AI computing card and computing cluster

Abstract

The present disclosure provides an AI computing card and a computing cluster, and relates to the field of computers, and the implementation scheme is as follows: the AI computing card comprises a first circuit board, a second circuit board, at least one cross-board connector, and a heat dissipation module; the first circuit board is provided with a first AI chip set, a first interface, and a second interface, the first AI chip set comprises at least one AI chip, and the first interface is configured to communicate with and / or supply power to an external device; the second circuit board is arranged opposite to and spaced apart from the first circuit board along the thickness direction, the second circuit board is provided with a second AI chip set and a third interface, and the second AI chip set comprises at least one AI chip; the cross-board connector is coupled and connected to the second interface and the third interface, respectively; and the heat dissipation module is clamped and arranged between the first circuit board and the second circuit board, and is simultaneously thermally coupled to the first AI chip set and the second AI chip set. The present disclosure can increase the integrated number of AI chips and improve the unit volume computing power density of the computing card.

Description

Technical Field

[0001] This disclosure relates to the field of computers, and in particular to an AI (Artificial Intelligence) computing card and computing cluster. Background Technology

[0002] Currently, AI chips are developing towards higher computing power and higher power consumption, and data centers urgently need to maximize the computing power of a single rack. AI accelerator cards mainly adopt a single-layer PCB (Printed Circuit Board) structure. Due to the space limitations of standard card slots, it is difficult to integrate multiple high-power AI chips and supporting DDR storage modules, resulting in limited computing power of a single card, which cannot meet the ever-increasing computing power demand. Summary of the Invention

[0003] This disclosure provides an AI computing card and computing cluster that can increase the number of AI chips integrated and improve the computing power density per unit volume of the computing card.

[0004] According to one aspect of this disclosure, an AI computing card is provided, comprising a first circuit board, a second circuit board, at least one cross-board connector, and a heat dissipation module; the first circuit board is provided with a first AI chipset, a first interface, and a second interface, the first AI chipset including at least one AI chip, and the first interface configured for communication and / or power coupling with an external device; the second circuit board is arranged opposite to and spaced apart from the first circuit board along the thickness direction, the second circuit board is provided with a second AI chipset and a third interface, the second AI chipset including at least one AI chip; the cross-board connector is respectively mated and coupled to the second interface and the third interface, the cross-board connector is provided with first type pins and second type pins, the first type pins are configured to realize interconnection communication between the first circuit board and the second circuit board, and the second type pins are configured to realize power transmission between the first circuit board and the second circuit board; the heat dissipation module is clamped between the first circuit board and the second circuit board, and is thermally coupled to both the first AI chipset and the second AI chipset.

[0005] In some embodiments, the heat dissipation module includes a first heat dissipation surface facing a first circuit board and a second heat dissipation surface facing a second circuit board; the first heat dissipation surface is attached to a first AI chipset, and the projection of the first AI chipset along the thickness direction of the circuit board falls entirely within the range of the first heat dissipation surface; the second heat dissipation surface is attached to a second AI chipset, and the projection of the second AI chipset along the thickness direction of the circuit board falls entirely within the range of the second heat dissipation surface.

[0006] In some embodiments, at least one of the first heat dissipation surface and the second heat dissipation surface is provided with a recess, and at least a portion of at least one of the first AI chipset and the second AI chipset is embedded in the corresponding recess.

[0007] In some embodiments, the heat dissipation module is a liquid-cooled radiator or an air-cooled radiator.

[0008] In some embodiments, the second interface is located on the side of the first circuit board, and the third interface is located on the side of the second circuit board; the cross-board connector is plugged into and unplugged with the second and third interfaces in a direction perpendicular to the thickness of the circuit board.

[0009] In some embodiments, at least one of the second interface and the third interface is a recessed connector, which is at least partially embedded in the interior of the corresponding circuit board along the thickness direction of the circuit board.

