AI server architecture system

Abstract

The invention relates to the technical field of computers, and discloses an AI server architecture system which comprises a case. The CPU module is installed on the lower layer of the case and comprises a mainboard, a first NVMe module, an IO board, a first PCIe module and an SATA module. The GPU module is installed on the upper layer of the case and comprises a UBB board card, a fixing frame and a plurality of AI board cards, the AI board cards are connected with slots in the UBB board card through golden finger dual-density connectors, and the adjacent AI board cards are in signal connection; the HIB module and the QSFP-56 module are installed on the upper layer of the case, the HIB module is connected with the UBB board card through the BTB connector, the HIB module is in signal connection with the mainboard, the second PCIe module and the second NVMe module, and the QSFP-56 module is in signal connection with the UBB board card. According to the invention, the computing and image processing capabilities of the server can be improved, the cost of switches and network cards is saved, the expansion is convenient, and the flexibility is good.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to an AI server architecture system. Background Art

[0002] With the rapid development of the Internet and the information industry, all kinds of sound, image, and video data have experienced explosive growth. Big data processing has gradually replaced traditional manual data processing, and the application of artificial intelligence (AI) technology has made big data analysis and processing capabilities leap forward again. Currently, artificial intelligence is developing very rapidly, creating a huge demand for computing power. It can be said that we have entered an era of explosive computing power.

[0003] Deep learning technology has triggered the rapid development of artificial intelligence applications, leading mankind from the information age to the intelligent age. Deep learning is essentially a machine learning technology that requires powerful hardware computing capabilities to complete complex data processing and operations. For such huge data processing and operations, existing artificial intelligence solutions use dedicated AI computing boards to perform deep learning operations, but even a single ultra-high-performance AI computing board has a processing power far from meeting the computing requirements, and the image processing capabilities are also insufficient. In traditional technologies, the most versatile CPU has the lowest energy efficiency ratio and cannot meet the needs of parallel processing of large amounts of data in AI computing; the hastily launched ASICs, although they solve the problem of energy efficiency, lose flexibility at the cost of not being able to meet complex real-world needs. Although GPUs are very powerful, they need to be interconnected through switches, network cards, etc. to expand to larger clusters, and this part is very costly.

[0004] Therefore, there is an urgent need for an AI server architecture system to solve the above technical problems. Summary of the invention

[0005] Based on the above, the purpose of the present invention is to provide an AI server architecture system with high-performance computing power. By integrating multiple AI boards, it can improve the server computing and image processing capabilities and meet the needs of big data parallel processing; and it is reasonably designed, which can reduce switches, network cards and other interconnection methods and save costs; at the same time, it is easy to expand and has good flexibility.

[0006] To achieve this object, the present invention adopts the following technical solutions:

[0007] AI server architecture system, including:

[0008] A chassis, comprising a base, a top cover, and a partition disposed between the base and the top cover, wherein the partition divides the interior of the chassis into an upper space and a lower space;

[0009] A CPU module is pluggably installed in the lower space of the chassis and is located at the front of the chassis. The CPU module includes a mainboard, a CPU connected to the mainboard signal, a first NVMe module, an IO board, a first PCIe module, and a SATA module;

[0010] A GPU module is installed in the upper space of the chassis, the GPU module includes a UBB board, a fixing frame and an AI module, the fixing frame is fixed on the UBB board, the AI ​​module includes a plurality of AI boards, each of the AI ​​boards is respectively plugged into the fixing frame, and the lower end of each of the AI ​​boards is respectively connected to the slot on the UBB board through a gold finger double-density connector, and adjacent AI boards are signal-connected;

[0011] The HIB and QSFP-56 modules are pluggably installed in the upper space of the chassis, including a HIB module and a QSFP-56 module, wherein the QSFP-56 module is located at the front of the chassis, and the HIB module is located between the GPU module and the QSFP-56 module; the HIB module is connected to the UBB board through a BTB connector, the HIB module is connected to the motherboard signal, and the HIB module is also signal-connected to a second PCIe module, and the QSFP-56 module is signal-connected to the UBB board;

[0012] The second NVMe module is arranged above the HIB and QSFP-56 modules and is signal-connected to the HIB module.

[0013] In some possible implementations, the AI ​​server architecture system further includes:

[0014] A first fan module is pluggably installed in the upper space of the chassis and is located at the rear of the chassis, and the first fan module is signal-connected to the HIB module;

[0015] The power supply module is pluggably installed in the lower space of the chassis and is located at the rear of the chassis. The power supply module is connected to the mainboard through a gold finger. The power supply module is electrically connected to the UBB board, the HIB module, the QSFP-56 module and the first fan module respectively.

