Cable backboard topological structure and artificial intelligence rack-type whole cabinet

Abstract

The invention provides a cable backboard topological structure and an artificial intelligence rack-mounted whole cabinet. The cable backboard topological structure comprises at least one graphic processor node and at least one data exchange node, the number of the graphics processor nodes is the same as that of the data ports of each data exchange node, and the data ports of the same serial number of different graphics processor nodes are connected with the data ports of different serial numbers of the same data exchange node in a one-to-one correspondence manner; according to the scheme, when the data exchange nodes need to be expanded, only the data exchange nodes need to be added, then all the data ports of the added data exchange nodes are fully connected with all the graphic processing nodes, and the expandability and compatibility of the topological structure of the Cable Tray are improved.

Description

Technical Field

[0001] The present disclosure relates to the field of cable topology backplanes for whole cabinets, and in particular to a cable backplane topology structure and an artificial intelligence rack-type whole cabinet. Background Art

[0002] Thanks to the development of artificial intelligence (AI) technology, various enterprises and scientific research institutions have an increasing demand for computing power, and AI Rack cabinets with powerful computing power have emerged. AI Rack cabinets mainly use graphics processing unit (GPU) nodes to provide computing services, and are equipped with central processing unit (CPU) nodes for control and data exchange (Switch) nodes for data exchange between nodes.

[0003] In related technologies, such as Figure 1 As shown in the figure, the GPU nodes, data exchange nodes and CPU nodes in the AI ​​Rack cabinet are all connected to the backplane connector through connectors to achieve signal interconnection. Among them, the signal interconnection between the GPU node and the data exchange node is also achieved through the following Figure 1 The cable tray topology shown in the figure is implemented. Figure 1 As shown, Figure 1 In the figure, an AI Rack cabinet including 8 graphics processor nodes GPU1 to GPU8 and 6 data exchange nodes Switch1 to Switch6 is taken as an example. Among the 8 graphics processor nodes, the data port of any serial number of each graphics processor node is connected to the same data port corresponding to all the data exchange nodes; for example, the data port 0 of the first graphics processor node GPU1 is connected to the data port 0 of the 6 data exchange nodes, the data port 1 of the first graphics processor node GPU1 is connected to the data port 1 of the 6 data exchange nodes, and the data port 7 of the first graphics processor node GPU1 is connected to the data port 7 of the 6 data exchange nodes. Figure 1 As shown, Figure 1Also shown are two reserved data exchange nodes Switch7 and Switch8 for expansion. When the data exchange nodes Switch7 and Switch8 need to be expanded, each graphics processor node in GPU1 to GPU8 needs to be connected, such as GPU1's 0 data port with Switch7 and Switch8's 0 data port, GPU1's 1 data port with Switch7 and Switch8's 1 data port, GPU2's 0 data port with Switch7 and Switch8's 0 data port, GPU2's 1 data port with Switch7 and Switch8's 1 data port, etc. It is foreseeable that the Cable Tray applicable to 6 data exchange nodes cannot be reused for 8 data exchange nodes after adding the data exchange nodes Switch7 and Switch8.

[0004] Therefore, the topological structure scalability and compatibility of the Cable Tray in the AI ​​Rack cabinet in the related art are poor. Therefore, the topological structure scalability and compatibility of the Cable Tray in the AI ​​Rack cabinet in the related art are poor. Summary of the invention

[0005] The present disclosure provides a cable backplane topology structure and an artificial intelligence rack-type cabinet, which can solve the problems of poor scalability and compatibility of the topology structure of the Cable Tray in the AI ​​Rack cabinet in the related technology, and improve the scalability and compatibility of the topology structure of the Cable Tray.

[0006] In order to solve the above technical problems, the present disclosure provides a cable backplane topology structure, including:

[0007] at least one graphics processor node and at least one data exchange node;

[0008] The number of the graphics processor nodes is the same as the number of data ports of each data exchange node, and the data ports with the same serial number of different graphics processor nodes are connected to the data ports with different serial numbers of the same data exchange node in a one-to-one correspondence.