[0010] In some embodiments, the heat dissipation module further includes a side surface connecting the first heat dissipation surface and the second heat dissipation surface, and the side surface is provided with a support structure for providing support to the cross-board connector.

[0011] In some embodiments, the first interface is a PCIe interface or an OAM interface.

[0012] In some embodiments, the first circuit board is further provided with a first power supply interface, and the second circuit board is further provided with a second power supply interface, wherein the second power supply interface and the first power supply interface are power supply interfaces that are redundant backups of each other.

[0013] According to another aspect of this disclosure, a computing cluster is provided, which includes multiple AI computing cards that are interconnected and coupled to each other.

[0014] According to one or more embodiments of this disclosure, the number of chips in a single card can be increased, thereby improving the computing power of a single computing card.

[0015] These and other aspects of this disclosure will be apparent from the embodiments described below, and will be elucidated with reference to the embodiments described below. Attached Figure Description

[0016] The accompanying drawings exemplify embodiments and form part of the specification, serving together with the textual description to explain exemplary implementations of the embodiments. The illustrated embodiments are for illustrative purposes only and do not limit the scope of this disclosure. Throughout the drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0017] Figure 1 This is a front view of an AI computing card according to some embodiments of this disclosure;

[0018] Figure 2 This is a top view of an AI computing card according to some embodiments of this disclosure;

[0019] Figure 3 This is an exploded structural diagram of an AI computing card according to some embodiments of this disclosure;

[0020] Figure 4 for Figure 2 A cross-sectional view along the AA direction;

[0021] Figure 5 For other embodiments of this disclosure Figure 2 Cross-sectional view along the AA direction;

[0022] Figure 6 for Figure 2 A cross-sectional view along the BB direction;

[0023] Figure 7 For other embodiments of this disclosure Figure 2 Cross-sectional view along the BB direction;

[0024] Figure 8 This is an exploded structural diagram of an AI computing card according to other embodiments of this disclosure;

[0025] Figure 9 This is a partial structural diagram of a computing cluster according to some embodiments of this disclosure. Detailed Implementation

[0026] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0027] In this disclosure, unless otherwise stated, the use of terms such as "first," "second," etc., to describe various elements is not intended to limit the positional, temporal, or importance relationships of these elements; such terms are merely used to distinguish one element from another. In some examples, the first element and the second element may refer to the same instance of that element, while in other cases, based on the context, they may refer to different instances.

[0028] The terminology used in the description of the various examples described in this disclosure is for the purpose of describing particular examples only and is not intended to be limiting. Unless the context explicitly indicates otherwise, an element may be one or more unless the number of elements is specifically limited. As used herein, the term "multiple" means two or more, and the term "based on" should be interpreted as "at least partially based on". Furthermore, the terms "and / or" and "at least one of..." cover any one of the listed items and all possible combinations thereof.

[0029] Please see Figures 1-3 , Figure 1 This is a front view of an AI computing card according to some embodiments of this disclosure; Figure 2 This is a top view of an AI computing card according to some embodiments of this disclosure; Figure 3 This is an exploded structural diagram of an AI computing card according to some embodiments of this disclosure.

[0030] One aspect of this disclosure provides an AI computing card 100, which includes a first circuit board 110, a second circuit board 120, at least one cross-board connector 130, and a heat dissipation module 140. The first circuit board 110 is provided with a first AI chipset 111, a first interface 112, and a second interface 113. The first AI chipset 111 includes at least one AI chip, and the first interface 112 is configured to communicate and / or power supply couple with an external device. The second circuit board 120 is arranged opposite to and spaced apart from the first circuit board 110 along the thickness direction, and the second circuit board 120 is provided with a second AI chipset 121 and a third interface. 122, the second AI chipset includes at least one AI chip; the cross-board connector 130 is respectively coupled to the second interface 113 and the third interface 122, the cross-board connector 130 is provided with a first type of pin and a second type of pin, the first type of pin is configured to realize the interconnection communication between the first circuit board 110 and the second circuit board 120, and the second type of pin is configured to realize the power transmission between the first circuit board 110 and the second circuit board 120; the heat dissipation module 140 is clamped between the first circuit board 110 and the second circuit board 120, and is thermally coupled to the first AI chipset 111 and the second AI chipset 121.

