Artificial intelligence inference computing mainboard based on VPX architecture

Abstract

The utility model discloses an artificial intelligence reasoning calculation mainboard based on VPX framework, include: the carrier board and heat abstract component and the AI core board of carrying neural processing unit that accord with VPX standard, install detachably on the carrier board in the way of laminating, and realize interconnection through the inter -plate insertion interface. The heat abstract component covers carrier board and AI core board, and contains at least one stereoscopic modeling phase change heat conduction element, the element absorbs its heat end and neural processing unit heat coupling, and extends its heat release end to the carrier board heat abstract area of more optimal heat abstract condition, to solve the high density laminated architecture heat abstract bottleneck under the restricted space. The utility model will general calculation and AI reasoning function high integration in single VPX slot, realized the nationalization artificial intelligence of traditional VPX hardware empowerment, improved the integrated level and heat -power coupling reliability of system, realized the quick upgrade and expansion of VPX platform AI reasoning calculation ability.

Description

Technical Field

[0001] This utility model relates to artificial intelligence hardware devices, and in particular to an artificial intelligence inference computing motherboard. Background Technology

[0002] With the rapid development and deep application of artificial intelligence (AI) technology, upgrading existing hardware systems with AI capabilities has become an urgent technical need in fields such as military aviation, communications, and industrial control, where reliability, integration, and independent control are extremely stringent. The VPX architecture, as a key infrastructure in these fields, has become the mainstream computing hardware platform standard due to its robustness, modularity, and high-bandwidth backplane interconnect. Therefore, how to efficiently and reliably integrate high-performance AI real-time inference computing functions within existing VPX hardware systems, while simultaneously meeting stringent environmental adaptability requirements, is a problem that needs to be solved.

[0003] Currently, the mainstream technical approach for AI hardware empowerment of the VPX platform mainly revolves around adding computing units to existing systems. One common solution is to add a dedicated, VPX-compliant AI inference accelerator card to the VPX chassis. This card typically carries a dedicated AI processor (such as an NPU or GPU) and interacts with the general-purpose computing motherboard (CPU board) in the chassis via the VPX backplane. Another more common approach is system-level expansion, which involves externally connecting a separate industrial control computer or server with AI inference capabilities to the entire VPX computing platform via an external interface. The external server usually uses a more general-purpose hardware architecture, such as an x86 processor-based platform, and is configured with high-performance commercial AI accelerator cards. These solutions, to a certain extent, achieve functional expansion, providing a preliminary technical approach for introducing AI computing capabilities into VPX systems and forming the foundation for current technological development.

[0004] However, the existing solutions mentioned above still have some problems when dealing with the deep application needs of specific fields. For example: the space for new equipment is limited, and the number of slots in the VPX enclosure is limited, which may leave no spare slots for board insertion and removal. Traditional server and industrial PC hardware is too large and heavy, with high heat dissipation, and heat dissipation becomes a problem in the limited space of the VPX enclosure; reliability and environmental adaptability are poor, and traditional industrial PCs and servers can only meet commercial or industrial needs, and are not suitable for harsh conditions such as high and low temperatures and vibration in some fields; the platform is not universally compatible, and traditional industrial PCs and servers need to add interfaces to achieve interconnection with the VPX architecture, which is less compatible than board interconnection through the VPX back panel. In addition, there is the risk of board-level thermal-mechanical coupling failure caused by high functional density integration, and the long-term reliability challenge of high-density interconnect components in harsh environments. Utility Model Content

[0005] The purpose of this utility model is to solve the above-mentioned problems existing in the prior art and to provide an artificial intelligence inference computing motherboard based on the VPX architecture.

[0006] Technical solution: An AI inference computing motherboard based on the VPX architecture, comprising:

[0007] A carrier board that conforms to the VPX standard and has a general-purpose processor installed on it;

[0008] Heat dissipation components used to dissipate heat from the carrier plate;

[0009] An AI core board equipped with a neural processing unit is detachably mounted on a carrier board in a stacked manner.

[0010] The inter-board interface, located between the carrier board and the AI ​​core board, is used to enable data and power transmission between the two.

[0011] The heat dissipation components are also used to dissipate heat for the AI ​​core board that is stacked on top of the carrier board.

[0012] Optionally, the heat dissipation component includes:

[0013] The first heat dissipation area is used to dissipate heat for the general-purpose processor;

[0014] The second heat dissipation area, located above the AI ​​core board, is used to dissipate heat for the neural processing unit;

[0015] A thermally conductive element is used to uniformly conduct and diffuse heat from the device along a plane to a first or second heat dissipation area, wherein at least one thermally conductive element is three-dimensional and has a heat absorption end thermally coupled to the neural processing unit and a heat release end.

[0016] The heat release end bypasses the physical shielding of the AI ​​core board, extends into the first heat dissipation area and is thermally coupled thereto, transferring part of the heat generated by the neural processing unit to the first heat dissipation area for dissipation;

[0017] The heat dissipation fins are located on the back of the first and second heat dissipation areas, and diffuse heat into the flowing air in the chassis heat dissipation duct.