[0016] In some possible implementations, the UBB board is arrayed with 32 slots, the number of the AI ​​boards is 32, and each AI board is plugged into each slot one by one;

[0017] The AI ​​board is integrated with a plurality of signal ports, the slots are signal-connected to the signal ports, each slot comprises a first interface and a second interface, the first interface is signal-connected to the second interface of the same slot, and the second interface is signal-connected to the first interface of an adjacent slot, so that adjacent AI boards are signal-connected.

[0018] In some possible implementations, the AI ​​board includes an AI card and a bracket, the AI ​​card is fixed on the bracket, and the bracket is vertically inserted into the fixing frame and is buckled and connected to the fixing frame.

[0019] In some possible implementations, the CPU module is integrated on the first tray, the HIB and QSFP-56 module are integrated on the second tray, the first fan module is integrated on the third tray, the power supply module is integrated on the fourth tray, and the first tray, the second tray, the third tray and the fourth tray are all pluggably installed in the chassis through a slide rail structure.

[0020] In some possible embodiments, the inner wall of the upper space of the chassis is rotatably connected to a power-assist wrench, and a card slot is provided at the bottom of the power-assist wrench. The power-assist wrench is in an inclined state under the action of gravity and the card slot is facing the direction in which the second tray is pushed; rotating fulcrums are provided on both sides of the second tray, and when the second tray, the HIB and the QSFP-56 module are pushed into the chassis, the rotating fulcrum can be inserted into the card slot and the power-assist wrench can be rotated, and then the second tray can be pushed into the chassis by pulling the power-assist wrench.

[0021] In some possible implementations, the CPU module further includes a second fan module and an air guide cover, wherein the second fan module is mounted on the first tray, the air guide cover is disposed on one side of the second fan module, and the air guide cover is used to divide the airflow of the second fan module into a memory air duct and a motherboard air duct;

[0022] A panel is detachably provided on the front side of the first tray, and eight cable sockets for connecting signals with the HIB module are provided on the front side of the mainboard. The panel is shielded on the outside of the cable sockets.

[0023] In some possible implementations, the UBB board is provided with a handle for taking and placing the GPU module; and / or,

[0024] The GPU module further includes an AI module cover, and the AI ​​module cover is detachably covered on the top of the fixing frame; and / or,

[0025] The GPU module further includes a reinforcement frame, and the reinforcement frame is fixedly connected to the fixing frame; and / or,

[0026] A cable protection cover is also provided in the chassis to protect the cables.

[0027] In some possible implementations, the first NVMe module includes four first NVMe mechanical hard disks located on the front window of the chassis;

[0028] The second NVMe module includes 8 second NVMe mechanical hard disks located on the front window of the chassis;

[0029] The SATA module includes 4 SATA solid state hard disks located on the front window of the chassis;

[0030] The first PCIe module includes a card tray and three first PCIe cards located on the front window of the chassis, the card tray is detachably mounted on the first tray, and the three first PCIe cards are pluggably mounted in the card tray;

[0031] The second PCIe module includes four second PCIe cards located on the front window of the chassis;

[0032] The QSFP-56 module includes 24 QSFP-56 optical modules and 24 CTC interconnection interfaces located on the front window of the chassis.

[0033] In some possible implementations, the first fan module includes a plurality of first fans, and each of the first fans is fixed to the third tray by a buckle;

[0034] The second fan module includes a plurality of second fans, and each of the second fans is fixed to the first tray by a buckle;

[0035] The power module is recessed inward relative to the rear window of the chassis, the power module includes 5 PSUCRPS power supplies, and the PSU CRPS power supply is pluggably arranged on the fourth tray;

[0036] The CPU module has 2 CPUs, 32 memory bars and several heat sinks;

[0037] The front of the QSFP-56 module is also provided with an AI card status indicator light and a power button and indicator light. The IO board is integrated with a UID button and indicator light, a system health indicator light, a BMC management network port, a USB interface and a VGA interface.

[0038] Beneficial effects of the present invention:

[0039] The AI ​​server architecture system provided by the present invention divides the chassis into an upper space and a lower space by a partition, and has a reasonable and compact structure, and the partition can play a good supporting role. The chassis is a 6U chassis, in which the upper space is a 4U chassis for installing the GPU module, so as to facilitate the vertical insertion of the AI ​​board into the UBB board from above; the lower space is a 2U chassis for installing the CPU module. Multiple AI boards are integrated in the GPU module, and each AI board is fixed in a fixed frame by plugging, which is convenient for plugging and unplugging and flexible in configuration; and the lower end of each AI board is connected to the slot on the UBB board through a gold finger dual-density connector, so that direct interconnection between the inside and outside of the chip is realized, and the whole is a data stream, thereby saving the expensive memory and switch costs; the signal connection between adjacent AI boards can form an AI card cluster, which greatly improves the computing power and image processing capabilities of the server, and is conducive to improving the speed of the server in processing massive data.