[0031] like Figure 3 As shown, the first circuit board 110 and the second circuit board 120 are arranged at intervals relative to each other along the first direction X, which is parallel to the thickness direction of the circuit board.

[0032] In some examples, the first AI chipset 111 includes multiple AI chips. Exemplarily, the multiple AI chips may be spaced apart along a second direction Y on the first circuit board 110. Similarly, the second AI chipset 121 may also include multiple AI chips spaced apart along a second direction Y. It is understood that the AI ​​chips in the first AI chipset 111 and the second AI chipset 121 may be chips of the same type or specifications, or they may be arranged in different types or specifications as needed. The number of AI chips in the first AI chipset 111 and the second AI chipset 121 may be the same or different.

[0033] The cross-board connector 130 is a connector used to enable signal transmission and power distribution between two different circuit boards. Both ends of the cross-board connector 130 have connector heads that couple to the second interface 113 of the first circuit board 110 and the third interface 122 of the second circuit board 120, respectively. The coupling method can be any feasible method, such as plug-in, soldering, or latching connection. There can be one or more cross-board connectors 130. Multiple cross-board connectors 130 can be located on the same side of the circuit board or on different sides. It is understood that multiple cross-board connectors 130 can increase the communication bandwidth between the two circuit boards.

[0034] The cross-board connector 130 can be equipped with two types of pins. The first type of pins are communication pins, which, for example, can be high-speed differential signal pins supporting the 802.3ck protocol to achieve high-bandwidth, low-latency communication between AI chips on different circuit boards. The second type of pins are power supply pins, which, for example, are 12V high-power power transmission pins, capable of carrying more than 10A of current per pin. Through a parallel design of multiple sets of pins, cross-board power supply requirements exceeding 1000W can be met. The first and second types of pins can be designed in a partitioned manner. In some embodiments, the first and second types of pins can be isolated from each other through a ground pin to reduce signal interference.

[0035] The first interface 112 on the first circuit board 110 is the interface for the AI ​​computing card to achieve data interaction and power access. It is configured at the docking end of the first circuit board 110, such as the side or bottom of the computing card. In some examples, the first interface 112 is compatible with the standard card slot interface protocol of a standard server or UBB (Universal Baseboard) substrate, and can achieve communication coupling and / or power supply coupling with external devices (such as UBB substrate, server backplane, switch), supporting single-function independent coupling or dual-function simultaneous coupling.

[0036] The communication transmission section of the first interface 112 is compatible with at least one high-speed interconnect protocol among PCIe, OAM, CXL, and GENZ. In terms of coupling, the first interface 112 can be adapted to the corresponding communication interface of external devices through any feasible method, such as gold finger insertion or physical coupling between boards, to achieve electrical connection and comply with the installation specifications of standard servers or OAM card slots. It should be noted that the first interface 112 can be an integrated interface with both communication and power supply functions, for example, using different pins to achieve signal transmission and power transmission respectively; the first interface 112 can also be a collective term for multiple interfaces containing independent communication and power interfaces.

[0037] The first interface 112 is coupled to an external device to realize signal interaction and power transmission between the AI ​​computing card 100 and the external device. Then, the power supplied by the external device is transmitted from the first circuit board 110 to the second circuit board 120 through the cross-board connector 130, thereby providing power to the second AI chipset.