[0018] Optionally, the three-dimensional heat-conducting element is specifically:

[0019] One or more heat pipes are pre-bent and contain a phase change medium. The bending shape of the heat pipe is adapted to the planar and three-dimensional spatial layout between its heat absorption end and heat release end, and heat can be uniformly conducted in the heat pipe cavity through the phase change medium.

[0020] Optionally, the three-dimensional heat-conducting element can also be specifically:

[0021] The vacuum cavity heat exchanger has a stepped thickness and contains a phase change medium. The stepped structure of the vacuum cavity heat exchanger allows it to physically overcome the spatial barrier between the AI ​​core board and the carrier board, and heat can be uniformly conducted in the cavity through the phase change medium.

[0022] Optionally, the three-dimensional heat-conducting element can also be specifically:

[0023] A micro-pump liquid cooling system with circulating flow channels, wherein at least one set of flow channels is spatially connected to the first and second heat dissipation areas, and heat can be uniformly conducted in a planar manner and spatially across the first and second heat dissipation areas by circulating coolant.

[0024] Optionally, the contact surface between the general-purpose processor and / or neural processing unit and the heat dissipation component has a microtexture on one side of the heat dissipation component;

[0025] Between the contact surfaces of the processor and the heat dissipation components, a self-locking thermal interface composite material is filled. The composite material includes a thermally conductive matrix and shape memory alloy microfilaments dispersed therein.

[0026] Optionally, an AI inference computing motherboard based on the VPX architecture further includes: at least one heat-conducting element is a planar heat-conducting element.

[0027] Beneficial effects: This utility model highly integrates general computing and AI inference functions into a single VPX slot, realizing the localization of artificial intelligence empowerment of traditional VPX hardware, and improving the system's integration and thermal-mechanical coupling reliability. Attached Figure Description

[0028] Figure 1 This is a schematic block diagram of the AI ​​VPX inference computing motherboard of this application.

[0029] Figure 2 This is a view showing the external shape and cross-section of this application.

[0030] Figure 3 This is a schematic diagram showing the mating of the COME and MXM core boards with the VPX carrier board.

[0031] Figure 4 This is a schematic diagram of the core board with COME and MXM interfaces.

[0032] Figure 5 This is a schematic diagram of the heat dissipation component.

[0033] Figure 6 This is a diagram illustrating the principle of phase change heat transfer.

[0034] Figure 7 This is a schematic diagram of a micro-pump liquid cooling system. Detailed Implementation

[0035] For clarity, one or more specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0036] Example 1: A motherboard for artificial intelligence inference computing based on VPX architecture is provided, referring to... Figures 1 to 6 .

[0037] The motherboard includes a VPX-compliant carrier board 1, on which general-purpose processors 1-4 are mounted; and a heat dissipation assembly 4 for cooling the carrier board.

[0038] Specifically, it also includes an AI core board 2 equipped with a neural processing unit 2-2, which is detachably mounted on the carrier board 1 in a stacked manner; an inter-board interface 3 is provided between the carrier board 1 and the AI ​​core board 2 to realize data and power transmission between the two; and a heat dissipation component 4 is also used to dissipate heat for the AI ​​core board 2 stacked on the carrier board 1.

[0039] like Figure 1 and Figure 2 As shown, the motherboard as a whole constitutes a 6U 5HP front-mounted air-cooled card structure that conforms to the VITA48 specification. The AI ​​core board 2 is fixed to the carrier board 1 by mounting screws 8. The heat dissipation component 4 covers the top of the carrier board 1 and the core board 2, and forms a complete and robust VPX hardware unit through the panel assembly 6 and the back cover 7.

[0040] Reference Figure 3 Carrier board 1 is a self-developed 6U general-purpose computing VPX board. The general-purpose processors 1-4 on it can be Phytium series CPU chips, such as FT2000, D2000, or D3000. Carrier board 1 also integrates memory modules, storage modules, PCIe modules, power supply modules, etc., which together constitute carrier board chipset 1-5, and run the Kylin operating system, forming a domestic VPX general-purpose computing platform.

[0041] Reference Figure 4 The AI ​​core board 2 is a standard packaged general-purpose artificial intelligence inference computing module. The neural processing unit 2-2 on it can be a specific Ascend series NPU, such as 310B or 310P. The core board also integrates a memory module, a codec module, a power supply module, etc., which together constitute the core board chipset 2-3.

[0042] The separate hardware design of the general-purpose computing VPX carrier board and the AI ​​core board allows self-developed hardware to be concentrated on the VPX carrier board, while the AI ​​core board can directly use mature commercially available modules, reducing the R&D threshold, risk and cycle.