[0040] In addition, the AI ​​board is connected to the UBB board through a gold finger dual-density connector, and the HIB module is connected to the UBB board through a BTB connector. Not only is the connection convenient and reliable, but it also greatly reduces the use of data cables, thereby reducing the ventilation area occupied by data cables and improving the heat dissipation performance of the server. Furthermore, by setting the first NVMe module, the second NVMe module, the first PCIe module, the second PCIe module, the SATA module and the IO board, a variety of flexible replaceable modules can be provided, realizing the flexible expansion and hardware redundancy functions of the server. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a first-perspective structural diagram of the AI ​​server architecture system provided by an embodiment of the present invention;

[0042] Figure 2 is an exploded view of an AI server architecture system provided by an embodiment of the present invention;

[0043] Figure 3 It is a structural diagram of the AI ​​server architecture system provided by an embodiment of the present invention when the second NVMe mechanical hard disk is extracted;

[0044] Figure 4 It is a structural schematic diagram of the front window position (after removing the panel) of the AI ​​server architecture system provided by an embodiment of the present invention;

[0045] Figure 5 It is a structural diagram of the AI ​​server architecture system provided by an embodiment of the present invention when the CPU module is extracted;

[0046] Figure 6 It is a structural schematic diagram of the AI ​​server architecture system provided by an embodiment of the present invention when the first PCIe module is disassembled;

[0047] Figure 7 is a structural schematic diagram of the AI ​​server architecture system provided by an embodiment of the present invention when the second fan is removed;

[0048] Figure 8 is a structural schematic diagram of the AI ​​server architecture system provided by an embodiment of the present invention when the air guide cover is disassembled;

[0049] Fig. 9 It is a structural schematic diagram of the AI ​​server architecture system provided by an embodiment of the present invention when the AI ​​board is disassembled;

[0050] Fig.10 2 is a schematic diagram of the structure of the AI ​​server architecture system provided by an embodiment of the present invention from a second perspective;

[0051] Fig.11 It is a structural diagram of the AI ​​server architecture system provided by an embodiment of the present invention when the PSU CRPS power supply is disassembled;

[0052] Fig.12 is a structural schematic diagram of the AI ​​server architecture system provided by an embodiment of the present invention when the first fan is disassembled;

[0053] Fig.13 It is a structural schematic diagram of a base, a partition and a power-assisting wrench involved in an embodiment of the present invention;

[0054] Fig.14 yes Fig.13 A partial enlarged view of the middle A;

[0055] Fig.15 is a schematic diagram of the structure of a CPU module involved in an embodiment of the present invention;

[0056] Fig.16 It is a schematic diagram of the structure of a UBB board involved in an embodiment of the present invention;

[0057] Fig.17 is a schematic structural diagram of a fixing frame involved in an embodiment of the present invention;

[0058] Fig.18 is a schematic diagram of the structure of an AI module involved in an embodiment of the present invention;

[0059] Fig.19 Schematic diagram of the structure of an AI board involved in an embodiment of the present invention;

[0060] Fig. 20 It is a schematic diagram of the structure of the HIB and QSFP-56 module involved in the embodiment of the present invention;

[0061] Fig.21 is a schematic diagram of the structure of a second NVMe module involved in an embodiment of the present invention;

[0062] Fig. 22 is a schematic diagram of the structure of a second NVMe mechanical hard disk involved in an embodiment of the present invention;

[0063] Fig.23 is a structural schematic diagram of a first fan module involved in an embodiment of the present invention;

[0064] Fig.24 It is a schematic diagram of the structure of a power supply module involved in an embodiment of the present invention.

[0065] In the figure:

[0066] 1. Chassis; 11. Base; 12. Top cover; 13. Partition; 14. Power wrench; 141. Card slot;

[0067] 2. CPU module; 20. First tray; 21. First NVMe module; 211. First NVMe mechanical hard disk; 22. First PCIe module; 221. Card tray; 222. First PCIe card; 23. SATA module; 231. SATA solid state drive; 24. Second fan module; 241. Second fan; 25. Air guide cover; 26. Panel; 27. Cable socket;

[0068] 3. GPU module; 31. UBB board; 311. slot; 312. handle; 32. fixing frame; 33. AI module; 331. AI board; 3311. AI card; 3312. bracket; 3313. heat sink module; 3314. fastener; 34. AI module cover; 35. reinforcement frame; 36. cable guard cover;

[0069] 4. HIB and QSFP-56 module; 40. second tray; 401. rotation fulcrum; 41. second PCIe module; 411. second PCIe card; 42. CTC interconnection interface;

[0070] 5. Second NVMe module; 51. Bracket; 52. Second NVMe mechanical hard disk;

[0071] 6. First fan module; 60. Third tray; 61. First fan;

[0072] 7. Power supply module; 70. Fourth tray; 71. PSU CRPS power supply. DETAILED DESCRIPTION

[0073] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only parts related to the present invention, rather than all structures, are shown in the accompanying drawings.

