AI module

By designing AI modules, using AI boards and matrix distribution slots to achieve communication connections between AI boards, the problem of insufficient processing capabilities of existing AI computing boards is solved, and high-performance computing and low-cost production are achieved.

CN119828863APending Publication Date: 2025-04-15DACHUANG PRECISION INTELLIGENT MANUFACTURING (KUNSHAN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411805251.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-12-09
Publication Date
2025-04-15

AI Technical Summary

Technical Problem

In the existing artificial intelligence solutions, the processing power of AI computing boards cannot meet the needs of parallel processing of large data volumes, and the energy efficiency ratio of traditional CPUs is low, ASICs lose flexibility, and GPU expansion cluster costs high.

Method used

Design an AI module to directly interconnect inside and outside the chip to eliminate the needs of memory and switches, and adopt multiple AI boards and a first slot design with matrix distribution to realize matrix layout communication connection between AI boards and cards.

Benefits of technology

It realizes high-performance computing, reduces production costs, improves data transmission efficiency and module scalability, and is suitable for AI application scenarios with high-speed data processing and modular requirements.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119828863A_ABST
    Figure CN119828863A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of computers, and particularly discloses an AI module. The AI module comprises a bottom plate and a plurality of AI board cards, wherein the AI board cards are integrated with signal ports; a plurality of first slots are distributed on the upper plate surface of the bottom plate in a matrix, the number of the first slots is the same as that of the AI board cards, the first slots are in one-to-one correspondence with the AI board cards, the AI board cards are vertically inserted into the first slots, the first slots are in communication connection with the signal ports, each first slot comprises a first interface and a second interface, and in the length direction of the matrix, the second interfaces are in one-to-one correspondence with the first interfaces. The first interface in one first slot is in communication connection with the second interface in another adjacent first slot in the width direction of the matrix, and the second interface in one first slot is in communication connection with the first interface in another adjacent first slot in the width direction of the matrix, so that the adjacent AI board cards are in communication connection. According to the AI board card, the inside and the outside of the chip are directly interconnected, so that the whole chip is in data flow, a memory and a switch do not need to be arranged, and the production cost is reduced while high-performance computing power is guaranteed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

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

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

[0003] Deep learning technology has triggered the rapid development of artificial intelligence applications, leading mankind from the information age to the intelligent age. Deep learning is essentially a machine learning technology that requires powerful hardware computing power to complete complex data processing and operations. For such huge data processing and operations, existing artificial intelligence solutions often use dedicated AI computing boards to perform deep learning operations, but even a single ultra-high-performance AI computing board has a processing power far from meeting the computing requirements, and the image processing capabilities are also insufficient. In traditional technologies, the most versatile CPU (Central Processing Unit) has the lowest energy efficiency ratio and cannot meet the requirements of large-scale parallel processing of large data in AI computing; ASIC (Application Specific Integrated Circuit) solves the problem of energy efficiency, but at the cost of losing flexibility, it also cannot meet complex real-world needs. GPU (Graphics Processing Unit) has strong processing power, but when expanding the GPU cluster, it still needs to cooperate with network cards and layered switches to achieve interconnection, and the cost of this part is very high, close to 50% of the total cost. Summary of the invention

[0004] The purpose of the present invention is to provide an AI module that directly interconnects the inside and outside of the chip so that the entire chip is a data flow, thereby eliminating the need to set up a memory and a switch, thereby ensuring high-performance computing power while reducing production costs.

[0005] To achieve this object, the present invention adopts the following technical solutions:

[0006] An AI module comprises a base plate and a plurality of AI boards; the AI ​​boards are integrated with signal ports; the upper surface of the base plate is provided with a plurality of first slots distributed in a matrix, the first slots are the same in number as the AI ​​boards and correspond one to one, the AI ​​boards are vertically inserted into the first slots, the first slots are communicatively connected with the signal ports, the first slots comprise a first interface and a second interface, in the length direction of the matrix, the first interface in one of the first slots is communicatively connected with the second interface in another adjacent first slot, in the width direction of the matrix, the second interface in one of the first slots is communicatively connected with the first interface in another adjacent first slot, so that the adjacent AI boards are communicatively connected.

[0007] As an optional technical solution for the AI ​​module, a fixing frame is further provided on the base plate, and the fixing frame and the base plate form an installation slot. The number of the installation slots is the same as that of the AI ​​board and the AI ​​board corresponds one to one. The AI ​​board is partially inserted into the installation slot.