[0038] The heat dissipation module 140 can be any feasible module unit with heat dissipation function. The heat dissipation module 140 is disposed between the first circuit board 110 and the second circuit board 120, forming a sandwich structure. The heat dissipation module 140 is simultaneously thermally coupled to the first AI chipset 111 and the second AI chipset 121, thereby simultaneously dissipating heat from the AI ​​chips on different circuit boards. It should be noted that the first AI chipset 111 and the second AI chipset 121 can be thermally coupled to the heat dissipation module 140 in any feasible manner, such as direct contact to achieve heat conduction, or indirect heat conduction through the provision of heat-conducting components or structures.

[0039] In some examples, the heat dissipation module 140 can be powered independently, for example, by a separate power supply unit providing the necessary energy, such as electrical energy or pressure energy. In other examples, the heat dissipation module 140 can also be powered by an electrical connection to the first circuit board 110 or the second circuit board 120. The connection between the heat dissipation module 140 and the circuit board can be achieved by a snap-fit ​​structure, a fastener connection, or any other feasible method, which will not be listed in this disclosure.

[0040] In some embodiments, the heat dissipation module serves as a clamping structure between the two boards, combining heat dissipation and structural support functions. This eliminates the need for additional inter-board support components, further optimizing the space utilization of the AI ​​computing card.

[0041] The AI ​​computing card disclosed herein improves the computing power density per unit volume through a double-layer circuit board stacking design. The first interface 112 and the cross-board connector 130 cooperate with each other to achieve integrated interconnection and power supply between multiple circuit boards while being compatible with external device interface specifications. The heat dissipation module 140 sandwiched between the two circuit boards achieves efficient heat dissipation on both sides, improving structural compactness and heat dissipation efficiency.

[0042] Please see Figure 4 , Figure 4 for Figure 2A cross-sectional view along the AA direction. In some embodiments, the heat dissipation module 140 includes a first heat dissipation surface 141 facing the first circuit board 110 and a second heat dissipation surface 142 facing the second circuit board 120; the first heat dissipation surface 141 is attached to the first AI chip group 111, and the projection of the first AI chip group 111 along the thickness direction of the circuit board falls completely within the range of the first heat dissipation surface 141; the second heat dissipation surface 142 is attached to the second AI chip group 121, and the projection of the second AI chip group 121 along the thickness direction of the circuit board falls completely within the range of the second heat dissipation surface 142.

[0043] The first heat dissipation surface 141 and the second heat dissipation surface 142 are two surfaces arranged facing away from each other. The first AI chipset 111 includes a plurality of first AI chips 1101 disposed on the side of the first circuit board 110 facing the heat dissipation module 140, and the second AI chipset 121 includes a plurality of second AI chips 1201 disposed on the side of the second circuit board 120 facing the heat dissipation module 140. The area of ​​the first heat dissipation surface 141 is greater than the sum of the orthographic projection areas of all the first AI chips 1101 along the first direction X, and the lower surface of each first AI chip 1101 is in contact with the first heat dissipation surface 141, which maximizes the thermal contact area and improves the heat conduction efficiency. Similarly, the second heat dissipation surface 142 and the second AI chipset 121 can also adopt the same design, which will not be described in detail here. In this way, each AI chip is thermally coupled to the heat dissipation module 140, meeting the thermal management requirements of multiple AI chips in the AI ​​computing card 100.

[0044] In some embodiments, the first AI chip 1101 can be bonded to the first heat dissipation surface 141 using thermally conductive adhesive, achieving bonding and fixation while improving heat conduction efficiency. Similarly, the second AI chip 1201 can also be connected to the second heat dissipation surface 142 using thermally conductive adhesive.

[0045] Please see Figure 5 , Figure 5 For other embodiments of this disclosure Figure 2 Cross-sectional view along the AA direction. In some embodiments, at least one of the first heat dissipation surface 141 and the second heat dissipation surface 142 is provided with a recess, and at least a portion of at least one of the first AI chipset 111 and the second AI chipset 121 is embedded in the corresponding recess.