[0074] In the description of the present invention, unless otherwise clearly specified and limited, the terms "connected", "connected", and "fixed" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0075] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include the first feature being in direct contact with the second feature, or may include the first feature being in contact with the second feature through another feature between them instead of being in direct contact. Moreover, a first feature being "above", "above" and "above" a second feature includes the first feature being directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes the first feature being directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0076] In the description of the present invention, the terms "upper", "lower", "left", "right" and other directions or positional relationships are based on the directions or positional relationships shown in the drawings, and are only for the convenience of description and simplified operation, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operate in a specific direction, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are only used to distinguish in the description and have no special meaning.

[0077] like Figures 1 to 24 As shown, this embodiment provides an AI server architecture system, including a chassis 1, a CPU module 2, a GPU module 3, a HIB and QSFP-56 module 4, a second NVMe module 5, a first fan module 6 and a power module 7. The chassis 1 includes a base 11, a top cover 12 detachably connected to the top of the base 11, and a partition 13 arranged between the base 11 and the top cover 12, and the partition 13 divides the interior of the chassis 1 into an upper space and a lower space. In this embodiment, the chassis 1 is a 6U industrial control chassis, which is used for a standard rack server with a 19-inch width rack; wherein the upper space of the chassis 1 is a 4U industrial control chassis for installing the GPU module 3, and the lower space of the chassis 1 is a 2U industrial control chassis for installing the CPU module 2. Here, the thickness of the chassis 1 is in U units, 1U = 44.5mm, that is, the thickness of the 4U industrial control chassis is 178mm, and the thickness of the 2U industrial control chassis is 89mm. In this embodiment, the chassis 1 is divided into an upper space and a lower space by a partition 13, and the structure is reasonable and compact. Moreover, the partition 13 can play a good supporting role for the module installed above it.

[0078] The CPU module 2 can be installed horizontally in the lower space of the chassis 1 and is located at the front of the chassis 1. It is mainly used to realize functions such as data interaction and global control. The CPU module 2 mainly includes a motherboard, a CPU connected to the motherboard signal, a memory bar, a first NVMe module 21, an IO board, a first PCIe module 22 and a SATA module 23. In this embodiment, the CPU module 2 is integrated as a whole on the first tray 20, and a slide rail extending along its depth direction is provided on the outer wall of the first tray 20. The side wall of the lower space of the chassis 1 is provided with a sliding assembly that cooperates with the slide rail, so that the first tray 20 can be pushed into the chassis 1 or pulled out from the chassis 1 along the front and rear direction of the chassis 1 to facilitate the installation and maintenance of the CPU module 2. The CPU module 2 of this embodiment includes two CPUs, and the two CPUs are connected by four sets of UPI (Ultra Path Interconnect, point-to-point processor interconnection technology), which has a higher communication rate and lower power consumption. Specifically, the CPU uses Intel EGS series CPU, can support 32 TX8 AI cards, and uses air cooling to dissipate heat. The CPU module 2 of this embodiment includes 32 memory sticks, using 2DPC DDR5 memory sticks.

[0079] The first NVMe module 21 includes four first NVMe mechanical hard disks 211, each of which is pluggable and arranged on the front window of the chassis 1 and locked by a buckle. The IO board is integrated with a UID button and indicator light, a system health indicator light, a BMC management network port, a USB interface and a VGA interface. The first PCIe module 22 includes a card tray 221 and three first PCIe cards 222 located on the front window of the chassis 1. The card tray 221 is detachably installed on the first tray 20, and the three first PCIe cards 222 are pluggable and installed in the card tray 221. In this way, the three first PCIe cards 222 can be disassembled at one time by disassembling the card tray 221, and it is also convenient for the rapid disassembly and assembly of a single first PCIe card 222, and maintenance is flexible. The SATA module 23 includes four SATA solid-state hard disks 231 located on the front window of the chassis 1. In this embodiment, the PCIe interface adopts the form of Slimline×8 to support PCIe 5.0, and is easy to flexibly configure through the form of MCIO to CEM. BMC uses AST2600 and AMI code base for system monitoring and remote management. The clock uses the PCH built-in clock generator, and the external clock buffer expands the PCIe 100M CLK. The CPLD controls the power startup timing and all reset control logic. Both BIOS and BMC use Dual Flash.

[0080] In this embodiment, a panel 26 is detachably provided on the front side of the first tray 20, and eight cable sockets 27 for connecting to the HIB module signals are provided on the front side of the mainboard, and the panel 26 is shielded on the outside of the cable sockets 27. Specifically, the panel 26 is connected to the first tray 20 by bolts, and is easy to disassemble and assemble; and the panel 26 is a perforated plate, which can protect the sockets and ensure the heat dissipation of the server, and also facilitate the technicians to observe whether there is any abnormality in the server through the front window.