[0008] As an optional technical solution for the AI ​​module, the AI ​​board includes a board body and a bracket, the signal port is arranged at the end of the board body, the board body is fixed on the bracket, and the bracket is snap-connected to the fixing frame.

[0009] As an optional technical solution for the AI ​​module, the bracket includes a handle, which is arranged at an end away from the signal port, and both ends of the handle are respectively connected to two sides of the bracket.

[0010] As an optional technical solution for the AI ​​module, an elastic clip is provided on each side of the bracket, and the two elastic clips can move closer to each other under the pulling force of the handle to make the bracket detach from the fixing frame.

[0011] As an optional technical solution for the AI ​​module, the AI ​​board also includes a heat sink module, the bracket is fixedly connected to the heat sink module by a number of spring screws, the bracket and the heat sink module form a accommodating cavity, and the board body is fixedly installed in the accommodating cavity.

[0012] As an optional technical solution for the AI ​​module, the AI ​​board also includes a chip and multiple memory elements, the chip and the memory elements are both arranged on the board body, the chip is connected to the signal port, and the chip is also provided with interfaces, the number of the interfaces is the same as the number of the memory elements and they are connected one-to-one; the memory elements are LPDDR.

[0013] As an optional technical solution for the AI ​​module, the first slot is distributed on the baseboard in the form of 25G NRE 2D mesh, and the first slot is distributed on the baseboard in 4×8 paths.

[0014] As an optional technical solution for the AI ​​module, the signal port and the first slot are connected using a double-density connector.

[0015] As an optional technical solution for the AI ​​module, a second slot is provided at one end of the base plate, a power interface is provided at the other end of the base plate, output interfaces are provided on both sides of the base plate, and the interface of the second slot is PCIe Gen4*16.

[0016] Beneficial effects of the present invention:

[0017] The AI ​​module uses a design of multiple AI boards and a first slot distributed in a matrix on the baseboard to define the specific way of connecting the AI ​​board and the baseboard. The design of connecting the first interface and the second interface in the length and width directions of the matrix realizes the communication connection between the AI ​​boards arranged in a matrix, improving the data transmission efficiency and module scalability. The above-mentioned AI module is reasonably designed, with the characteristics of high efficiency, stability and scalability, which can meet the needs of high-speed data processing and modularization, provide strong hardware support for AI applications, and is suitable for various scenarios requiring high-performance computing. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic diagram of the structure of an AI module provided by an embodiment of the present invention;

[0019] Figure 2 is an exploded diagram of an AI module provided by an embodiment of the present invention;

[0020] Figure 3 is an exploded diagram of an AI board provided by an embodiment of the present invention;

[0021] Figure 4 yes Figure 3 A partial enlarged view of middle A;

[0022] Figure 5 It is a structural schematic diagram of a board main body provided by an embodiment of the present invention;

[0023] Figure 6 is a schematic diagram of a board main body and a base plate provided in an embodiment of the present invention;

[0024] Figure 7 It is a schematic diagram of the topology of PCIe provided in an embodiment of the present invention;

[0025] Figure 8 It is a schematic diagram of the topology of the x8 25G NRE 2D mesh provided in an embodiment of the present invention.

[0026] In the figure:

[0027] 100, AI board; 101, signal port; 1, board body; 11, memory element; 2, bracket; 21, handle; 22, elastic buckle; 3, heat sink module; 4, spring screw;

[0028] 200, bottom plate; 201, first slot; 2011, first interface; 2012, second interface; 202, second slot; 203, output interface; 204, power interface; 210, fixing frame. DETAILED DESCRIPTION

[0029] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] In the description of the present invention, it should be noted that the orientations or positional relationships indicated by the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside", etc. 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, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. Among them, the terms "first position" and "second position" are two different positions, and the first feature "above", "above" and "above" the second feature include the first feature being directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. The first feature "below", "below" and "below" the second feature include the first feature being directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0031] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0032] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0033] like Figures 1 to 8 As shown, this embodiment provides an AI module, including a base plate 200 and a plurality of AI boards 100, the AI ​​boards 100 are integrated with signal ports 101; the upper surface of the base plate 200 is provided with a plurality of first slots 201 distributed in a matrix, the number of the first slots 201 and the number of the AI ​​boards 100 are the same and correspond one to one, the AI ​​boards 100 are vertically inserted in the first slots 201, the first slots 201 are communicatively connected with the signal ports 101, the first slots 201 include a first interface 2011 and a second interface 2012, in the length direction of the matrix, the first interface 2011 in a first slot 201 is communicatively connected with the second interface 2012 in another adjacent first slot 201, in the width direction of the matrix, the second interface 2012 in a first slot 201 is communicatively connected with the first interface 2011 in another adjacent first slot 201, so that the adjacent AI boards 100 are communicatively connected. In this embodiment, the base plate 200 is a UBB board.