[0046] like Figure 5As shown, the first heat dissipation surface 141 is provided with a first recess 1401, and at least a portion of at least one first AI chip 1101 extends into the first recess 1401. In some examples, the number of first recesses 1401 is the same as the number of first AI chips 1101, and they are arranged in a one-to-one correspondence, with each first AI chip 1101 extending at least partially into the corresponding first recess 1401 along the first direction X. Similarly, the second heat dissipation surface 142 is provided with a second recess 1402, and a second AI chip 1201 extends at least partially into the second recess 1402 along the first direction X.

[0047] In some embodiments, the depth of the recess along the first direction X is equal to the height of the AI ​​chip protruding along the first direction X. This allows the heat dissipation module 140 to not only adhere to the AI ​​chip but also contact the surfaces of the first circuit board 110 and the second circuit board 120, thereby providing support for the two circuit boards and dissipating heat for other circuit components on the circuit boards.

[0048] In some embodiments, the heat dissipation module 140 is a liquid-cooled heat sink or an air-cooled heat sink.

[0049] A liquid-cooled heat sink is a device that dissipates heat from a target component through heat exchange via a flowing liquid. For example, a liquid-cooled heat sink has a coolant inlet, an outlet, and a closed flow channel connecting the inlet and outlet. The flow channel can be arranged in a serpentine pattern and is correspondingly positioned to correspond to all chips in the first AI chipset 111 and the second AI chipset 121, enabling heat exchange between them. Thermally conductive silicone or thermal pads can be attached to the heat dissipation surfaces on both sides of the liquid-cooled heat sink to fill the gap between the heat dissipation surface and the chip, reducing thermal resistance and improving heat conduction efficiency.

[0050] An air-cooled heat sink refers to a device that uses airflow to dissipate heat from a target component. For example, an air-cooled heat sink includes a thermally conductive substrate, heat sink fins, and a fan. The thermally conductive substrate has a double-sided panel structure, with the first AI chipset 111 and the second AI chipset 121 attached to its two sides respectively. The heat sink fins are located in the middle of the thermally conductive substrate, and the fan is arranged correspondingly to the heat sink fins, providing forced airflow to the interior of the AI ​​computing card 100. The thermally conductive substrate is made of a material with high thermal conductivity, such as copper or aluminum, which allows for rapid heat transfer from the AI ​​chips to the central heat sink fins. The fan drives airflow through the heat sink fins, and through convection heat transfer, the heat is dissipated to the outside of the AI ​​computing card 100, achieving simultaneous heat dissipation for both chips.

[0051] Please see Figure 3 and Figure 6 , Figure 6 for Figure 2A cross-sectional view along the BB direction. In some embodiments, the second interface 113 is located on the side of the first circuit board 110, and the third interface 122 is located on the side of the second circuit board 120; the cross-board connector 130 is inserted and removed from the second interface 113 and the third interface 122 in a direction perpendicular to the thickness direction of the circuit board.

[0052] The second interface 113 can be integrally formed or welded to the side edge area of ​​the first circuit board 110, and the third interface 122 can be integrally formed or welded to the side edge area of ​​the second circuit board 120. Both are arranged vertically in the overall structure of the AI ​​computing card 100, and are located on the same side of the circuit boards opposite each other. Both the second interface 113 and the third interface 122 can be horizontal slot structures, with the slot opening direction set along a horizontal direction perpendicular to the circuit board thickness, for example, along the third direction Z. The slot has pin contacts that match the cross-board connector 130, and are electrically connected to the internal traces of the first circuit board 110 and the second circuit board 120, respectively.

[0053] The cross-board connector 130 is a long, strip-shaped board-to-board interconnection structure. Its mating end structure matches the horizontal slots of the second interface 113 and the third interface 122. The entire connector is positioned horizontally perpendicular to the thickness of the circuit board, for example, in the third direction (Z), to facilitate insertion and removal of the two interfaces. In some examples, the cross-board connector 130 has mating gold fingers at both its top and bottom ends.