[0009] Optionally, data ports with different serial numbers of the same graphics processor node are connected to data ports with the same serial numbers of different data exchange nodes in a one-to-one correspondence via the same cable bridge.

[0010] Optionally, the cable backplane topology structure further includes the same number of first connectors as the number of the graphic connector nodes and the same number of second connectors as the number of the data exchange nodes;

[0011] The graphics processor node is plugged into the corresponding first connector, and the data exchange node is plugged into the corresponding second connector; the same first connector is connected to different second connectors through the same cable bridge.

[0012] Optionally, the cable backplane topology structure also includes a first number of reserved first connectors and a second number of reserved second connectors, the reserved first connectors are used to plug in the added graphics processor nodes when the graphics processor nodes are expanded, and the reserved second connectors are used to plug in the added data exchange nodes when the data exchange nodes are expanded.

[0013] The present disclosure also provides an artificial intelligence rack-type whole cabinet, comprising:

[0014] A cabinet and a backplane connector installed in the cabinet;

[0015] At least one central processing unit node, at least one graphics processing unit node and at least one data exchange node for signal interconnection are plugged into the backplane connector; wherein the at least one graphics processing unit node and the at least one data exchange node are signal interconnected through the cable backplane topology structure provided in the present disclosure.

[0016] Optionally, two cable backplanes are further included; the two cable backplanes are symmetrically distributed on both sides of the cabinet, and the number of the central processing unit nodes, the graphics processing unit nodes and the data exchange nodes connected to each cable backplane is the same.

[0017] Optionally, the artificial intelligence rack-type cabinet further includes a pluggable power supply module installed in the cabinet;

[0018] The backplane connector is provided with a power connector, and the power module is plugged and matched with the power connector.

[0019] Optionally, the power module includes a first power module and a second power module;

[0020] The first power module, the central processing unit node, the graphics processing unit node, the data exchange node and the second power module are distributed in the cabinet from top to bottom.

[0021] Optionally, the artificial intelligence rack-type cabinet also includes a busbar arranged in the center of the cabinet, and the power module provides power to the central processing unit node, the graphics processing unit node and the data exchange node through the busbar.

[0022] Optionally, the artificial intelligence rack-type cabinet also includes a heat dissipation module, and the heat dissipation module includes a liquid cooling module and an air cooling module.

[0023] The present disclosure provides a cable backplane topology structure, comprising: at least one graphics processor node and at least one data exchange node; the number of the graphics processor nodes is the same as the number of data ports of each of the data exchange nodes, and the data ports with the same serial number of different graphics processor nodes are connected one-to-one with the data ports with different serial numbers of the same data exchange node.

[0024] According to the scheme disclosed in the present invention, since the number of the graphics processor nodes is the same as the number of data ports of each of the data exchange nodes, the data ports with the same serial number of different graphics processor nodes are connected one-to-one with the data ports with different serial numbers of the same data exchange node. Therefore, each data exchange node can be connected to all the graphics processing nodes through its data port; when the data exchange nodes need to be expanded, it is only necessary to add data exchange nodes and then connect all the data ports of the added data exchange nodes to all the graphics processing nodes, thereby improving the scalability and compatibility of the topological structure of the Cable Tray.

[0025] Furthermore, since the lengths of cables connecting data ports with different serial numbers of the same data exchange node and different graphics processor nodes are slightly different, the links between different data exchange nodes have better consistency and higher stability.

[0026] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0028] Figure 1 It is a schematic diagram of the topological structure of a cable backplane in the related art;

[0029] Figure 2 A schematic diagram of the topological structure of the cable backplane provided in an embodiment of the present disclosure;

[0030] Figure 3 A front view of the cabinet structure of the artificial intelligence rack-type whole cabinet provided in an embodiment of the present disclosure;

[0031] Figure 4 This is a rear view of the cabinet structure of the artificial intelligence rack-type whole cabinet provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0032] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0033] Thanks to the development of AI technology, various enterprises and scientific research institutions have an increasing demand for computing power, and AI Rack cabinets with powerful computing power have emerged. AI Rack cabinets mainly use GPU nodes to provide computing services, while carrying CPU nodes for control and Switch nodes for data exchange between nodes.