[0081] The GPU module 3 is installed in the upper space of the chassis 1, and is mainly used for AI graphics data calculation. The GPU module 3 includes a UBB board 31, a fixed frame 32 and an AI module 33. The fixed frame 32 is fixed on the UBB board 31. The bottom of the fixed frame 32 is hollowed out. A plurality of plug-in positions are provided in the fixed frame 32, and each plug-in position is provided with a snap-on structure. The AI ​​module 33 includes a plurality of AI boards 331, and each AI board 331 is respectively plugged into the fixed frame 32 and fixed to the fixed frame 32 by a snap-on structure. The lower end of each AI board 331 is respectively plugged into the slot 311 on the UBB board 31 through a gold finger dual-density connector, and the signals between adjacent AI boards 331 are connected. In this embodiment, the GPU module 3 is installed in the upper space of the chassis 1, so that the AI ​​board 331 can be inserted vertically into the UBB board 31 from above; the AI ​​board 331 and the UBB board 31 are connected by a gold finger double-density connector, which realizes direct interconnection between the inside and outside of the chip, and the entire chip is a data stream, thereby saving the expensive memory and switch costs, and reducing the use of data cables; the signal connection between adjacent AI boards 331 can form an AI card cluster, which greatly improves the computing power and image processing capabilities of the server, and is conducive to improving the speed of the server in processing massive data. Optionally, 32 slots 311 are arranged in an array on the UBB board 31, and the number of AI boards 331 is 32, and each AI board 331 is vertically plugged into each slot 311 one by one. In this embodiment, the AI ​​boards 331 are interconnected at high speed through the NVLink interface. Specifically, each AI board 331 is integrated with a plurality of signal ports, the slot 311 is signal-connected to the signal ports, each slot 311 includes a first interface and a second interface, the first interface is signal-connected to the second interface of the same slot 311, and the second interface is signal-connected to the first interface of the adjacent slot 311, so that adjacent AI boards 331 are signal-connected. Exemplarily, the AI ​​board 331 of this embodiment integrates 8 LPDDR5s and is connected to adjacent AI boards 331 through 4-way × 8 25G NRE 2D mesh expansion.

[0082] Preferably, the AI ​​board 331 includes an AI card 3311 and a bracket 3312, the AI ​​card 3311 is fixed to the bracket 3312 by a fastener 3314, the bracket 3312 is vertically plugged into the fixed frame 32 and fixed to the fixed frame 32 by a snap structure, so that the AI ​​board 331 can be plugged into the UBB board 31 as a whole, and the AI ​​board 331 is easy to plug and unplug, and the configuration is flexible. The AI ​​board 331 does not need to be fixed to the fixed frame 32 by screws, and is easy to disassemble and assemble, and is reliably fixed, and can effectively avoid the impact of vibration on the system caused by the AI ​​card 3311 during transportation. Further, the AI ​​board 331 also includes a heat sink module 3313, the heat sink module 3313 is fixedly connected to the bracket 3312, and the heat sink module 3313 is attached to the AI ​​card 3311 for dissipating heat for the AI ​​card 3311. In this embodiment, the PCIe uplink interface of the AI ​​card 3311 is PCIe gen4 x8. The AI ​​card 3311 is designed to have a maximum power consumption of 300W / 450W. Considering the CPU performance, the interface between each AI card 3311 and the CPU is PCIe Gen4×16. In this embodiment, the system's artificial intelligence large model training and reasoning are mainly carried out through 32 AI cards 3311. The TX8 can reshape the AI ​​chip and organically combine multiple AI cards 3311 in a flexible expansion method to achieve excellent performance of the entire machine.

[0083] Furthermore, a handle 312 is also provided on the UBB board 31 to facilitate the operator to put the entire GPU module 3 into or take it out of the chassis 1. Furthermore, the GPU module 3 also includes an AI module cover 34, which is detachably covered on the top of the fixed frame 32. Specifically, the AI ​​module cover 34 can be a whole cover or a plurality of sub-covers. Figure 2 and Fig. 9 As shown, this embodiment is provided with eight sub-covers, each of which is provided above four corresponding AI boards 331. In addition, the GPU module 3 also includes a reinforcing frame 35, which is fixedly connected to the outside of the fixing frame 32 to increase the strength of the fixing frame 32 to improve the reliability of the system. Furthermore, a cable protection cover 36 is also provided in the chassis 1 of this embodiment to fix and protect the cables to ensure the neatness and aesthetics of the cables and improve the stability of the system.

