second hanging ear 22 is arranged at the left side of the front end of the first box body 4 and corresponds to the first hanging ear 3; a USB and VGA signal plate is arranged in the second hanging ear 22, and the second hanging ear 22 is connected with the mainboard 6 through a cable;

[0038] In the middle section of the first box body 4, between the back plate 2 and the main plate 6, a fan frame 5 is arranged for installing fans;

[0039] A plurality of first fan modules 20 can be installed in the fan frame 5. The first fan modules 20 are installed in the fan frame 5, and the power supply connectors on the fans are directly connected to the main plate 6 to realize power supply of the fans;

[0040] As a preferred embodiment, the number of first fan modules 20 installed in the fan frame 5 is four, and each first fan module 20 is independently controlled in terms of start-stop and rotation speed. Each first fan module 20 is independently controlled, so that the heat dissipation strategy can be adjusted according to the actual situation in the first box body 4, thereby improving the heat dissipation efficiency; because the number of first fan modules 20 is redundant, the stability of heat dissipation is improved;

[0041] Further, the first fan module 20 has two fans, namely an upper fan and a lower fan, and each fan has a power supply connector fixed on the first fan module 20. The upper fan and the lower fan are independently controlled; during use, the rotation speed of the upper fan or the lower fan corresponding to a component can be adjusted according to the temperature of the component, so as to achieve precise control.

[0042] Further, the double-rotor fan with model number 8056 has the advantages of high air output, high efficiency, low power consumption, and high redundancy. As a preferred embodiment, the first fan module 20 selects the 8056 fan module.

[0043] At the rear end of the first box body 4, a plurality of first power supply modules 7 are arranged, and the first power supply modules 7 are directly connected to the power supply connectors on the main plate 6;

[0044] The number of first power supply modules 7 can be set according to actual conditions, and is preferably four, which can ensure the reliability of power supply through redundant design;

[0045] At the rear end of the first box body 4, above the first power supply modules 7, a first PCIe adapter plate 8 is arranged, and a plurality of PCIe slots are arranged on the first PCIe adapter plate 8. The first PCIe adapter plate 8 is connected to the main plate 6 through a cable to realize PCIe signal transmission and power supply. A first GPU card 9 is inserted into the PCIe slot of the first PCIe adapter plate 8, and is connected to the main plate 6 through a cable to realize auxiliary power supply of the first GPU card 9;

[0046] As a preferred embodiment, 11 PCIe slots are arranged on the first PCIe adapter plate 8, and 10 first GPU cards 9 are inserted into the PCIe slots of the first PCIe adapter plate 8.

[0047] As a preferred embodiment, the first GPU card 9 selects a double-wide GPU card.

[0048] In the middle of the first box 4, the air deflector 19 is arranged above the mainboard 6; the air deflector 19 is located above the CPU and the memory, separates the air ducts of the lower fans and the upper fans in the plurality of first fan modules 20 in the first box 4, and forms two upper and lower air flow channels; the lower fans in the first fan modules 20 blow directly on the CPU and the memory through the lower air flow channel, so that the heat dissipation of the CPU and the memory is solved; the upper fans in the first fan modules 20 blow directly on the installed first GPU card 9 through the upper air flow channel, so that the heat dissipation of the plurality of first GPU cards 9 is solved.

[0049] The first box middle cover 18 is arranged above the middle of the first box 4, and the two are fixedly connected through screws.

[0050] The first box front cover 21 is arranged above the front end of the first box 4, and the two are fixedly connected through screws.

[0051] The first box middle cover 18 and the first box front cover 21 are arranged respectively, so that the two can be conveniently disassembled.

[0052] The second box upper cover 15 is arranged above the rear end of the first box 4, and the two are fixedly connected through screws.

[0053] The foregoing hard disk group 1, the hard disk group back plate 2, the first hanging ear 3, the first box 4, the fan frame 5, the mainboard 6, the first power supply module 7, the first PCIe adapter board 8, the first GPU card 9,

[0054] The first box middle cover 18, the air deflector 19, the first fan module 20, the first box front cover 21, the second hanging ear 22, and the second box upper cover 15

[0055] The basic part of the GPU server, the basic part can be used as an independent 4U server, the first PCIe adapter board 8 can be connected with 10 first GPU cards 9, and the computing power of 10 GPU cards is provided.

[0056] In some scenarios, the server with 10 GPU cards cannot meet the demand, in order to flexibly increase the computing power, the GPU server also provides an expansion part.