[0034] The AI ​​module defines the specific connection between the AI ​​board 100 and the base plate 200 through the design of a plurality of AI boards 100 and a first slot 201 distributed in a matrix on the base plate 200, and utilizes the design of connecting the first interface 2011 and the second interface 2012 in the length direction and the width direction of the matrix to realize the communication connection between the AI ​​boards 100 arranged in a matrix, thereby improving the data transmission efficiency and module scalability. The above-mentioned AI module is reasonably designed, has the characteristics of high efficiency, stability and scalability, can meet the needs of high-speed data processing and modularization, provides strong hardware support for AI applications, and is suitable for various scenarios requiring high-performance computing.

[0035] In this embodiment, a fixing frame 210 is further provided on the base plate 200. The fixing frame 210 and the base plate 200 form a mounting slot. The number of the mounting slots is the same as that of the AI ​​boards 100 and they correspond one to one. The AI ​​boards 100 are partially inserted into the mounting slots.

[0036] The fixing frame 210 on the base plate 200 cooperates with the AI ​​board 100 to ensure the stable installation of the AI ​​board 100 and improve the overall reliability and stability of the AI ​​module.

[0037] Furthermore, the AI ​​board 100 includes a board body 1 and a bracket 2 , the signal port 101 is disposed at the end of the board body 1 , the board body 1 is fixed on the bracket 2 , and the bracket 2 is snap-connected to the fixing frame 210 .

[0038] The design of the AI ​​board 100 includes a board body 1 and a bracket 2, so that the signal port 101 is arranged reasonably. At the same time, the bracket 2 is connected to the fixed frame 210 by a snap, which facilitates the installation and removal of the AI ​​board 100 on the base plate 200 and improves the convenience of AI module maintenance.

[0039] Furthermore, the bracket 2 includes a handle 21 . The handle 21 is disposed at an end away from the signal port 101 , and two ends of the handle 21 are respectively connected to two sides of the bracket 2 .

[0040] The handle 21 on the bracket 2 is designed to facilitate the user to carry and operate the AI ​​module, improve the operating efficiency, and ensure the convenience of use.

[0041] Furthermore, an elastic buckle 22 is respectively provided on both sides of the bracket 2 , and the two elastic buckles 22 can be moved closer to each other under the pulling force of the handle 21 , so that the bracket 2 can be separated from the fixing frame 210 .

[0042] The design of the elastic buckle 22 on the bracket 2 facilitates the quick detachment of the bracket 2 from the fixing frame 210, which is convenient for the replacement and maintenance of the AI ​​board 100, further facilitates the disassembly and repair of the AI ​​module, and improves the maintenance efficiency.

[0043] In this embodiment, the AI ​​board 100 also includes a heat sink module 3, the bracket 2 is fixedly connected to the heat sink module 3 by a plurality of spring screws 4, the bracket 2 and the heat sink module 3 form a receiving cavity, and the board body 1 is fixedly installed in the receiving cavity.

[0044] The design of the heat sink module 3 and the bracket 2 being fixedly connected in the AI ​​board 100 effectively improves the heat dissipation performance of the AI ​​board 100, ensures the stability of the long-term operation of the AI ​​module, and extends the working stability and life of the AI ​​module. In addition, the installation of the board body 1 in the accommodating cavity improves the compactness and safety of the overall structure.

[0045] By providing the heat sink module 3, the heat dissipation effect in the accommodating cavity is guaranteed, and the over-temperature of the board body 1 is avoided, thereby ensuring the long-term stable operation of the board body 1. At the same time, the heat sink module 3 and the bracket 2 provide a space for accommodating the board body 1, achieving the purpose of positioning and protecting the board body 1, and reducing the risk of damage to the board body 1.

[0046] In this embodiment, the AI ​​board 100 also includes a chip and multiple memory elements 11; the chip and the memory elements 11 are both arranged on the board body 1, the chip is connected to the signal port 101, and the chip is also provided with multiple interfaces, each interface is connected to the multiple memory elements 11 one by one.