[0084] The HIB and QSFP-56 modules 4 can be installed horizontally in the upper space of the chassis 1 in a pluggable manner, including the HIB module and the QSFP-56 module, wherein the QSFP-56 module is located at the front of the chassis 1, and the HIB module is located between the GPU module 3 and the QSFP-56 module. Specifically, the HIB and QSFP-56 modules 4 are integrated as a whole on the second tray 40, and the outer side wall of the second tray 40 is provided with a slide rail extending along its depth direction, and the side wall of the upper space of the chassis 1 is provided with a sliding assembly that cooperates with the slide rail, so that the second tray 40 can be pushed into the chassis 1 or pulled out from the chassis 1 along the front and rear direction of the chassis 1, so as to facilitate the installation and maintenance of the HIB module and the QSFP-56 module. In this embodiment, the HIB module is connected to the UBB board 31 through a BTB connector, and has high speed and high density performance. It is not only convenient and reliable to connect, but also greatly reduces the use of data cables, thereby reducing the ventilation area occupied by data cables, and improving the heat dissipation performance of the server. The HIB module is signal-connected to the eight cable sockets 27 on the front side of the mainboard via cables, and the HIB module is also signal-connected to the second PCIe module 41. The QSFP-56 module is signal-connected to the UBB board 31. The QSFP-56 module includes 24 QSFP-56 optical modules and 24 CTC interconnection interfaces 42 located on the front window of the chassis 1. In addition, 32 AI card status indicator lights and a power button and indicator light are also provided on the front of the QSFP-56 module. The second PCIe module 41 includes four second PCIe cards 411 located on the front window of the chassis 1, of which two second PCIe cards 411 are installed on the left side of the front window, and two second PCIe cards 411 are installed on the right side of the front window.

[0085] The second NVMe module 5 is arranged above the HIB and QSFP-56 module 4, and is connected to the HIB module signal. The second NVMe module 5 includes eight second NVMe mechanical hard disks 52 located on the front window of the chassis 1. The eight second NVMe mechanical hard disks 52 can be integrated as a whole on a bracket 51 to facilitate the installation and disassembly of the entire module; each second NVMe mechanical hard disk 52 is plugged into the bracket 51 through an elastic buckle, so as to facilitate the loading or extraction of each second NVMe mechanical hard disk 52, and the expansion is convenient and the configuration is flexible. This embodiment can provide a variety of flexible replaceable modules by setting the first NVMe module 21, the second NVMe module 5, the first PCIe module 22, the second PCIe module 41, the SATA module 23 and the IO board, thereby realizing the flexible expansion and hardware redundancy functions of the server.

[0086] The first fan module 6 can be installed in the upper space of the chassis 1 in a horizontal pluggable manner and is located at the rear of the chassis 1 for heat dissipation of the entire server. The first fan module 6 includes a plurality of first fans 61, which are integrated on the third tray 60. The outer wall of the third tray 60 is provided with a slide rail extending along its depth direction, and the side wall of the upper space of the chassis 1 is provided with a sliding assembly that cooperates with the slide rail, so that the third tray 60 can be pushed into the chassis 1 or pulled out from the chassis 1 along the front-back direction of the chassis 1, so as to facilitate the installation and maintenance of the first fan module 6. The first fan module 6 is connected to the HIB module signal. In this embodiment, the plurality of first fans 61 are fixed to the third tray 60 by means of snaps, which are convenient and flexible to disassemble and assemble, and support hot plugging of each fan. Specifically, the first fan module 6 includes ten 8086 fans.

[0087] The power module 7 can be installed horizontally in the lower space of the chassis 1 and is located at the rear of the chassis 1. The power module 7 is connected to the mainboard through a gold finger, and its connection is convenient and reliable; the power module 7 is electrically connected to the UBB board 31, the HIB module, the QSFP-56 module and the first fan module 6 respectively, and is used to realize the power supply of the corresponding modules. The power module 7 is integrated as a whole on the fourth tray 70, and a slide rail extending along its depth direction is provided on the outer wall of the fourth tray 70, and a sliding assembly cooperating with the slide rail is provided on the side wall of the lower space of the chassis 1, so that the fourth tray 70 can be pushed into the chassis 1 or pulled out from the chassis 1 along the front and rear direction of the chassis 1, so as to facilitate the installation and maintenance of the power module 7. In this embodiment, the power module 7 is recessed 50mm inward relative to the rear window of the chassis 1, and the length is 1050mm, thereby solving the problem of easy cable outlet when assembled into a 1.2m cabinet. Optionally, the power supply module 7 includes five PSU CRPS power supplies 71, and each PSU CRPS power supply 71 is pluggable and disposed on the fourth tray 70, so that disassembly and assembly are convenient and flexible.

[0088] As can be seen from the above, this embodiment is provided with 4 SFF modules (i.e., 4 first NVMe mechanical hard disks 211 + 8 second NVMe mechanical hard disks 52 + 4 SATA solid-state hard disks 231), 24 QSFP-56 optical modules, 1×3 first PCIe cards 222, 2×2 second PCIe cards 411, power buttons and indicator lights, etc. on the front window of the chassis 1, and the rear window is provided with a power module 7 and a first fan module 6.