[0057] The expansion part comprises a second box 10 with a height of 4U, the second box 10 is located above the rear end of the first box 4, after the second box upper cover 15 arranged above the rear end of the first box 4 is removed, the second box 10 is fixed on the first box 4 through screws, the height and the width of the second box are the same as those of a standard 4U case in the field, and the length can be telescopic according to actual conditions;

[0058] The power adapter board 11 and the second power supply module 12 are arranged at the rear end of the second box 10, the number of the second power supply module 12 can be set according to actual conditions, and preferably is 4, and the reliability of power supply can be ensured through redundant design.

[0059] The second power module 12 can be directly connected with the power connector on the power adapter board 11, and the power adapter board 11 is connected with the mainboard 6 through a cable to realize the transmission of management signals.

[0060] The second GPU card 14 is connected with the PCIe slot of the second PCIe adapter board 13, and is connected with the power adapter board 11 through a cable to realize the auxiliary power supply of the second GPU card 14.

[0061] The second PCIe adapter board 13 is arranged above the second power module 12 at the rear end of the second box 10, and a plurality of PCIe slots are arranged on the second PCIe adapter board 13. The second PCIe adapter board 13 is connected with the power adapter board 11 through a cable, and the second GPU card 14 is connected with the PCIe slot of the second PCIe adapter board 13, and is connected with the power adapter board 11 through a cable to realize the auxiliary power supply of the second GPU card 14.

[0062] The second PCIe adapter board 13 is connected with the mainboard 6 in the first box 4 through a cable to realize the transmission of PCIe signals.

[0063] As a preferred embodiment, 11 PCIe slots are provided on the second PCIe adapter board 13, and 10 second GPU cards 14 are connected with the PCIe slots on the second PCIe adapter board 13.

[0064] As a preferred embodiment, the second GPU card 14 is a double-wide GPU card.

[0065] A second box upper cover 15 is arranged above the second box 10, and the second box upper cover 15 is fixed on the top of the 4U upper box 10 through screws to prevent dust from entering.

[0066] A second box front panel 17 is arranged at the front end of the second box 10, and a plurality of second fan modules 16 can be fixed on the second box front panel 17. The second fan modules 16 are connected with the power adapter board 11 through connectors, and the second fan modules 16 provide heat dissipation for the second GPU card 14.

[0067] As a preferred embodiment, the number of the second fan modules 16 is 3, and each second fan module 16 is independently controlled in terms of start-stop and rotation speed. In this way, the heat dissipation strategy can be adjusted according to the actual situation in the second box 10, the heat dissipation efficiency is improved, and the stability of heat dissipation is improved due to the redundancy.

[0068] Further, the fan with model number 12038 has good stability and durability, has high air volume and air pressure, and as a preferred embodiment, the second fan module 16 uses a 12038 fan module.

[0069] The second box body 10, the power adapter board 11, the second power module 12, the second PCIe adapter board 13, the second GPU card 14, the second box body upper cover 15, the second fan module 16 and the second box body front panel 17

[0070] The expansion part constitutes the GPU server of the utility model, the expansion part cannot be used independently, can be used together with the basic part as an 8U server, and the total highest power of 20 GPU cards is provided.

[0071] The first box body 4 and the second box body 10 are detachable designs, when the power of the basic part can meet the requirements, the basic part is used as an independent 4U height GPU server, and the power of 10 GPU cards can be provided, when the power of the basic part does not meet the requirements, the first box body 4 and the second box body 10 are connected, and the total highest power of 20 GPU cards is provided by the basic part and the expansion part.

[0072] The first GPU card 9 and the second GPU card 14 are located in two different cases, and different heat dissipation systems are used for heat dissipation, so that the problem of difficult heat dissipation of a large number of high-energy-consumption boards in a small space is solved.

[0073] It should be noted that in this paper, the terms "upper", "lower", "left", "right", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, and are only used for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, relationship terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations, nor can it be understood as indicating or implying relative importance. Moreover, the term "includes" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed or inherent to such process, method, article or terminal device.

[0074] The above describes in detail the 8U height split GPU server provided by the application. The principles and implementation manners of the application are described by using specific examples. The above description of the embodiments is only used to help

[0075] The content of the specification should not be understood as a limitation of the application. For those skilled in the art, according to the application, there will be different forms of changes in the specific implementation manners and application ranges, which do not need and cannot be exhausted here, and the obvious changes or variations derived therefrom are still within the protection scope of the application.