[0047] Specifically, an interface is set on the chip, and the memory element 11 is directly accessed through the interface, which omits the need for the existing chip to access the memory stick through PCLe or the like, that is, a switch or a network card is saved, thereby saving space and cost.

[0048] Exemplarily, the memory element 11 is LPDDR.

[0049] The memory element 11 adopts the LPDDR design, which improves the data transmission rate and bandwidth, enhances the operating performance of the AI ​​module, improves the data transmission rate and stability, and meets the needs of high-speed data processing.

[0050] Furthermore, the board body 1 also includes a voltage stabilizer and a resistor fuse (Electronic-Fuse, EFuse); the chip is provided with a first integrated circuit interface (inter-integrated circuit, I2C_0), a second integrated circuit interface (inter-integrated circuit, I2C_1) and a third integrated circuit interface (inter-integrated circuit, I2C_2).

[0051] Furthermore, the first integrated circuit interface and the third integrated circuit interface are both connected to the signal port 101, wherein the first integrated circuit interface is used to receive signals transmitted by the UBB board, and the second integrated circuit is used to output signals to the UBB board.

[0052] One end of the resistor fuse is connected to the signal port 101, and the other end of the resistor fuse is connected to the voltage regulator; the voltage regulator is connected to the second integrated circuit interface; the voltage regulator is connected to the LPDDR; the UBB board provides power to the voltage regulator, and the voltage regulator then provides power to the chip and the LPDDR.

[0053] Furthermore, the uplink interface of the signal port 101 adopts PCIe Gen4*8, and the designed maximum power consumption is 300W / 450W.

[0054] Furthermore, the first slot 201 is distributed on the baseboard 200 in the form of 25G NRE 2D mesh; further, the first slot 201 is distributed on the baseboard 200 in 4×8 ways. It should be noted that different numbers of AI cards are arranged on the baseboard 200 in different array forms. This solution uses thirty-two AI cards in 4×8 ways.

[0055] Thirty-two AI boards 100 are arranged and integrated in a 4×8 high-density matrix. Each AI board 100 integrates multiple LPDDRs and is interconnected with adjacent boards through a four-way x8 25G NRE 2D mesh expansion. The above improvements achieve efficient data transmission and processing, as well as interconnection communication between modules, thereby improving the overall performance of the AI ​​module. The design of the first slot 201 distributed on the baseboard 200 in the form of a 4×8 25G NRE 2D mesh can increase the number of AI boards 100, that is, it can expand more AI boards 100 to integrate more LPDDRs, thereby providing higher-speed data transmission and processing capabilities, and realizing high-speed interconnection between each board body 1, so that the board body 1 can adapt to high-load AI computing tasks and improve the overall performance of the board body 1.

[0056] Furthermore, the signal port 101 and the first slot 201 are connected by a double-density connector.

[0057] Furthermore, interfaces are provided around the periphery of the base plate 200 .

[0058] A second slot 202 is disposed at one end of the bottom plate 200 , a power interface 204 is disposed at the other end of the bottom plate 200 , and output interfaces 203 are disposed on both sides of the bottom plate 200 .

[0059] The second slot 202 is used to connect to the CPU; specifically, by setting the second slot 202 to connect to the CPU, the use of wires is reduced, which saves space and reduces costs.

[0060] The interface of the second slot 202 is PCIe Gen4*16.

[0061] The output interface 203 is connected to each optical module interface through a wiring harness.

[0062] In this embodiment, the chip is not the focus of protection of this application, and is a conventional setting in the field. Its specific structure and working principle are common knowledge in this field and are well known to technicians in this field, so no further details will be given here.

[0063] In this embodiment, PCIe topology is as follows Figure 7 As shown, the 2D Mesh / 25G NRE topology is as follows Figure 8 shown.

[0064] Continue to refer Figures 1 to 6 , the chip is TX8 chip.

[0065] The TX8 chip can be used to reconfigure the AI ​​chip and organically combine multiple board bodies 1 in a flexible bus expansion method, thereby expanding the interconnection capabilities between adjacent board bodies 1, enabling the board body 1 to adapt to high-load AI computing tasks, combined with the system's artificial intelligence large model training and reasoning, thereby improving the overall working performance of the board body 1. It has the advantages of high performance and low power consumption, as well as high integration and high stability, can meet the needs of high-intensity computing, and ensure the overall reliability and performance of the board body 1.