[0034] The topology of the Cable Tray is extremely important in the design of the AI ​​Rack cabinet. Different Cable Tray topologies will affect the performance of the entire cabinet, system reliability, and communication costs. Therefore, the topology design of the Cable Tray of the AI ​​Rack cabinet is an important part of the design of the AI ​​Rack cabinet. It determines the connection method and communication path of each component inside the cabinet, and has a significant impact on the performance, reliability, and future scalability of the entire cabinet.

[0035] In the related art, the commonly used Cable Tray topologies include star topology, tree topology and hybrid topology; each GPU node in the star topology is connected to the data exchange node through an independent cable tray chain to achieve good fault isolation; the tree topology includes multiple sub-star topologies to achieve hierarchical management; the hybrid topology combines the advantages of multiple topologies to meet specific system requirements. For example, the star topology can be used for the connection between central processing nodes or GPU nodes, while the mesh topology can be used for redundant connections between key nodes to improve the reliability and fault tolerance of the system.

[0036] in Figure 1 The topology of the cable backplane in is a hybrid topology. Figure 1 As shown in the figure, the GPU nodes, data exchange nodes and CPU nodes in the AI ​​Rack cabinet are all connected to the backplane connector through connectors to achieve signal interconnection. Among them, the signal interconnection between the GPU node and the data exchange node is also achieved through the following Figure 1 The cable tray topology shown in the figure is implemented. Figure 1 As shown, Figure 1In the figure, an AI Rack cabinet including 8 graphics processor nodes GPU1 to GPU8 and 6 data exchange nodes Switch1 to Switch6 is taken as an example. Among the 8 graphics processor nodes, the data port of any serial number of each graphics processor node is connected to the same data port corresponding to all the data exchange nodes; for example, the data port 0 of the first graphics processor node GPU1 is connected to the data port 0 of the 6 data exchange nodes, the data port 1 of the first graphics processor node GPU1 is connected to the data port 1 of the 6 data exchange nodes, and the data port 7 of the first graphics processor node GPU1 is connected to the data port 7 of the 6 data exchange nodes. Figure 1 As shown, Figure 1 Also shown are two reserved data exchange nodes Switch7 and Switch8 for expansion. When the data exchange nodes Switch7 and Switch8 need to be expanded, each graphics processor node in GPU1 to GPU8 needs to be connected, such as GPU1's 0 data port with Switch7 and Switch8's 0 data port, GPU1's 1 data port with Switch7 and Switch8's 1 data port, GPU2's 0 data port with Switch7 and Switch8's 0 data port, GPU2's 1 data port with Switch7 and Switch8's 1 data port, etc. It is foreseeable that the Cable Tray applicable to 6 data exchange nodes cannot be reused for 8 data exchange nodes after adding the data exchange nodes Switch7 and Switch8.

[0037] Therefore, the topological structure scalability and compatibility of the Cable Tray in the AI ​​Rack cabinet in the related art are poor.

[0038] In order to solve the problem of poor scalability and compatibility of the topological structure of the Cable Tray in the related art, the present disclosure provides a cable backplane topological structure. Figure 2 As shown, the cable backplane topology structure provided by the embodiment of the present disclosure includes:

[0039] At least one graphics processor node and at least one data exchange node; the number of the graphics processor nodes is the same as the number of data ports of each data exchange node, and the data ports with the same serial number of different graphics processor nodes are connected to the data ports with different serial numbers of the same data exchange node in a one-to-one correspondence.

[0040] In one embodiment, the number of graphics processor nodes may be 4, 8, 16, etc., which is not limited in the present disclosure. Similarly, the number of data exchange nodes may also be 4, 8, 16, etc., which is not limited in the present disclosure. Figure 2 As shown, in the embodiment of the present disclosure, eight graphics processor nodes GPU1 to GPU8, six data switching nodes SWITCH1 to SWITCH6, and two extended data switching nodes SWITCH7 and SWITCH8 are taken as an example for description.

