CN216133362U

CN216133362U - 4-node GPU server

Info

Publication number: CN216133362U
Application number: CN202121047326.1U
Authority: CN
Inventors: 崔海英; 区志辉; 李明春; 郭立书; 臧红卫
Original assignee: Well Innovation Tianjin Science And Technology Development Co ltd
Current assignee: Kuangzhi Zhongke Beijing Technology Co ltd
Priority date: 2021-05-15
Filing date: 2021-05-15
Publication date: 2022-03-25
Anticipated expiration: 2031-05-15

Abstract

The utility model belongs to the technical field of data processing, and particularly relates to a 4-node GPU server which comprises a chassis, 4 IO modules, a group of disk modules, 4 GPU nodes, 4 fan modules, 1 backboard system and 2 power supply modules. This 4 node GPU server, through setting up quick-witted case, 4 IO modules, a set of disk module, 4 GPU nodes, 4 fan modules, 1 backplate system, 2 power module, reached through this practical realization solution complete machine adopts the modularized design, but GPU operation node direct hot plug, the GPU trouble only influences the monolithic GPU card, it is convenient to maintain, every GPU corresponds a slice network card, complete machine network bandwidth maximize, the GPU node supports GPU, the FPGA mixes with higher speed, and it is nimble to use, the effect of two PCI-E equipment of maximum support.

Description

4-node GPU server

Technical Field

The utility model relates to the technical field of data processing, in particular to a 4-node GPU server.

Background

The 4U multislot GPU server chassis can support a plurality of high-performance GPU cards and PCIe card expansion supporting other functions in a 4U standard chassis, and is widely applied to the fields of high-performance calculation, big data analysis and artificial intelligence; the system plays an important role in high-performance environment and super computing field by matching with the latest English WEI GPU card.

The traditional rack-mounted GPU server can only support +1 GPU cards of a single node, and each GPU server comprises power supply equipment inside. Because the number of GPU cards which can be used in a limited space is very limited, the processing capacity of a GPU server is greatly limited, and meanwhile, the following defects exist:

1. the GPU of the conventional GPU operation server is directly arranged in a case, the fault influence surface is too large when the GPU is damaged and maintained, and the case needs to be opened after the whole GPU operation server is shut down to be maintained difficultly.

2. The whole layout of the conventional GPU operation server is designed according to the 2 slot positions of the public version card, and the width of the whole GPU operation server is seriously wasted when a display card of the 3 slot positions of a new version is installed.

SUMMERY OF THE UTILITY MODEL

The utility model provides a 4-node GPU server based on the technical problems that the GPU of the existing GPU operation server is directly arranged in a case, the fault influence surface is too large when the GPU is damaged and maintained, the case is required to be opened after the whole GPU operation server is shut down and the maintenance is not good, and the whole GPU operation server is integrally arranged according to the design of a public version card 2 slot position and the waste of the whole GPU width is serious when a new version 3 slot position display card is installed.

The utility model provides a 4-node GPU server which comprises a case, 4 IO modules, a group of disk modules, 4 GPU nodes, 4 fan modules, 1 back plate system and 2 power modules, wherein the IO modules are installed at the lower front end of the case, the disk modules are installed in the middle of the vertical direction in front of the case, the GPU nodes are installed at the rear end inside the case, the fan modules are installed at the upper front end of the case, the back plate system is installed in the longitudinal middle of the case, the power modules are installed at the rear of the case, and guide rails are arranged inside the case.

Preferably, the IO module corresponds to the GPU node located at the rear end inside the chassis, and provides network and display support for each GPU node, the IO module is divided into four independent units, and signal connection ends thereof are connected to the backplane system through the guide rail located in the chassis;

through the technical scheme, 4 IO modules, a group of disk modules, 4 GPU nodes, 4 fan modules, 1 backboard system and 2 power modules are composed of 1 PCBA board, and connector interfaces for connecting different modules are correspondingly arranged on the PCBA board.

Preferably, the disk module specifically comprises eight independent 3.5-inch disks, and corresponds to the GPU nodes located at the rear end inside the chassis, wherein each GPU node is allocated with two independent 3.5-inch disks, and a signal connection end of the GPU node is connected to the backplane system through a hot-plug hard disk bracket and the guide rail located in the chassis;

through the technical scheme, each GPU node is distributed with two independent 3.5-inch disks to meet the storage requirement of each GPU node, and each disk module is maintained before use.