[0066] Exemplarily, one of the dual-density connector and the board body 1 is provided with a spring clip seat, and the other is provided with a hook spring clip; the spring clip seat includes two elastic arms arranged opposite to each other, and a movable opening is formed between the two elastic arms. The hook spring clip can be inserted or pulled out relative to the spring clip seat through the movable opening, and the movable opening is elastically expanded when the hook spring clip moves.

[0067] The design of the dual-density connector and the detachable connection method of the hook spring piece facilitate the flexible disassembly and assembly of the hook spring piece structure, thereby improving the stability and reliability of the connection. This design can ensure that the connection between the board body 1 and the base plate 200 is more stable and reliable, while also facilitating user maintenance and upgrades and avoiding interference and errors in data transmission.

[0068] Furthermore, the hook spring is made of plastic.

[0069] The hook spring piece is made of plastic, which is light and durable, ensuring the stability of the connection and having good elasticity, thereby reducing the manufacturing cost and improving durability. The above design makes the hook spring piece easier to install and remove, thereby improving the maintainability and scalability of the board body 1.

[0070] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. An AI module, characterized in that: include: A plurality of AI boards (100), each of the AI ​​boards (100) being integrated with a signal port (101); A base plate (200), wherein the upper plate surface of the base plate (200) is provided with a plurality of first slots (201) distributed in a matrix, the number of the first slots (201) being the same as the number of the AI ​​boards (100) and corresponding to each other, the AI ​​boards (100) being vertically inserted into the first slots (201), the first slots (201) being communicatively connected to the signal port (101), the first slots (201) comprising a first interface (2011) and a second interface (2012), in the length direction of the matrix, the first interface (2011) in one of the first slots (201) is communicatively connected to the second interface (2012) in another adjacent first slot (201), and in the width direction of the matrix, the second interface (2012) in one of the first slots (201) is communicatively connected to the first interface (2011) in another adjacent first slot (201), so that the adjacent AI boards (100) are communicatively connected.

2. The AI ​​module according to claim 1, characterized in that: A fixing frame (210) is also provided on the base plate (200), and the fixing frame (210) and the base plate (200) form a mounting groove, the number of the mounting grooves is the same as that of the AI ​​boards (100) and they correspond one to one, and the AI ​​boards (100) are partially inserted into the mounting grooves.

3. The AI ​​module according to claim 2, characterized in that: The AI ​​board (100) comprises a board body (1) and a bracket (2); the signal port (101) is arranged at the end of the board body (1); the board body (1) is fixed on the bracket (2); and the bracket (2) is snap-connected to the fixing frame (210).

4. The AI ​​module according to claim 3, characterized in that: The bracket (2) comprises a handle (21), the handle (21) is arranged at an end away from the signal port (101), and the two ends of the handle (21) are respectively connected to two sides of the bracket (2).

5. The AI ​​module according to claim 4, characterized in that: An elastic buckle (22) is respectively provided on both sides of the bracket (2), and the two elastic buckles (22) can be moved closer to each other under the pulling force of the handle (21) so that the bracket (2) can be separated from the fixing frame (210).

6. The AI ​​module according to claim 3, characterized in that: The AI ​​board (100) further comprises a heat sink module (3); the bracket (2) is fixedly connected to the heat sink module (3) via a plurality of spring screws (4); the bracket (2) and the heat sink module (3) form a receiving cavity; and the board body (1) is fixedly installed in the receiving cavity.

7. The AI ​​module according to claim 3, characterized in that: The AI ​​board (100) further comprises a chip and a plurality of memory elements (11), wherein the chip and the memory elements (11) are both arranged on the board body (1), the chip is connected to the signal port (101), and the chip is further provided with interfaces, wherein the number of the interfaces is the same as that of the memory elements (11) and they are connected in a one-to-one correspondence; the memory elements (11) are LPDDR.

8. The AI ​​module according to claim 7, characterized in that: The first slots (201) are distributed on the baseboard (200) in the form of 25G NRE 2Dmesh, and the first slots (201) are distributed on the baseboard (200) in a 4×8 manner.

9. The AI ​​module according to claim 8, characterized in that: The signal port (101) and the first slot (201) are connected by a double-density connector.

10. The AI ​​module according to claim 1, characterized in that: A second slot (202) is provided at one end of the base plate (200), a power interface (204) is provided at the other end of the base plate (200), output interfaces (203) are provided on both sides of the base plate (200), and the interface of the second slot (202) is PCIe Gen4*16.