[0041] like Figure 2 As shown, each data switching node has 8 data ports from 0 to 7, for example, SWITCH1 has 8 data ports A0, A1, ..., A7, for example, SWITCH2 has 8 data ports B0, B1, ..., B7, for example, SWITCH3 has 8 data ports CO, C1, ..., C7, for example, SWITCH4 has 8 data ports D0, D1, ..., D7, for example, SWITCH5 has 8 data ports E0, E1, ..., E7, for example, SWITCH6 has 8 data ports F0, F1, ..., F7, for example. The graphics processor node has 8 data ports from A to H, for example, GPU1 has 8 data ports from A0 to HO, GPU2 has 8 data ports from A1 to H1, GPU3 has 8 data ports from A2 to H2, GPU4 has 8 data ports from A3 to H3, GPU5 has 8 data ports from A4 to H4, GPU6 has 8 data ports from A5 to H5, GPU7 has 8 data ports from A6 to H6, and GPU8 has 8 data ports from A7 to H7. The number of graphics processor nodes is the same as the number of data ports of the data exchange node, that is, Figure 2 As shown, there are 8 graphics processor nodes GPU1 to GPU1 in total, and each data exchange node also has 8 data ports.

[0042] The data ports with the same serial number of different graphics processors, such as the data ports A1 to A7 of the A serial number of GPU1 to GPU8, are respectively connected to the data ports A0, A1, ..., A7 of the data switching node SWITCH1; it is foreseeable that the data ports B1 to B7 of the B serial number of GPU1 to GPU8 can be respectively connected to the data ports B0, B1, ..., B7 of the data switching node SWITCH2, ..., the data ports F1 to F7 of the F serial number of GPU1 to GPU8 can be respectively connected to the data ports B0, B1, ..., B7 of the data switching node SWITCH2, ..., the data ports F1 to F7 of the F serial number of GPU1 to GPU8 can be respectively connected to the data switching node The F0, F1, ..., F7 data ports of SWITCH6 are connected respectively; when the data switching nodes SWITCH7 and SWITCH8 need to be expanded, it is only necessary to connect the data ports G1 to G7 of the G sequence number of GPU1 to GPU8 with the G0, G1, ..., G7 data ports of the data switching node SWITCH7 respectively, and connect the data ports F1 to F7 of the F sequence number of GPU1 to GPU8 with the F0, F1, ..., F7 data ports of the data switching node SWITCH8F respectively.

[0043] Therefore, according to the solution disclosed in the present invention, since the number of the graphics processor nodes is the same as the number of data ports of each of the data exchange nodes, the data ports with the same serial number of different graphics processor nodes are connected one-to-one with the data ports with different serial numbers of the same data exchange node. Therefore, each data exchange node can be connected to all graphics processing nodes through its data port; when the data exchange nodes need to be expanded, it is only necessary to add data exchange nodes, and then all data ports of the added data exchange nodes can be connected to all graphics processing nodes, thereby improving the scalability and compatibility of the topological structure of CableTray.

[0044] Furthermore, since the lengths of cables connecting data ports with different serial numbers of the same data exchange node and different graphics processor nodes are slightly different, the links between different data exchange nodes have better consistency and higher stability.

[0045] In one embodiment, in order to meet the high-speed data transmission requirements between the graphics processor node and the data exchange node, the data ports with different serial numbers of the same graphics processor node are connected to the data ports with the same serial numbers of different data exchange nodes in a one-to-one correspondence, such as Figure 2As shown, the A0 data port of the graphics processor node GPU1 is connected to the serial number data port 0 of the data switching node SWITCH1, that is, the A0 data port of the graphics processor node GPU1 is connected to the serial number data port A0 of the data switching node SWITCH1, and the B0 data port of the graphics processor node GPU1 is connected to the serial number data port 0 of the data switching node SWITCH2, that is, the B0 data port of the graphics processor node GPU1 is connected to the serial number data port B0 of the data switching node SWITCH2.

[0046] In one embodiment, data ports with different serial numbers of the same graphics processor node are connected to data ports with the same serial numbers of different data exchange nodes in a one-to-one correspondence via the same cable tray.