[0043] Reference Figure 3 and Figure 4 The inter-board mating interface 3 is a crucial component for enabling the combined functions of the two boards. Its connector 3-1 is located at the bottom of the AI ​​core board 2, and its socket 3-2 is located at the top of the carrier board 1. When mated, they enable high-speed data interconnection and power transmission. In optional implementations, the specific standard of this interface can be MXM or COME, and the mating height is preferably less than or equal to 6mm to meet the stringent hardware stacking space constraints and heat dissipation requirements of the 5HP board.

[0044] Reference Figure 2 , Figure 5 and Figure 6 The heat dissipation component 4 is a crucial component in solving the heat dissipation challenge of this highly integrated, high-power multilayer board. Specifically, the heat dissipation component 4 defines a first heat dissipation area (first heat dissipation surface) 4-2 for dissipating heat for the general-purpose processor 1-4 and the carrier chipset 1-5, and a second heat dissipation area (second heat dissipation surface) 4-3 located above the AI ​​core board 2 for dissipating heat for the neural processing unit 2-2 and the core board chipset 2-3. The heat dissipation component also includes heat dissipation fins located on the back of the first and second heat dissipation areas, which diffuse heat into the flowing air within the chassis's airflow channels for exhaust.

[0045] Due to the stacked structure of the AI ​​core board 2, the height of the heat dissipation fins 4-9 above the second heat dissipation area 4-3 is limited, resulting in a naturally weaker heat dissipation capacity compared to the first heat dissipation area 4-2, which has taller fins. To address this issue, further:

[0046] The heat dissipation assembly 4 also includes a heat-conducting element for uniformly conducting and diffusing the heat of the device along the plane to the first or second heat dissipation area, wherein at least one heat-conducting element has a three-dimensional shape.

[0047] The three-dimensional heat-conducting element 4-4 has a heat absorption end that is thermally coupled to the neural processing unit 2-2. The three-dimensional heat-conducting element 4-4 also has a heat release end, which bypasses the physical shielding of the AI ​​core board 2 and extends into and thermally couples with the first heat dissipation area 4-2 through its three-dimensional shape, transferring some of the heat generated by the neural processing unit 2-2 to the first heat dissipation area 4-2 for dissipation. Through this cross-zone heat transfer design, the concentrated heat generated by the high-power NPU is efficiently guided to an area with stronger heat dissipation capabilities, overcoming the physical bottleneck of heat dissipation for the motherboard and daughterboard in a confined space, and ensuring that the temperature of the entire board is controlled within a stable operating range.

[0048] In this embodiment, the three-dimensional heat-conducting element 4-4 is specifically: one or more heat pipes that have been pre-bent and contain a phase change medium. The bending shape of the heat pipe is adapted to the plane and three-dimensional spatial layout between the heat absorption end and the heat release end of the heat pipe. Heat can be uniformly conducted in the heat pipe cavity through the phase change medium.

[0049] The aforementioned heat pipes are embedded and welded into the groove of the AL6063 high thermal conductivity aluminum alloy finned cold plate 4-1 to reduce contact thermal resistance and achieve efficient heat transfer.

[0050] The heat transfer principle of a heat pipe is as follows: the inside of the heat pipe is a vacuum environment, the inner wall of the pipe has a capillary structure, and the pipe is filled with a phase change medium. When the heat absorption end is heated, the medium rapidly vaporizes and changes phase, and when it encounters cold at the release end, it quickly liquefies and releases heat, transferring the heat to the release end. At the same time, the liquefied medium flows back to the heat absorption end through the capillary structure of the pipe wall, repeating the next cycle.

[0051] Furthermore, at least one of the heat-conducting elements is a planar heat-conducting element.

[0052] This embodiment is designed to address the following technical bottlenecks in existing solutions, specifically:

[0053] Uneven heat dissipation in high-density stacked architectures is a primary challenge restricting performance and reliability. To achieve functional integration within a single VPX slot, a stacked design of carrier board + core board has become a trend. However, in compact structures, the upper AI core board is typically the concentrated area for high-power devices such as the NPU, while the heat dissipation components directly above it, due to space constraints, have significantly smaller heat dissipation fin height and effective heat dissipation area than the carrier board area, forming a natural and severe heat dissipation bottleneck. Conventional planar heat dissipation designs cannot effectively dissipate the concentrated heat in this area, leading to excessively high local temperatures in the NPU chip. This not only limits its performance but also accelerates device aging, seriously threatening its long-term reliability under wide-temperature and vibration / shock environments.

[0054] Example 2: This example is based on Example 1 and provides optional or further optimized technical solutions for the three-dimensional heat-conducting elements in the heat dissipation assembly and the heat-conducting interface material between the processor and the heat sink.

[0055] As an alternative to the bent heat pipe in Embodiment 1, the three-dimensional heat-conducting element 4-4 can also be specifically:

[0056] A vacuum cavity heat exchanger (VC) with stepped thickness and containing a phase change medium has a stepped structure that allows it to physically overcome the spatial barrier between the AI ​​core plate 2 and the carrier plate 1, and heat can be uniformly conducted in the cavity through the phase change medium.