[0054] The insertion and removal direction between the cross-board connector 130 and multiple circuit boards is parallel to the surface of the circuit board and perpendicular to the thickness direction and stacking direction of the circuit board. This structure allows the arrangement and connection of the cross-board connector 130 to not occupy the thickness space between the two-layer circuit boards, thereby effectively controlling the external dimensions of the stacked AI computing card 100. This avoids the problem of the thickness of the stacked two-layer circuit boards increasing further and becoming incompatible with standard card slots, thus improving the applicability of the AI ​​computing card 100. Moreover, the flexible insertion and removal method can also meet diverse connection needs.

[0055] Please see Figure 7 , Figure 7 For other embodiments of this disclosure Figure 2 Cross-sectional view along the BB direction. In some embodiments, at least one of the second interface 113 and the third interface 122 is a recessed connector, which is at least partially embedded in the interior of the corresponding circuit board along the thickness direction of the circuit board.

[0056] The main structure of the recessed connector is at least partially embedded in the interior of the circuit board along its thickness direction. The embedded portion is adapted to the recessed or cutout area of ​​the circuit board, with only the slot or gold finger end required for mating exposed. The second interface 113 and the third interface 122 can either be configured as a recessed connector or both can be configured as recessed connectors.

[0057] The recessed connector can be positioned on the surface of the corresponding circuit board facing the heat dissipation module 140. For the recessed connector, the insertion / removal direction of the cross-board connector 130 and the circuit board becomes parallel to the first direction X. In this way, the connection between the cross-board connector 130 and the circuit board can be arranged entirely within the scope of a single circuit board without exceeding its boundaries. This allows for better control over the size of a single AI computing card, thus improving compatibility with server architectures.

[0058] In some embodiments, the heat dissipation module 140 further includes a support structure 144 for providing support to the cross-board connector 130.

[0059] Please continue reading. Figure 7 The heat dissipation module 140 also includes a side surface 143 connecting the first heat dissipation surface 141 and the second heat dissipation surface 142. The side surface 143 can be a surface parallel to the first direction X. A support structure 144 is disposed on the side surface 143 to provide support for the crossboard connector 130. For example, the support structure 144 can be a snap-fit ​​structure protruding from the side surface 143, which provides support for the crossboard connector 130 or limits the crossboard connector 130 by snap-fit. In other embodiments, the support structure 144 can also be a through hole penetrating the heat dissipation module 140, through which the crossboard connector 130 passes and is connected to two circuit boards at both ends respectively.

[0060] In some embodiments, the support structure 144 provides support for the crossboard connector 130 while also being thermally coupled to the crossboard connector 130, thereby further dissipating heat from the crossboard connector 130.

[0061] The support structure 144 provides support for the cross-board connector 130, making the overall structure of the single AI computing card 100 more compact, improving the reliability of the connection between components, and controlling the size of the single card. On the other hand, the heat dissipation module 140 can also dissipate heat from the cross-board connector 130, thereby improving the thermal management capability of the AI ​​computing card 100.

[0062] In some embodiments, the first interface 112 adopts a PCIe interface or an OAM interface.

[0063] Please see Figure 2 and Figure 3For example, the first interface 112 can be configured with PCIe full-height, full-length, double-width standard gold fingers, thus adapting to the PCIe slots of servers. The first circuit board 110 and the second circuit board 120 are aligned by positioning posts, and the heat dissipation module 140 is fixed to the two circuit boards by bolts. The overall structural dimensions conform to the PCIe full-height, full-length, double-width card standard.

[0064] Please see Figure 8 , Figure 8 This is an exploded structural diagram of an AI computing card according to other embodiments of this disclosure. For example, the first interface 112 of the first circuit board 110 can be configured with an OAM standard interface and can be inserted into a UBB substrate of the OAM architecture. The second circuit board 120 is disposed on the side of the first circuit board 110 away from the first interface 112, so that the AI ​​computing card 100 can be adapted to the OAM modular card slot space as a whole.