[0089] Optionally, the CPU module 2 of this embodiment also includes a second fan module 24 and an air guide cover 25. The second fan module 24 is installed on the first tray 20 for dissipating heat for the CPU module 2. The air guide cover 25 is arranged on one side of the second fan module 24. The air guide cover 25 can divide the airflow of the second fan module 24 into a memory air duct and a motherboard air duct to improve the heat dissipation performance of the server and ensure the stable and high-performance operation of the server. Specifically, the second fan module 24 includes a plurality of second fans 241, each of which is fixed to the first tray 20 by a snap-fit ​​method without screw fastening, and the disassembly and assembly are flexible and convenient. Specifically, the second fan module 24 includes six 6056 fans. Further, a mounting buckle is provided on the first tray 20, and the air guide cover 25 is installed on the first tray 20 by the mounting buckle. The air guide cover 25 can be installed in place by pressing down, eliminating the complicated bolt connection, greatly improving the installation efficiency, saving assembly time, and reducing production labor costs. Furthermore, a heat sink is also provided on the first tray 20, which may specifically include a heat sink seat, a heat-conducting copper tube, heat sink fins, etc., so as to improve the heat dissipation capacity of the CPU.

[0090] Since the HIB and QSFP-56 module 4 are heavy and the HIB module needs to be plugged and connected to the UBB board 31 through the BTB connector, in order to facilitate the HIB and QSFP-56 module 4 to be installed into the chassis 1 and to be smoothly connected to the UBB board 31, the present embodiment is further provided with a power assist mechanism. The power assist mechanism includes a power assist wrench 14 and a rotation fulcrum 401. The two power-assisting wrenches 14 are respectively hinged to the two inner side walls of the upper space of the chassis 1. A card slot 141 is provided at the bottom of the power-assisting wrench 14. The power-assisting wrench 14 is in an inclined state under the action of natural gravity, and the card slot 141 is facing the direction in which the second tray 40 is pushed; rotation fulcrums 401 are respectively provided on both sides of the second tray 40. When the second tray 40 drives the HIB and the QSFP-56 module 4 to be pushed into the chassis 1, each rotation fulcrum 401 can be inserted into the card slot 141 of the corresponding power-assisting wrench 14. When the second tray 40 is continued to be pushed into the chassis 1, the rotation fulcrum 401 can make the power-assisting wrench 14 rotate to a vertical state. At this time, the card slot 141 completely hooks the rotation fulcrum 401 on the second tray 40. After that, the operator can further push the second tray 40 into the chassis 1 by leveraging the power-assisting wrench 14 to realize the plug-in connection between the HIB module and the UBB board 31. In addition, the power wrench 14 is also provided with a fixing hole, and the two inner side walls of the upper space of the chassis 1 are also provided with threaded holes. When the second tray 40 is installed in place, the power wrench 14 is fixed to the side wall of the chassis 1 by bolts, thereby locking the position of the second tray 40, ensuring the reliability of the connection between the HIB module and the UBB board 31, reducing the server failure problem caused by insufficient connection, and greatly improving the stability of the server. Furthermore, a handle is also provided on the outside of the HIB and QSFP-56 module 4 to facilitate the HIB and QSFP-56 module 4 to be pushed into and pulled out of the chassis 1.

[0091] The AI ​​server architecture system provided in this embodiment is a large-model GPU training AI server model. Through the design of 32 high-density AI TX8 (reshapeable) cards, it takes into account the expansion of NVMe mechanical hard drives and PCIe cards, and integrates Intel EGS computing modules, it can achieve leading performance and low cost and other optimization indicators; by providing a variety of flexible replaceable modules, it has balanced computing power and high-density AI computing modules, as well as flexible IO expansion and hardware redundancy.

[0092] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, various obvious changes, readjustments and substitutions can be made without departing from the protection scope of the present invention. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. AI server architecture system, characterized in that: include: A chassis, comprising a base, a top cover, and a partition disposed between the base and the top cover, wherein the partition divides the interior of the chassis into an upper space and a lower space; A CPU module is pluggably installed in the lower space of the chassis and is located at the front of the chassis. The CPU module includes a mainboard, a CPU connected to the mainboard signal, a first NVMe module, an IO board, a first PCIe module, and a SATA module; A GPU module is installed in the upper space of the chassis, the GPU module includes a UBB board, a fixing frame and an AI module, the fixing frame is fixed on the UBB board, the AI ​​module includes a plurality of AI boards, each of the AI ​​boards is respectively plugged into the fixing frame, and the lower end of each of the AI ​​boards is respectively connected to the slot on the UBB board through a gold finger double-density connector, and adjacent AI boards are signal-connected; The HIB and QSFP-56 modules are pluggably installed in the upper space of the chassis, including a HIB module and a QSFP-56 module, wherein the QSFP-56 module is located at the front of the chassis, and the HIB module is located between the GPU module and the QSFP-56 module; the HIB module is connected to the UBB board through a BTB connector, the HIB module is connected to the motherboard signal, and the HIB module is also signal-connected to a second PCIe module, and the QSFP-56 module is signal-connected to the UBB board; The second NVMe module is arranged above the HIB and QSFP-56 modules and is signal-connected to the HIB module.