[0057] Specifically, the VC heat spreader is thinner in the portion corresponding to the second heat dissipation area 4-3, while it thickens accordingly in the portion extending to the first heat dissipation area 4-2, forming a certain height difference with the second heat dissipation area 4-3. Its internal vacuum chamber is entirely interconnected, allowing the phase change medium to fully absorb heat, vaporize, and liquefy, releasing heat. This stepped, integrated structure also enables rapid, low-thermal-resistance heat transfer from the second heat dissipation area (which has weaker heat dissipation capacity due to limited fin height) to the first heat dissipation area, and its temperature uniformity may be superior to solutions with multiple heat pipes.

[0058] To further optimize the heat transfer efficiency and long-term reliability between the processor (such as general-purpose processor 1-4 or neural processing unit 2-2) and the heat dissipation component 4, this invention also provides a preferred thermal interface material scheme, as follows:

[0059] In the contact surface between the general-purpose processor 1-4 and / or neural processing unit 2-2 and the heat dissipation component 4, a micro-texture is processed on one side of the heat dissipation component; between the contact surface between the processor and the heat dissipation component 4, a self-locking thermal interface composite material 5 is filled, the composite material comprising a thermally conductive matrix and shape memory alloy microfilaments dispersed therein. For example, a biomimetic shark skin-like micro-texture (such as a micro pyramid array with a depth of 50 μm and a spacing of 100 μm) can be processed on the bottom surface of the thermally conductive protrusions / grooves 4-6 of the heat dissipation component 4 (grooves are opened if the device height exceeds the heat dissipation surface) and the thermally conductive grease filled between it and the chip surface, then pretreated nickel-titanium alloy microfilaments with a specific volume fraction (such as 8-12%) are mixed in. Its working mechanism is as follows:

[0060] When the operating temperature exceeds the phase transition point of the shape memory alloy microfilament (e.g., 55°C), the microfilament undergoes a morphological change (e.g., from a straight state at low temperature to a curved state at high temperature) to generate mechanical locking force within the microstructure, suppressing the flow failure of the thermally conductive matrix during thermal cycling (i.e., pumping effect).

[0061] Furthermore, the microtexture on the processor contact surface and the microtexture on the heat dissipation component contact surface are geometrically complementary; wherein, after the shape memory alloy microfilament undergoes morphological changes, its geometry can form a mechanical embedding or wedging fit with the complementary microtexture to more firmly anchor the heat-conducting substrate between the contact surfaces.

[0062] Utilizing the micro-mechanical self-locking effect generated by material phase transition, it has been applied to thermal interface materials for the first time. It can improve the long-term reliability of the thermal interface and prevent the degradation of interface thermal resistance caused by silicone grease pumping out. It is especially suitable for harsh working environments that need to withstand severe temperature cycling.

[0063] Example 3: This example is based on Examples 1 and 2, and provides optional or further optimized technical solutions for three-dimensional heat-conducting elements in heat dissipation components.

[0064] As an alternative to the bent heat pipe and stepped thickness vacuum cavity heat spreader in Embodiments 1 and 2, the three-dimensional heat-conducting element 4-4 can also be specifically: a micro-pump liquid cooling system with circulating channels, and at least one set of channels spatially connecting the first and second heat dissipation areas, so that heat can be uniformly conducted in a planar manner and spatially across the first and second heat dissipation areas by circulating coolant.

[0065] The micro-pump liquid cooling system includes:

[0066] The flow channel is a coolant flow channel embedded in the first and second heat dissipation zones of the cold plate, and connects the first and second heat dissipation zones through a three-dimensional shape. It can be a pipe embedded in a pre-set groove in the cold plate through pipe flattening and welding processes, or it can be a flow channel embedded in the cold plate formed by machining and welding processes, with no obvious pipe visible from the outside.

[0067] Coolant, a medium used to conduct and transfer heat, can be an aqueous solution of ethylene glycol or glycerol with a freezing point of -60 to -40°C, meeting the requirements for use in low-temperature environments.

[0068] Micropumps are small pumps embedded in cold plates to circulate coolant in pipelines. They can be ultra-thin piezoelectric ceramic pumps or small impeller pumps.

[0069] like Figure 7 As shown, the micro-pump liquid cooling system circulates coolant within the flow channel using a micro-pump, uniformly transferring heat from the device to the first or second heat dissipation area where the flow channel is located. Furthermore, it achieves spatial cross-regional heat transfer through the three-dimensional flow channel, transferring some of the heat from the second heat dissipation area to the first heat dissipation area, which has better heat dissipation conditions, via the circulating coolant.

[0070] Example 4: Under certain operating conditions, a domestically produced AI VPX inference computing motherboard with a carrier board + core board architecture is provided, including a VPX carrier board, an AI core board, a mating interface, a cold plate assembly, a thermal interface material, a panel assembly, and a back cover.