Preferably, the GPU nodes are computing modules arranged in an array, adjacent GPU nodes are equidistant, and the GPU nodes are parallel to each other;

through the technical scheme, the 4 GPU node modules are parallel to each other, and the distances between any two GPU nodes are equal, so that a neat array is formed. Each GPU node module shell adopts a semi-closed shell, and a certain heat dissipation hole is formed in the back of the shell, so that heat dissipation can be achieved for installing a GPU and an FPGA accelerator card in the GPU and through a fan unit and a fan arranged in each GPU node, each GPU node module is provided with an independent locking device, and the GPU node module can be inserted into a backboard system or pulled out of the backboard system in the horizontal direction.

Preferably, the fan modules are four independent parts, the signal connection ends of the fan modules are electrically connected with the signal connection end of the back panel system, and each fan module comprises two high-speed fans;

through the technical scheme, a certain space distance is reserved between the air outlet of the fan and the part needing heat dissipation, so that the heat dissipation of the part needing heat dissipation in the whole air duct is facilitated. The fan module adopts a clamping structure, so that quick online replacement and maintenance can be realized.

Preferably, the power module supplies power to the IO module, the disk module, the GPU node, and the fan module through corresponding connection terminals on the backplane;

through the technical scheme, the power module adopts a standard crps power supply, two groups of power supplies are arranged according to a 1+1 hot standby mode, the normal operation of the whole machine can still be guaranteed after a single fault occurs, and the power module is provided with the locking mechanism, so that tool-free quick installation and replacement can be realized.

The beneficial effects of the utility model are as follows:

through setting up the quick-witted case, 4 IO modules, a set of disk module, 4 GPU nodes, 4 fan module, 1 backplate system, 2 power module, reached through the utility model discloses the solution complete machine adopts the modularized design, GPU operation node can direct hot plug, GPU trouble only influences the monolithic GPU card, it is convenient to maintain, every GPU corresponds a slice network card, complete machine network bandwidth maximize, GPU node support GPU, FPGA mixes with higher speed, use in a flexible way, the effect of the biggest support two PCI-E equipment, current GPU operation server GPU direct dress in quick-witted case has been solved, the trouble influence plane is too big when GPU damages the maintenance, it is not good to maintain to open quick-witted case maintenance after needing the complete machine to shut down, and current GPU operation server wholly arranges according to public version card 2 slot designs, the extravagant serious problem of complete machine 3 slot width when installing new edition display card.

Drawings

Fig. 1 is a schematic diagram of a 4-node GPU server according to the present invention;

fig. 2 is a perspective view of a power module structure of a 4-node GPU server according to the present invention.

In the figure: 1. a chassis; 2. an IO module; 3. a disk module; 4. a GPU node; 5. a fan module; 6. a backplane system; 7. and a power supply module.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments.

Referring to fig. 1-2, a 4-node GPU server includes a

chassis

1, 4 IO modules 2, a set of

disk modules

3, 4

GPU nodes

4, 4

fan modules

5, 1

backplane system

6, and 2 power modules 7, where the IO modules 2 are installed at the front lower end of the chassis 1, the disk modules 3 are installed in the front vertical middle of the chassis 1, the GPU nodes 4 are installed at the internal rear end of the chassis 1, the fan modules 5 are installed at the front upper end of the chassis 1, the backplane system 6 is installed in the longitudinal middle of the chassis 1, and is a key bridge component of the system, the front fan modules 5, the disk modules 3, and the IO modules 2 are inserted into the backplane system 6 at corresponding positions, and the rear GPU nodes 4 and the power modules 7 are inserted into the corresponding positions of the backplane system 6 at set positions; power module 7 installs in quick-witted case 1's rear, and for the convenience of plug node module, the inside of quick-witted case 1 is provided with the guide rail.

Furthermore, the IO module 2 corresponds to the GPU nodes 4 located at the rear end inside the chassis 1, and provides network and display support for each GPU node 4, the IO module 2 is divided into four independent units, and signal connection ends thereof are connected to the backplane system 6 through guide rails located in the chassis 1;

4 IO modules 2, a set of

disk modules

3, 4

GPU nodes

4, 4

fan modules

5, 1

backplane system

6, 2 power modules 7 are composed of 1 PCBA board, and connector interfaces for connecting different modules are correspondingly arranged on the PCBA board.

Further, the disk module 3 specifically includes eight independent 3.5-inch disks, and corresponds to the GPU node 4 located at the rear end inside the chassis 1, wherein each GPU node 4 is allocated with two independent 3.5-inch disks, and a signal connection end thereof is connected to the backplane system 6 through a hot-plug hard disk bracket and a guide rail located in the chassis 1;

each GPU node 4 is allocated two independent 3.5-inch disks to meet the storage requirements of each GPU node 4, and in addition, each disk module 3 is maintained before use.