[0065] Correspondingly, the cross-board connector 130 can be a GENZ interconnect panel that supports power transmission with a total power of not less than 1000W and a voltage of up to 54V.

[0066] In some embodiments, the first circuit board 110 is further provided with a first power supply interface 114, and the second circuit board 120 is further provided with a second power supply interface 123. The second power supply interface 123 and the first power supply interface 114 are power supply interfaces that are redundant backups of each other.

[0067] like Figure 3 As shown, in addition to the first interface 112, the first circuit board 110 also has an independent first power supply interface 114, which can be connected to an external power source to receive externally supplied power. The second power supply interface 123 is an independent power supply interface located on the second circuit board 120. For example, the first power supply interface 114 and the second power supply interface 123 can be 16-pin high-current power supply connectors.

[0068] It should be noted that the "mutual redundancy backup" claimed in this embodiment does not mean that only one of the two power supply interfaces can be used. Specifically, the second power supply interface 123 and the first power supply interface 114 can receive external power individually or simultaneously. For example, the power supply interfaces of the two circuit boards can draw power from external sources, such as the UBB substrate or server power supply, to independently power their respective AI chipsets. This halves the single-path power supply current, reducing the current transmission burden on the power supply pins of the cross-board connector, reducing line loss and heat generation, and improving the power supply stability of the dual AI chipsets operating at full load. In other examples, one of the circuit board's power supply interfaces, such as the first power supply interface 114, can draw power from the outside and then supply power from the first circuit board 110 to the second circuit board 120 through the cross-board connector 130. In this case, the second power supply interface 123 serves as a redundant backup for the first power supply interface 114, improving the reliability of power supply when the first power supply interface 114 fails.

[0069] Please see Figure 9 The image shows a partial structural diagram of a computing cluster according to some embodiments of this disclosure. According to another aspect of this disclosure, a computing cluster 1000 is provided, which includes multiple AI computing cards 100, and the multiple AI computing cards 100 are communicatively coupled to each other.

[0070] The computing cluster 1000 is a distributed computing architecture built by using multiple AI computing cards 100 as core computing units and connecting them through communication. Multiple AI computing cards 100 can achieve a preset interconnection topology through interconnection backplanes and switching devices. For example, multiple AI computing cards 100 can be connected to form at least one of several interconnection topologies, such as full connectivity, dragon topology, and star topology.

[0071] AI computing cards 100 are configured with an OAM standard structure and are respectively inserted into UBB substrates 200, which also have an OAM architecture. For example, multiple AI computing cards 100 are compatible with at least one of the communication protocols of GENZ, CXL, PCIe 6.0 and 802.3ck.

[0072] Because each AI computing card 100 has a double-layer high-density design, the computing power per unit volume is doubled. The overall computing power of a cluster built by coupling multiple cards is significantly improved in the same physical space. The hardware interface for communication coupling of multiple AI computing cards 100 can directly reuse the existing data center infrastructure without the need for reconstruction, reducing the cost of building and deploying the cluster.

[0073] In this specification, multiple instances may implement components, operations, or structures described as a single instance. Although individual operations of one or more methods are described and illustrated as independent operations, one or more individual operations may be performed concurrently, and these operations are not required to be performed in the order shown. Structures and functionalities presented as independent components in the example configuration may be implemented as combined structures or components. Similarly, structures and functionalities presented as individual components may be implemented as independent components. These and other variations, modifications, additions, and improvements are all within the scope of this document.

[0074] As used herein, “or” is inclusive rather than exclusive unless explicitly stated or indicated by context. Furthermore, “and” is both common and individual unless explicitly stated or indicated by context. Moreover, multiple instances may be provided for the resources, operations, or structures described herein as a single example. Furthermore, the boundaries between various resources, operations, engines, and data stores are somewhat arbitrary, and specific operations are illustrated within the context of a particular illustrative configuration. The allocation of other functionalities is conceivable and may fall within the scope of various embodiments of this disclosure. Generally, structures and functionalities presented as independent resources in example configurations may be implemented as combined structures or resources. Similarly, structures and functionalities presented as individual resources may be implemented as independent resources. These and other variations, modifications, additions, and improvements are all within the scope of embodiments of this disclosure. Therefore, this specification and accompanying drawings should be viewed in an illustrative rather than restrictive sense.