[0071] The VPX inference computing motherboard is a 6U 5HP front-mounted air-cooled structure compliant with the VITA48 specification. The VPX carrier board is a domestically produced general-purpose computing motherboard, equipped with a Phytium series CPU and the Kylin operating system. The AI ​​core board is a domestically produced standard-packaged general-purpose artificial intelligence inference computing module, equipped with an Ascend series domestic NPU. The motherboard and core board are stacked and plugged in through a mating interface, supplemented by a high-performance air-cooled heat dissipation plate assembly with phase change heat conduction elements. This allows a single 6U 5HP VPX motherboard to integrate both general-purpose computing and artificial intelligence inference computing functions. It can be plugged into an existing VPX hardware platform chassis to enable autonomous and controllable artificial intelligence upgrades.

[0072] The VPX carrier board is a self-developed 6U general-purpose computing VPX board. One side of the PCB has the VPX connector, and the other side has the panel interface area. The PCB is equipped with Phytium series (FT2000, D2000, D3000) CPU chips, bridge chips, memory modules, storage modules, PCIe modules, BMC modules, serial port modules, 40G modules, power systems and other supporting chips and components. The VPX carrier board runs the domestic Galaxy Kylin operating system.

[0073] The AI ​​core board is a standard packaged general-purpose artificial intelligence inference computing module, equipped with Ascend series (310B, 310P) NPU chips, memory modules, encoding and decoding modules, power supply modules and related supporting chips and components.

[0074] The mating interface is a high-speed inter-board transmission interface, with the plug end located on one side of the AI ​​core board and the socket end located on the other side of the VPX carrier board. The mating interface adopts the COME or MXM standard form, and the mating height is ≤6mm (the distance between the BOT side of the AI ​​core board and the TOP side of the VPX carrier board after mating).

[0075] The cold plate assembly is 5HP thick, using AL6063 high thermal conductivity aluminum alloy as the base material. The top features heat dissipation fins, while the bottom has grooves cut to accommodate the carrier board, core board, and their components. For low-heat-dissipation devices, thermally conductive bosses or grooves are created on the aluminum-based cold plate. For high-heat-dissipation devices such as CPUs and NPUs, phase-change thermal conductive elements (PCTs) are embedded in corresponding locations; these can be heat pipes or vapor chambers (VCs). Planar PCTs are laid on the cold plate's heat dissipation surface for high-heat-dissipation devices on the carrier board, such as CPUs. For high-heat-dissipation devices on the core board, such as NPUs, due to their stacked interlocking structure above the carrier board, the limited thickness of the top cold plate and the low fin height result in poor heat dissipation. Therefore, in addition to planar PCTs, a certain number of three-dimensional PCTs are also installed on their corresponding cold plate's heat dissipation surface to transfer some heat to the carrier board's heat dissipation surface, which has higher fin height and lower temperature, thus improving the NPU's heat dissipation performance. For devices on the carrier board and core board located below the PCTs with a gap exceeding 1mm, copper thermal blocks are welded at the corresponding PCT locations. All heat dissipation components have a small gap between them and the heat-conducting surfaces of the cold plate bosses / grooves to compensate for component packaging size errors and cold plate processing errors. The gaps are filled with thermal interface material to improve thermal conductivity.

[0076] The finned cold plate 4-1 has several side ventilation openings 4-7 on both sides, allowing some of the heat dissipation airflow to enter the interior of the board along the ventilation openings to assist in the heat dissipation of low-power devices and the PCB. The thermal interface material can be a thermal pad, or TIM graphene, thermally conductive silicone sludge, silicone grease, etc., which densely fills the assembly gaps and tiny pits between the cold plate and the device, improving the contact area and thermal conductivity.

[0077] The panel assembly is located on the front of the board, including the panel and pull-out aid, enabling external insertion and removal. The panel has a slotted design to avoid interference with the VPX carrier board's front panel interfaces. The back cover is made of AL6063 high thermal conductivity aluminum alloy, used for shielding and protecting components on the back of the carrier board and for heat dissipation. The carrier board features a series of mating interfaces, allowing mating with MXM or COME interface core boards. It achieves 20 TOPS-INT8 computing power when paired with the standard Ascend 310B AI core board, and 176 TOPS-INT8 computing power when paired with the standard Ascend 310P AI core board. Users can choose the appropriate configuration based on their application needs, enabling the installation of different quantized versions of large AI models to perform offline image, video, and speech recognition, analysis, and generation; natural language processing; deep data analysis, screening, and classification; and agent-based intelligent decision-making and task scheduling, among other AI-related tasks.

[0078] The VPX inference computing motherboard adopts a 6U 5HP front-mounted air-cooled board design that conforms to the VITA48 specification. It integrates both general computing and artificial intelligence inference computing functions to form a complete AI-enabled hardware unit. A single slot can enable the artificial intelligence-enabled upgrade of the existing VPX hardware platform.

[0079] The VPX inference computing motherboard with a carrier board + core board architecture includes: VPX carrier board 1, AI core board 2, mating interface 3, cold plate assembly (heat dissipation assembly) 4, thermal interface material 5, front panel assembly 6, back cover 7, and mounting screws 8.