[0047] like Figure 2 As shown, taking the graphics processor node GPU1 as an example, the connection cables between the A0 to FO data ports of the graphics processor node GPU1 and the serial number 0 data ports of the data switching nodes SWITCH1 to SWITCH6 use the same cable tray, thereby effectively improving the fault isolation performance of the network. When a fault occurs in the graphics processor node GPU1 or its corresponding cable tray, it will only affect the connection between the graphics processor node GPU1 and the data processing node, and will not affect the operation of the entire network, thereby simplifying the diagnosis and repair process of the network fault; at the same time, since each graphics processor node has an independent connection path, it helps to reduce network congestion and data conflicts, thereby improving the speed and stability of data transmission.

[0048] Furthermore, since data ports with different serial numbers of the same graphics processor node are connected one-to-one with data ports with the same serial numbers of different data exchange nodes through the same cable tray, when it is necessary to reduce graphics processor nodes, it is only necessary to remove the corresponding cable tray, thereby achieving backward compatibility.

[0049] In one embodiment, the cable backplane topology structure further includes the same number of first connectors as the number of the graphic connector nodes and the same number of second connectors as the number of the data exchange nodes;

[0050] The graphics processor node is plugged into the corresponding first connector, and the data exchange node is plugged into the corresponding second connector; the same first connector is connected to different second connectors through the same cable bridge.

[0051] In one embodiment, the provision of the first connector and the second connector makes the installation of the graphics processor node and the data exchange node faster; if the graphics processor node or the data exchange node needs to be expanded, it is only necessary to add the corresponding first connector or the second connector, and then plug in the graphics processor node and the data exchange node.

[0052] In one embodiment, the cable backplane topology structure also includes a first number of reserved first connectors and a second number of reserved second connectors, the reserved first connectors are used to plug in the added graphics processor nodes when expanding the graphics processor nodes, and the reserved second connectors are used to plug in the added data exchange nodes when expanding the data exchange nodes.

[0053] In one embodiment, a maximum of 16 graphics processor nodes and 8 data exchange nodes are usually set in the AI ​​Rack cabinet. Therefore, if the current AI Rack cabinet is provided with 8 graphics processor nodes and 6 data exchange nodes, the first number of reserved first connectors is determined to be 8, and the first number of reserved second connectors is determined to be 2. The setting of the reserved first connector and the reserved second connector allows the graphics processor node or the data exchange node to be plugged into the corresponding reserved first connector or reserved second connector when the graphics processor node or the data exchange node needs to be expanded, thereby improving the convenience of expansion.

[0054] In one embodiment, after the specifications of the first connector and the second connector are selected, PINMAP design is also required. The requirement of PINMAP design is to maximize the configuration of all nodes of all connectors to be in place.

[0055] The disclosed embodiment also provides an artificial intelligence rack-type whole cabinet, such as Figure 3 As shown, the AI ​​Rack cabinet includes a cabinet body 101 and a backplane connector installed in the cabinet body 101;

[0056] At least one central processing unit node, at least one graphics processing unit node and at least one data exchange node 104 are plugged into the backplane connector for signal interconnection; wherein, the at least one graphics processing unit node and the at least one data exchange node 104 are signal interconnected through a cable backplane topology structure as provided in an embodiment of the present disclosure.

[0057] like Figure 3 As shown, Figure 3The size of the Host node area 102 is 4U, and the size of the business Box node area 103 is 4U. The Host node area 102 can be used to accommodate a central processing unit node, and the business Box node area 103 can be used to accommodate a graphics processing unit node. Since a central processing unit node can be interconnected with multiple graphics processing unit nodes through a peripheral component interconnection PCIE assembly part, the Host node area 102 and the business Box node area 103 are set in the middle.

[0058] Since the AI ​​Rack cabinet provided by the embodiment of the present disclosure adopts the cable backplane topology structure provided by the embodiment of the present disclosure to realize the signal interconnection between the graphics processor node and the data exchange node 104, if it is necessary to expand the data exchange node 104 in the AI ​​Rack cabinet, it is only necessary to add the data exchange node 104 and then connect all the data ports of the added data exchange node 104 to all the graphics processing nodes, thereby improving the scalability and compatibility of the AI ​​Rack cabinet.