[0075] The terms “comprising” or “including” are used to indicate the presence of a subsequently stated feature, but do not preclude the addition of other features. Conditional language, in particular, such as “may,” “can,” or “may,” unless specifically stated or otherwise understood in the context of use, is generally intended to express that certain embodiments include certain features, elements, and / or steps, while other embodiments do not. Therefore, such conditional language generally does not imply that a feature, element, and / or step is necessary in any way for one or more embodiments, or that one or more embodiments must include logic that, with or without user input or prompting, determines whether such features, elements, and / or steps are included in any particular embodiment, or whether they are to be performed in any particular embodiment.

Claims

1. An AI computing power card, characterized in that, include: The first circuit board is provided with a first AI chipset, a first interface and a second interface. The first AI chipset includes at least one AI chip, and the first interface is configured to communicate and / or power supply couple with an external device. The second circuit board is arranged opposite to and spaced from the first circuit board along the thickness direction. The second circuit board is provided with a second AI chipset and a third interface. The second AI chipset includes at least one AI chip. At least one crossboard connector, which is mated and coupled to the second interface and the third interface respectively. The crossboard connector has first-type pins and second-type pins. The first-type pins are configured to enable interconnection communication between the first circuit board and the second circuit board, and the second-type pins are configured to enable power transmission between the first circuit board and the second circuit board. A heat dissipation module is sandwiched between a first circuit board and a second circuit board, and is thermally coupled to both the first AI chipset and the second AI chipset. Wherein, the second interface and the third interface are both recessed connectors and are positioned facing the heat dissipation module. The recessed connector is at least partially embedded inside the corresponding circuit board along the thickness direction of the circuit board. The cross-board connector is inserted into the second interface and the third interface respectively along the thickness direction of the circuit board. The heat dissipation module is provided with a support structure, which is used to provide support to the cross-board connector, and the heat dissipation module is thermally coupled to the cross-board connector.

2. The AI ​​computing card according to claim 1, characterized in that, The heat dissipation module includes a first heat dissipation surface facing the first circuit board and a second heat dissipation surface facing the second circuit board; The first heat dissipation surface is attached to the first AI chipset, and the projection of the first AI chipset along the thickness direction of the circuit board falls completely within the range of the first heat dissipation surface. The second heat dissipation surface is attached to the second AI chipset, and the projection of the second AI chipset along the thickness direction of the circuit board falls completely within the range of the second heat dissipation surface.

3. The AI ​​computing card according to claim 2, characterized in that, At least one of the first heat dissipation surface and the second heat dissipation surface is provided with a recess, and at least a portion of the first AI chip group and the second AI chip group is embedded in the corresponding recess.

4. The AI ​​computing card according to any one of claims 1-3, characterized in that, The heat dissipation module is a liquid-cooled radiator or an air-cooled radiator.

5. The AI ​​computing card according to claim 2 or 3, characterized in that, The heat dissipation module also includes a side surface connecting the first heat dissipation surface and the second heat dissipation surface, and the support structure is disposed on the side surface.

6. The AI ​​computing card according to any one of claims 1-3, characterized in that, The first interface uses either a PCIe interface or an OAM interface.

7. The AI ​​computing card according to any one of claims 1-3, characterized in that, The first circuit board is further provided with a first power supply interface, and the second circuit board is further provided with a second power supply interface; The second power supply interface and the first power supply interface are redundant backup power supply interfaces.

8. A computing cluster, characterized in that, It includes multiple AI computing cards as described in any one of claims 1-7, and the multiple AI computing cards are mutually communicative and coupled.

Images