[0080] like Figure 3 As shown, the VPX carrier board is a 6UVPX board, including the carrier PCB 1-1, VPX connector 1-2, front panel interface area 1-3, CPU chip (general-purpose processor) 1-4, carrier chipset 1-5, and nut pillars 1-6. The CPU chip uses the Phytium series (FT2000, D2000, D3000) domestic CPU. The carrier chipset 1-5 includes supporting chips and components such as bridge chip, memory module, storage module, PCIe module, BMC module, network module, serial port module, and power system. The carrier board runs the Galaxy Kylin domestic operating system.

[0081] like Figure 3 , Figure 4 As shown, AI Core Board 2 is a standard packaged general-purpose artificial intelligence inference computing module, requiring no in-house design. Core Board 2 includes Core Board PCB 2-1, NPU chip (neural processing unit) 2-2, and Core Board chipset 2-3. The NPU chip uses the Ascend series (310B, 310P) domestic NPU. Core Board chipset 2-3 includes memory modules, encoding / decoding modules, power modules, and other supporting chips and components. The core board using the 310B NPU has a computing power of 20 TOPS-INT8, and the core board using the 310P NPU has a computing power of 176 TOPS-INT8. Users can choose the computing power as needed to adapt to different application scenarios.

[0082] High-speed interconnection between the AI ​​core board 2 and the VPX carrier board 1 is achieved using a mating interface 3, which can be either an MXM interface or a COME interface. The mating height is ≤6mm (the distance between the BOT side of the AI ​​core board and the TOP side of the VPX carrier board after mating). The plug 3-1 is located on one side of the core board PCB2-1, and the socket 3-2 is located on one side of the carrier board PCB1-1. After mating, the BOT side of the core board contacts the nut posts 1-6 on the TOP side of the carrier board and aligns with the mounting holes, then is secured with mounting screws. The VPX carrier board 1 and the cold plate assembly (heat dissipation assembly) 4 are configured in a series according to the type of core board 2, allowing users to select the appropriate configuration based on their needs.

[0083] The cold plate assembly (heat dissipation assembly) 4 adopts a 5HP high-performance air-cooled design, including a finned cold plate 4-1, a first heat dissipation surface 4-2, a second heat dissipation surface 4-3, a phase change heat conduction element 4-4, a heat conduction block 4-5, a heat conduction boss / groove 4-6, a side vent 4-7, an interface clearance groove 4-8, heat dissipation fins 4-9, and mounting studs 4-10. The finned cold plate 4-1 uses AL6063 high thermal conductivity aluminum alloy as the base material. The top is machined with heat dissipation fins 4-9 to achieve heat dissipation through heat exchange with the air. The bottom is grooved to accommodate the positions of the carrier board 1, the AI ​​core board 2, and their components, forming a first heat dissipation surface 4-2 and a second heat dissipation surface 4-3 at a certain distance from the top surface of the carrier board PCB1-1 and the core board PCB2-1. For low-power heat dissipation devices, thermally conductive protrusions / grooves 4-6 of appropriate sizes are machined on the first heat dissipation surface 4-2 and the second heat dissipation surface 4-3 (grooves are made if the device height exceeds the heat dissipation surface) to conduct heat from the device to the heat dissipation surface and diffuse it into the convective air through the back heat dissipation fins 4-9. For high-power heat dissipation devices such as CPU chips (general-purpose processors) 1-4 and NPU chips (neural processing units) 2-2, grooves are made at corresponding positions on the heat dissipation surface of the finned cold plate 1 and phase change thermal conductive elements 4-4 are laid. These can be heat pipes or VC vapor chambers. Through the phase change heat transfer principle of internal medium evaporation and condensation, the heat from the concentrated chip position is quickly conducted to other positions on the first heat dissipation surface 4-2 and the second heat dissipation surface 4-3, increasing the heat dissipation area and thus improving heat dissipation efficiency. The phase change thermal conductive elements 4-4 are connected to the finned cold plate 1 by welding to reduce contact thermal resistance. For the part of the heat dissipation device that falls below the phase change thermal conductive element and the gap between it and the phase change thermal conductive element is >1mm, a copper thermal conductive block 4-5 is welded to conduct heat from the device to the phase change thermal conductive element. Because the AI ​​core board is stacked and inserted on top of the VPX carrier board, the heat dissipation fins 4-9 in the second heat dissipation surface 4-3 area above the core board are relatively low, resulting in poor heat dissipation. The phase change heat conduction elements 4-4 are not limited to a planar shape; a certain number of three-dimensional phase change heat conduction elements 4-4 are also provided to conduct some of the heat from the NPU chip (neural processing unit) 2-2 on the second heat dissipation surface 4-3 to the first heat dissipation surface 4-2, which has a higher fin height and lower temperature, thereby improving overall heat dissipation performance. The finned cold plate 4-1 has several side ventilation openings 4-7 on both sides, allowing some of the heat dissipation airflow to enter the board's interior along these openings, assisting in the heat dissipation of low-power devices and the PCB. Multiple mounting studs 4-10 are provided around and in the center of the finned cold plate 4-1 to fix the VPX carrier board and core board, and to maintain relatively uniform contact pressure between the cold plate assembly and the PCB, as well as among the heat dissipation devices, preventing PCB bending and excessive / insufficient contact force.