Furthermore, the GPU nodes 4 are computing modules arranged in an array, the adjacent GPU nodes 4 are equidistant, and the GPU nodes 4 are parallel to each other;

the 4 GPU node 4 modules are parallel to each other and any two GPU nodes 4 are equidistant from each other, thus forming an ordered array. Each GPU node 4 module shell adopts a semi-closed shell, and a certain heat dissipation hole is formed in the back of the shell, so that heat dissipation can be achieved for installing a GPU and an FPGA accelerator card in the GPU and the fan built in each GPU node 4, each GPU node 4 module is provided with an independent locking device, and the GPU nodes can be inserted into a backboard system 6 or pulled out of the backboard system 6 in the horizontal direction.

Furthermore, the fan modules 5 are four independent parts, the signal connection end of the fan module 5 is electrically connected with the signal connection end of the back panel system 6, and each fan module 5 comprises two high-speed fans;

a certain space distance is reserved between the air outlet of the fan and the part needing heat dissipation, so that the heat dissipation of the part needing heat dissipation in the whole air duct is facilitated. The fan module 5 adopts a clamping structure, and can realize quick online replacement and maintenance.

Further, the power module 7 supplies power to the IO module 2, the disk module 3, the GPU node 4, and the fan module 5 through corresponding connection ends on the backplane;

the power module 7 adopts a standard crps power supply, two groups of power supplies are arranged in a 1+1 hot standby mode, normal operation of the whole machine can still be guaranteed after a single fault occurs, and the power module 7 is provided with a locking mechanism, so that tool-free rapid installation and replacement can be realized.

By arranging the case 1, the 4 IO modules 2, the group of disk modules 3, the 4 GPU nodes 4, the 4 fan modules 5, the 1 backboard system 6 and the 2 power supply modules 7, the modular design of the whole machine is solved, the GPU operation nodes can be directly hot-plugged, GPU faults only affect a single GPU card, the maintenance is convenient, each GPU corresponds to one network card, the network bandwidth of the whole machine is maximized, the GPU nodes 4 support the GPUs, the FPGA is accelerated in mixing and flexible in application, the effect of maximally supporting two pieces of PCI-E equipment is achieved, the problems that the GPU of the conventional GPU operation server is directly arranged in the case 1, the fault influence surface is too large when the GPU is damaged and maintained, the case 1 is opened after the whole machine is shut down, and the maintenance is not good, and the problem that the whole layout of the conventional GPU operation server is designed according to the 2 slot positions of the public version card, and the width of the whole GPU operation server is seriously wasted when a display card of the 3 slot positions of a new version is installed.

The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art should be considered to be within the technical scope of the present invention, and equivalent alternatives or modifications according to the technical solution of the present invention and the inventive concept thereof should be covered by the scope of the present invention.

Claims

1. The utility model provides a 4 node GPU server, includes quick-witted case (1), 4 IO modules (2), a set of disk module (3), 4 GPU nodes (4), 4 fan module (5), 1 backplate system (6) and 2 power module (7), its characterized in that: the intelligent computer case is characterized in that the IO module (2) is installed at the front lower end of the case (1), the disk module (3) is installed in the middle of the front vertical direction of the case (1), the GPU node (4) is installed at the rear end of the inside of the case (1), the fan module (5) is installed at the upper end of the front of the case (1), the backboard system (6) is installed at the longitudinal middle of the case (1), the power module (7) is installed at the rear of the case (1), and a guide rail is arranged inside the case (1).

2. The 4-node GPU server of claim 1, wherein: IO module (2) with be located quick-witted case (1) inside rear end GPU node (4) are corresponding, for every GPU node (4) provide network, show and support, IO module (2) divide into four independent units, its signal connection end is through being located quick-witted case (1) guide rail access backplate system (6).

3. The 4-node GPU server of claim 1, wherein: the disk module (3) specifically comprises eight independent 3.5-inch disks and corresponds to the GPU nodes (4) located at the rear end inside the case (1), wherein each GPU node (4) is distributed with two independent 3.5-inch disks, and the signal connection end of each GPU node is connected into the backplane system (6) through a hot plug hard disk bracket and the guide rail located in the case (1).

4. The 4-node GPU server of claim 1, wherein: the GPU nodes (4) are computing modules arranged in an array mode, the adjacent GPU nodes (4) are equidistant, and the GPU nodes (4) are parallel to each other.

5. The 4-node GPU server of claim 1, wherein: the fan modules (5) are four independent parts, the signal connecting ends of the fan modules (5) are electrically connected with the signal connecting ends of the back panel system (6), and each fan module (5) comprises two high-speed fans.

6. The 4-node GPU server of claim 1, wherein: and the power supply module (7) supplies power to the IO module (2), the disk module (3), the GPU node (4) and the fan module (5) through corresponding connecting ends on the backboard.