Claims

1. A GPU server, comprising: A first box (4) with a height of 4U; A mainboard (6) is provided in the first box (4), and a hard disk assembly backboard (2) is provided in front of the mainboard (6); A hard disk assembly (1) is provided at the front end of the first box body (4), and the hard disk assembly (1) is movably connected to the hard disk assembly backplane (2); A first power supply module (7) is provided at the rear end of the first box (4), and the first power supply module (7) is connected to the power supply connector of the mainboard (6); A first PCIe adapter board (8) is provided at the rear end of the first box (4) and above the first power supply module (7), and a plurality of PCIe slots are provided on the first PCIe adapter board (8); The first PCIe adapter board (8) is connected to the mainboard (6) via a cable to achieve PCIe signal transmission and power supply; The first GPU card (9) is plugged into the PCIe slot of the first PCIe adapter board (8) and is connected to the mainboard (6) via a cable for auxiliary power supply to the first GPU card (9); It is characterized by: also including A second box (10) with a height of 4U is located above the rear end of the first box (4) and is fixedly connected to the first box (4); The rear end of the second box (10) is provided with a power adapter plate (11) and a second power module (12); The second power module (12) is connected to the power connector on the power adapter board (11), and the power adapter board (11) is connected to the main board (6) via a cable to achieve management signal transmission; A second PCIe adapter board (13) is provided at the rear end of the second box (10) and above the second power supply module (12), and a plurality of PCIe slots are provided on the second PCIe adapter board (13); The second PCIe adapter board (13) is connected to the power adapter board (11) via a cable; The second GPU card (14) is plugged into the PCIe slot of the second PCIe adapter board (13) and is connected to the power adapter board (11) via a cable to provide auxiliary power to the second GPU card (14); The second PCIe adapter board (13) is connected to the main board (6) via a cable to achieve PCIe signal transmission.

2. The GPU server according to claim 1, wherein: A fan frame (5) is provided in the middle section of the first housing (4), between the hard disk assembly backplane (2) and the mainboard (6); A plurality of first fan modules (20) can be installed in the fan frame (5), and a power supply connector on the fan is connected to the mainboard (6) to realize power supply for the fan.

3. The GPU server according to claim 2, wherein: Four first fan modules (20) are installed in the fan frame (5), and each first fan module (20) is independently controlled to start, stop and rotate speed; The first fan module (20) includes an upper fan and a lower fan, and the upper fan and the lower fan are controlled separately; An air guide cover (19) is provided in the middle of the first housing (4) above the mainboard (6); the air guide cover (19) is located above the CPU and the memory and is used to isolate the upper and lower fan ducts in the first fan module (20).

4. The GPU server according to claim 1, wherein: A second box front panel (17) is provided at the front end of the second box (10); a plurality of second fan modules (16) are fixed on the second box front panel (17), and the second fan modules (16) provide heat dissipation for the second GPU card (14).

5. The GPU server according to claim 4, wherein: The number of the second fan modules (16) is 3, and each second fan module (16) is independently controlled to start, stop and rotate speed.

6. The GPU server according to claim 1, wherein: Eleven PCIe slots are provided on the first PCIe adapter board (8), and ten of the first GPU cards (9) are plugged into the PCIe slots on the first PCIe adapter board (8); The first GPU card (9) is a double-width GPU card; Eleven PCIe slots are provided on the second PCIe adapter board (13), and ten of the second GPU cards (14) are plugged into the PCIe slots on the second PCIe adapter board (13); The second GPU card (14) is a double-width GPU card.

7. The GPU server according to claim 1, wherein: The first power supply modules (7) are provided with four groups; The second power supply modules (12) are provided in four groups.

8. The GPU server according to any one of claims 1 to 7, wherein: A first hanging ear (3) is provided on the right side of the front end of the first box body (4), a power switch and a signal indicator light panel are provided inside the first hanging ear (3), and the first hanging ear (3) is connected to the main board (6) via a cable.

9. The GPU server according to claim 8, wherein: A second hanging ear (22) is provided on the left side of the front end of the first box body (4); a USB and VGA signal board is provided inside the second hanging ear (22), and the second hanging ear (22) is connected to the mainboard (6) via a cable.

10. The GPU server according to claim 9, wherein: A first box body middle cover (18) is provided above the middle portion of the first box body (4), and the two are fixedly connected; A first box body front cover (21) is provided above the front end of the first box body (4), and the two are fixedly connected; A second box upper cover (15) is provided above the second box (10), and the two are fixedly connected.

Images