2. The AI ​​server architecture system according to claim 1, characterized in that: Also includes: A first fan module is pluggably installed in the upper space of the chassis and is located at the rear of the chassis, and the first fan module is signal-connected to the HIB module; The power supply module is pluggably installed in the lower space of the chassis and is located at the rear of the chassis. The power supply module is connected to the mainboard through a gold finger. The power supply module is electrically connected to the UBB board, the HIB module, the QSFP-56 module and the first fan module respectively.

3. The AI ​​server architecture system according to claim 1, characterized in that: The UBB card array is provided with 32 slots, the number of the AI ​​cards is 32, and each AI card is plugged into each slot one by one; The AI ​​board is integrated with a plurality of signal ports, the slots are signal-connected to the signal ports, each slot comprises a first interface and a second interface, the first interface is signal-connected to the second interface of the same slot, and the second interface is signal-connected to the first interface of an adjacent slot, so that adjacent AI boards are signal-connected.

4. The AI ​​server architecture system according to claim 1, characterized in that: The AI ​​board includes an AI card and a bracket, the AI ​​card is fixed on the bracket, and the bracket is vertically inserted into the fixing frame and connected with the fixing frame by buckles.

5. The AI ​​server architecture system according to claim 2, characterized in that: The CPU module is integrated on the first tray, the HIB and QSFP-56 module are integrated on the second tray, the first fan module is integrated on the third tray, and the power supply module is integrated on the fourth tray. The first tray, the second tray, the third tray and the fourth tray are all pluggably installed in the chassis through a slide rail structure.

6. The AI ​​server architecture system according to claim 5, characterized in that: The inner wall of the upper space of the chassis is rotatably connected with a power-assist wrench, and a card slot is provided at the bottom of the power-assist wrench. The power-assist wrench is in an inclined state under the action of gravity and the card slot is facing the direction in which the second tray is pushed; rotation fulcrums are provided on both sides of the second tray. When the second tray, the HIB and the QSFP-56 module are pushed into the chassis, the rotation fulcrum can be inserted into the card slot and the power-assist wrench can be rotated, and then the second tray can be pushed into the chassis by pulling the power-assist wrench.

7. The AI ​​server architecture system according to claim 5, characterized in that: The CPU module further includes a second fan module and an air guide cover, wherein the second fan module is mounted on the first tray, the air guide cover is arranged on one side of the second fan module, and the air guide cover is used to divide the air flow of the second fan module into a memory air duct and a motherboard air duct; A panel is detachably provided on the front side of the first tray, and eight cable sockets for connecting signals with the HIB module are provided on the front side of the mainboard. The panel is shielded on the outside of the cable sockets.

8. The AI ​​server architecture system according to claim 1, characterized in that: The UBB board is provided with a handle for taking and placing the GPU module; and / or, The GPU module further includes an AI module cover, and the AI ​​module cover is detachably covered on the top of the fixing frame; and / or, The GPU module further includes a reinforcement frame, and the reinforcement frame is fixedly connected to the fixing frame; and / or, A cable protection cover is also provided in the chassis to protect the cables.

9. The AI ​​server architecture system according to claim 5, characterized in that: The first NVMe module includes four first NVMe mechanical hard disks located on the front window of the chassis; The second NVMe module includes 8 second NVMe mechanical hard disks located on the front window of the chassis; The SATA module includes 4 SATA solid state hard disks located on the front window of the chassis; The first PCIe module includes a card tray and three first PCIe cards located on the front window of the chassis, the card tray is detachably mounted on the first tray, and the three first PCIe cards are pluggably mounted in the card tray; The second PCIe module includes four second PCIe cards located on the front window of the chassis; The QSFP-56 module includes 24 QSFP-56 optical modules and 24 CTC interconnection interfaces located on the front window of the chassis.

10. The AI ​​server architecture system according to claim 7, characterized in that: The first fan module includes a plurality of first fans, and each of the first fans is fixed to the third tray by a buckle; The second fan module includes a plurality of second fans, and each of the second fans is fixed to the first tray by a buckle; The power module is recessed inward relative to the rear window of the chassis, and the power module includes five PSU CRPS power supplies, which are pluggably arranged on the fourth tray; The CPU module has 2 CPUs, 32 memory bars and several heat sinks; The front of the QSFP-56 module is also provided with an AI card status indicator light and a power button and indicator light. The IO board is integrated with a UID button and indicator light, a system health indicator light, a BMC management network port, a USB interface and a VGA interface.

Images