[0059] In one embodiment, if Figure 4 As shown, the AI ​​Rack cabinet provided in the embodiment of the present disclosure further includes two cable backplanes; the two cable backplanes are symmetrically distributed on both sides of the cabinet body 101, and the number of the central processing unit nodes, the graphics processing unit nodes, and the data exchange nodes 104 connected to each cable backplane is the same. Specifically, taking the AI ​​Rack cabinet containing 8 graphics processing unit nodes as an example, each cable backplane is connected to 4 graphics processing unit nodes.

[0060] The two cable trays 106 are provided to realize a redundant design. When one cable tray 106 fails or needs maintenance, the other cable tray 106 can continue to work, thereby ensuring the stable operation of the entire network. Furthermore, the two cable trays 106 are symmetrically distributed on both sides of the cabinet 101, which helps to optimize the wiring, reduce the crossing and interference of cables, and improve the neatness and maintainability of the wiring.

[0061] In one embodiment, the artificial intelligence rack-type cabinet further includes a pluggable power supply module installed in the cabinet 101;

[0062] The backplane connector is provided with a power connector, and the power module is plugged and matched with the power connector.

[0063] In one embodiment, in order to ensure power supply, a power module is added in this embodiment. The power module is installed in the AI ​​Rack cabinet and forms a pluggable connection with the AI ​​Rack cabinet, that is, it can realize the insertion and installation operation and the removal and removal operation by reciprocating linear movement in the AI ​​Rack cabinet, and can conveniently perform disassembly and assembly operations on the AI ​​Rack cabinet. The power module is mainly used to realize power supply to other AI Rack cabinet components, such as powering the central processing unit node, graphics processing unit node and data exchange node of the AI ​​Rack cabinet.

[0064] In one embodiment, the power module includes a first power module and a second power module;

[0065] In one embodiment, if Figure 3 As shown, the power module is installed in the AI ​​Rack cabinet in the form of a power plug-in box Power Shelf105, and the number of Power Shelf105 is 2; the AI ​​Rack cabinet is connected to different AI Rack cabinet components through two Power Shelf105s, ensuring that when one Power Shelf105 input fails, the other Power Shelf105 input can still continue to supply power, thereby reducing the risk of power outages caused by a single Power Shelf105 failure, thereby ensuring the continuous operation of the AI ​​Rack cabinet. Furthermore, the AI ​​Rack cabinet needs to have a power distribution function, that is, it can effectively distribute the power provided by the two Power Shelf105s to each device, while ensuring the load balance between the two Power Shelf105s. This helps to avoid overloading of a single Power Shelf105 and improve the reliability of the entire AI Rack cabinet power system.

[0066] The first power module, the central processing unit node, the graphics processing unit node, the data exchange node 104 and the second power module are distributed from top to bottom in the cabinet 101 .

[0067] The first power module and the second power module are respectively arranged at the upper and lower parts of the cabinet 101, which facilitates the wiring between the power module and the central processing unit node, the graphics processing unit node and the data exchange node 104, and improves the rationality of the wiring. At the same time, the graphics processing unit node can be evenly distributed into two parts, and the two parts of the graphics processing unit nodes can be respectively arranged at the upper and lower parts of the central processing unit node, and the central processing unit node and the graphics processing unit node are arranged.

[0068] In one embodiment, the AI ​​Rack cabinet further comprises a busbar disposed in the center of the cabinet 101, and the power module provides power to the CPU node, the GPU node and the data exchange node 104 through the busbar. The cross-sectional area and length of the busbar 107 can be determined based on the maximum power consumption of the AI ​​Rack cabinet. Figure 4 As shown, two cable trays 106 are respectively arranged on both sides of the busbar 107.

[0069] In one embodiment, the artificial intelligence rack-type cabinet also includes a heat dissipation module, and the heat dissipation module includes a liquid cooling module and an air cooling module.