[0084] A small gap (0.5~1mm) exists between the thermally conductive bosses / grooves 4-6 and the thermally conductive blocks 4-5 of the cold plate assembly (heat dissipation assembly) 4 and the heat dissipation device. This gap is used to compensate for component packaging size errors and cold plate processing errors. The gap is filled with thermally conductive interface material 5, which densely fills the assembly gap and small pits between the cold plate and the device, improving the contact area and thermal conductivity. The thermally conductive interface material 5 can be a thermally conductive pad, or it can be TIM graphene, thermally conductive silicone sludge, silicone grease, or other materials.

[0085] The panel assembly 6 is located on the front of the board and includes panel 6-1 and pull-out aid 6-2, enabling external insertion and removal. Panel 6-1 has a slot to avoid the panel interface area 1-3 of the VPX carrier board 1. The back cover 7 is made of AL6063 high thermal conductivity aluminum alloy and is used to shield and protect the components on the back of the VPX carrier board 1 and dissipate heat. It has heat dissipation bosses / grooves machined at the positions of heat dissipation components on the BOT side of the carrier board, and holes are cut out to avoid the positions of replaceable components such as SSD hard drives, so that hard drives can be replaced without disassembling the machine, which is convenient for hardware, data maintenance and system and model upgrades.

[0086] The VPX inference computing motherboard features a series of interlocking interfaces, allowing it to be plugged into core boards with either MXM or COME interfaces. This enables tiered computing power of 20 TOPS-INT8 and 176 TOPS-INT8, which users can select according to their application needs. It can be equipped with different quantized versions of large AI models to perform image, video, and speech recognition, analysis, and generation in offline environments; natural language processing; deep data analysis, screening, and classification; and AI services such as agent-based intelligent decision-making and task scheduling.

[0087] This utility model adopts a standard air-cooled VPX hardware architecture that conforms to the VITA48 specification, which can be quickly adapted and compatible with existing VPX platform equipment, meet the reliability and environmental adaptability requirements of specific fields, and enable rapid upgrades and empowerment of artificial intelligence.

[0088] This invention integrates a general-purpose computing (Phytium series CPU) and a standard AI core inference card (Ascend series NPU) into a single 6U5HP VPX motherboard through a stacking and plugging method. A single slot can enable the VPX system to upgrade and empower artificial intelligence. Compared with the dual-board design of general-purpose computing motherboard + AI inference board and interconnection through VPX backplane, it occupies fewer slots and has lower data transmission latency.

[0089] This invention adopts a composite heat dissipation method of AL6063 aluminum alloy + phase change heat conduction element (heat pipe / VC vapor chamber) or micro pump liquid cooling system. Combined with planar + three-dimensional heat conduction structure design, it breaks through the heat dissipation problem of mother-daughter card under the limited space of 5HP, realizes efficient heat conduction and diffusion, controls the temperature of NPU and CPU within an appropriate range, and meets the stable operation requirements of high power consumption chip.

[0090] The carrier board is compatible with Ascend 310B (20TOPS-INT8 computing power) and 310P (176TOPS-INT8 computing power) core boards through the serialized design of the COME / MXM interface. Users can choose the computing power as needed to adapt to different application scenarios.

[0091] This invention solves the problem of uneven heat dissipation in high-density stacked architectures under limited space, especially the heat dissipation bottleneck, by setting up a heat dissipation component containing three-dimensional heat-conducting elements. Specifically, it achieves this through the following synergistic effect: the heat dissipation component is divided into a first heat dissipation area with superior heat dissipation conditions (corresponding to the CPU) and a second heat dissipation area with limited space (corresponding to the NPU); secondly, the three-dimensional heat-conducting elements (such as pre-bent heat pipes, stepped VC vapor chambers, or three-dimensional liquid cooling pipelines) construct a thermally efficient heat highway that physically bypasses the shielding of the AI ​​core board. Its heat absorption end is tightly thermally coupled to the NPU, utilizing the efficient phase change heat transfer or liquid cooling circulation principle of the internal medium to rapidly, with low thermal resistance, and continuously transport the concentrated heat generated by the NPU to its heat release end, which is located in the first heat dissipation area with higher heat sink fins and better airflow. In military and aerospace applications that require the highest computing power per slot, this design enables the NPU with a power consumption of up to 100 watts to operate stably for a long time within a small 5HP space at a safe temperature without performance throttling. This ensures the computing power output and task reliability of the entire AI inference computing motherboard throughout its entire life cycle, especially under extreme conditions such as high altitude and high temperature. This is something that traditional planar heat dissipation solutions cannot achieve.