[0070] In one embodiment, the heat dissipation module is installed in the AI ​​Rack cabinet and forms a pluggable connection with the AI ​​Rack cabinet, that is, it can realize the insertion and installation operation and the removal and removal operation by reciprocating linear movement in the AI ​​Rack cabinet, and can conveniently perform disassembly and assembly operations on the AI ​​Rack cabinet. The heat dissipation module is mainly used to realize air cooling of the remaining AI Rack cabinet components. Of course, the heat dissipation module can also dissipate the heat of the remaining server components by liquid cooling. For example, the AI ​​Rack cabinet can adopt a heat dissipation method that mainly uses liquid cooling and supplemented by air cooling.

[0071] In the description of the present invention, it needs to be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0072] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, "multiple" means at least two, such as two, three, etc., unless otherwise clearly defined. In the present invention, unless otherwise clearly specified and defined, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, or a detachable connection, or integrated; it can be a mechanical connection, or an electrical connection or can communicate with each other; it can be directly connected, or indirectly connected through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements, unless otherwise clearly defined. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0073] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. Moreover, a first feature being "above", "above" and "above" a second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0074] In the present invention, the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" and the like mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the described specific features, structures, materials or characteristics may be combined in a suitable manner in any one or more embodiments or examples. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory. Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be construed as limitations on the present invention, and those of ordinary skill in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A cable backplane topology structure, characterized in that: include: at least one graphics processor node and at least one data exchange node; The number of the graphics processor nodes is the same as the number of data ports of each data exchange node, and the data ports with the same serial number of different graphics processor nodes are connected to the data ports with different serial numbers of the same data exchange node in a one-to-one correspondence.

2. The cable backplane topology structure according to claim 1, characterized in that: Data ports with different serial numbers of the same graphics processor node are connected to data ports with the same serial numbers of different data exchange nodes in a one-to-one correspondence via the same cable bridge.

3. The cable backplane topology structure according to claim 1, characterized in that: The cable backplane topology structure also includes the same number of first connectors as the number of the graphic connector nodes and the same number of second connectors as the number of the data exchange nodes; The graphics processor node is plugged into the corresponding first connector, and the data exchange node is plugged into the corresponding second connector; the same first connector is connected to different second connectors through the same cable bridge.

4. The cable backplane topology structure according to claim 1, characterized in that: The cable backplane topology structure also includes a first number of reserved first connectors and a second number of reserved second connectors, wherein the reserved first connectors are used to plug in the added graphics processor nodes when the graphics processor nodes are expanded, and the reserved second connectors are used to plug in the added data exchange nodes when the data exchange nodes are expanded.

5. An artificial intelligence rack-mounted cabinet, characterized in that: include: A cabinet and a backplane connector installed in the cabinet; At least one central processing unit node, at least one graphics processing unit node and at least one data exchange node for signal interconnection are plugged into the backplane connector; wherein, the at least one graphics processing unit node and the at least one data exchange node are signal interconnected through the cable backplane topology structure described in any one of claims 1 to 4.

6. The artificial intelligence rack-mounted whole cabinet according to claim 5, characterized in that: It also includes two cable backplanes; the two cable backplanes are symmetrically distributed on both sides of the cabinet, and the number of the central processing unit nodes, the graphics processing unit nodes and the data exchange nodes connected to each cable backplane is the same.

7. The artificial intelligence rack-mounted whole cabinet according to claim 5, characterized in that: The artificial intelligence rack-type cabinet also includes a pluggable power supply module installed in the cabinet.

8. The artificial intelligence rack-mounted whole cabinet according to claim 5, characterized in that: The power module includes a first power module and a second power module; The first power module, the central processing unit node, the graphics processing unit node, the data exchange node and the second power module are distributed in the cabinet from top to bottom.

9. The artificial intelligence rack-mounted whole cabinet according to claim 8, characterized in that: The artificial intelligence rack-type cabinet also includes a busbar arranged in the center of the cabinet, and the power module provides power to the central processing unit node, the graphics processing unit node and the data exchange node through the busbar.

10. The artificial intelligence rack-mounted whole cabinet according to claim 5, characterized in that: The artificial intelligence rack-type whole cabinet also includes a heat dissipation module, and the heat dissipation module includes a liquid cooling heat dissipation module and an air cooling heat dissipation module.

Images