[0092] In summary, this application enables the upgrading and expansion of AI computing inference capabilities on the VPX platform at minimal cost, without compromising reliability and environmental adaptability; it allows for functional expansion with minimal space (a single 5HP board), or the replacement of existing general-purpose computing boards with general-purpose computing boards containing AI core boards, without excessively occupying or wasting the limited space resources of the VPX chassis; it solves the problem of insufficient heat dissipation space caused by the stacking and insertion of 5HP download boards and core boards, and achieves high heat flux density heat dissipation under the condition of limited cold plate size.

[0093] This invention effectively solves the problems of heat-conducting interface material pump-out failure and thermal resistance degradation caused by thermo-mechanical cycling through a self-locking thermal interface composite material and its corresponding micro-textured surface. This effect stems from a micro-scale thermo-mechanical adaptive locking mechanism: the micro-texture (such as grooves or honeycomb pits) pre-processed on the contact surface of the heat sink increases the heat-conducting area while providing initial physical anchoring for the heat-conducting substrate; shape memory alloy microfilaments dispersed in the heat-conducting substrate undergo a preset morphological change (such as bending or twisting) when the chip's operating temperature exceeds its phase transition point, generating active mechanical locking force in the grooves of the micro-texture. This temperature-driven micro-anchoring effect effectively resists the shear force generated by the differential thermal expansion between the two interfaces, firmly locking the heat-conducting substrate within the interface. In scenarios where equipment needs to undergo multiple severe temperature shocks from -40°C to +50°C, the interface thermal resistance can be guaranteed to remain stable throughout the entire service life, avoiding chip temperature rise caused by the gradual deterioration of thermal resistance, breaking the vicious cycle of thermal-mechanical coupling failure, and improving the reliability and predictability of the product throughout its entire life cycle.

[0094] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and all such equivalent transformations fall within the protection scope of the present invention.

Claims

1. An AI inference computing motherboard based on the VPX architecture, comprising: A carrier board that conforms to the VPX standard and has a general-purpose processor installed on it; Heat dissipation components used to dissipate heat from the carrier plate; Its characteristic is that it further includes: An AI core board equipped with a neural processing unit is detachably mounted on a carrier board in a stacked manner. The inter-board interface, located between the carrier board and the AI ​​core board, is used to enable data and power transmission between the two. The heat dissipation components are also used to dissipate heat for the AI ​​core board that is stacked on top of the carrier board.

2. The AI ​​inference computing motherboard based on the VPX architecture according to claim 1, characterized in that, The heat dissipation components include: The first heat dissipation area is used to dissipate heat for the general-purpose processor; The second heat dissipation area, located above the AI ​​core board, is used to dissipate heat for the neural processing unit; A thermally conductive element is used to uniformly conduct and diffuse heat from the device along a plane to a first or second heat dissipation area, wherein at least one thermally conductive element is three-dimensional and has a heat absorption end thermally coupled to the neural processing unit and a heat release end. The heat release end bypasses the physical shielding of the AI ​​core board, extends into the first heat dissipation area and is thermally coupled thereto, transferring some of the heat generated by the neural processing unit to the first heat dissipation area for dissipation; The heat dissipation fins are located on the back of the first and second heat dissipation areas, and diffuse heat into the flowing air in the chassis heat dissipation duct.

3. The AI ​​inference computing motherboard based on the VPX architecture according to claim 2, characterized in that, The three-dimensional heat-conducting element is as follows: One or more heat pipes are pre-bent and contain a phase change medium. The bending shape of the heat pipe is adapted to the planar and three-dimensional spatial layout between its heat absorption end and heat release end. Heat can be uniformly conducted in the heat pipe cavity through the phase change medium.

4. The AI ​​inference computing motherboard based on the VPX architecture according to claim 2, characterized in that, Three-dimensional heat-conducting elements can also be specifically: The vacuum cavity heat exchanger has a stepped thickness and contains a phase change medium. The stepped structure of the vacuum cavity heat exchanger allows it to physically overcome the spatial barrier between the AI ​​core board and the carrier board, and heat can be uniformly conducted in the cavity through the phase change medium.

5. The AI ​​inference computing motherboard based on the VPX architecture according to claim 2, characterized in that, Three-dimensional heat-conducting elements can also be specifically: A micro-pump liquid cooling system with circulating flow channels, wherein at least one set of flow channels is spatially connected to the first and second heat dissipation areas, and heat can be uniformly conducted in a planar manner and spatially across the first and second heat dissipation areas by circulating coolant.

6. The AI ​​inference computing motherboard based on the VPX architecture according to claim 1, characterized in that: In the contact surface between the general-purpose processor and / or neural processing unit and the heat dissipation component, the side of the heat dissipation component is processed with a microtexture; Between the contact surfaces of the processor and the heat dissipation components, a self-locking thermal interface composite material is filled. The composite material includes a thermally conductive matrix and shape memory alloy microfilaments dispersed therein.

7. The AI ​​inference computing motherboard based on the VPX architecture according to claim 2, characterized in that, Also includes: At least one heat-conducting element is a planar heat-conducting element.

Images