High-performance AI computing power card

Abstract

The utility model belongs to the technical field of AI acceleration modules, and particularly relates to a high-performance AI computing power card, which comprises a PCB (printed circuit board), a processor, a power management module, a memory and an interface module used for being connected with external equipment are arranged on the PCB, the interface module comprises an M.2 M Key interface and a PCIe Gen 3 * 4 channel interface, the processor adopts an AI processor with the model of Hailo-8, and the interface module adopts an M.2M Key interface and a PCIe Gen 3 * 4 channel interface. The high-performance AI computing power card adopts an M.2 M Key interface form, constructs core computing power depending on a Hailo-8TMAI processor, has strong computing power, is provided with a complete PCIe Gen 3 * 4 channel interface, can stably operate in a wide working temperature range of-40 DEG C to 85 DEG C so as to adapt to various complex environments, can also achieve high-throughput data transmission, and has wide application prospects. The real-time and low-delay neural network reasoning task is powerfully supported, in an artificial intelligence application scene, the calculation efficiency and the response speed are remarkably improved, meanwhile, the development difficulty is greatly reduced, and a solid guarantee is provided for rapidly pushing products to the market.

Description

TECHNICAL FIELD

[0001] The utility model belongs to the field of edge computing artificial intelligence application, specifically relates to a kind of high-performance AI computing power card. BACKGROUND

[0002] Artificial intelligence (Artificial Intelligence) English abbreviation is AI, in the development process of AI related fields such as machine learning, deep learning, computer vision, hardware equipment plays a vital role.AI acceleration module (such as computing power card) is one of the key hardware equipment, its essence is the module that can accelerate the calculation of AI algorithm.

[0003] AI acceleration module is a kind of hardware accelerator, compared with the traditional calculation mode of only relying on CPU or GPU, has obvious advantages.The traditional calculation mode often faces the problems of low computing efficiency and high power consumption when processing large-scale AI computing tasks.AI acceleration module can realize efficient AI calculation under lower power consumption through special hardware architecture and algorithm optimization.In many AI application scenarios, such as autonomous driving, intelligent security, industrial automation, there are higher requirements for real-time performance and low latency.AI acceleration module can quickly process input data and perform neural network inference to meet the real-time requirements of these scenarios.

[0004] For many enterprises and research and development teams, the ability in hardware design and underlying optimization is limited.If AI acceleration module is designed from scratch, technical problems are often encountered, which leads to the extension of project cycle and even failure.Using mature AI acceleration module, such as a high-performance AI computing power card mentioned in this paper, can reduce the difficulty of research and development and accelerate product launch.

[0005] In the prior art, some AI acceleration modules have some deficiencies, for example, the interface design of some modules is not reasonable, which leads to poor compatibility when connecting with other devices;Or there are defects in the heat dissipation design, and overheating phenomenon is easy to occur during high-load operation, which affects performance and stability. UTILITY MODEL CONTENTS

[0006] The utility model aims at providing a kind of high-performance AI computing power card, solves the technical problems of poor interface compatibility and heat dissipation design defects of part of AI acceleration module in prior art.

[0007] To solve the above technical problems, the utility model adopts the following technical solutions:

[0008] The utility model provides a high -performance AI algorithm power card, including PCB board, which is provided with processor, power management module and memory on the PCB board, and interface module for connecting with external device, the interface module includes M.2M Key interface and PCIe Gen 3x4 channel interface, the pin corresponding to multiple functions is defined on the M.2M Key interface, can realize the function including high -speed data transmission, equipment identification, the processor adopts the AI processor of model Hailo -8.

[0009] Preferably, the processor, power management module and memory are centrally arranged in the middle region of the PCB board, and the interface module is located at the edge of the PCB board.

[0010] Preferably, the M.2M Key interface and the PCIe Gen 3x4 channel interface are respectively designed with independent signal transmission lines to reduce signal interference.

[0011] Preferably, the interface module further comprises a debugging interface for hardware debugging to facilitate development and troubleshooting.

[0012] Preferably, the memory comprises a cache module and a large-capacity storage module to meet different data processing requirements.

[0013] Compared with the prior art, the high-performance AI algorithm power card has the following advantages: the high-performance AI algorithm power card includes a PCB board, which is provided with a processor, a power management module and a memory, and an interface module for connecting with external devices, the interface module includes a M.2M Key interface and a PCIe Gen 3x4 channel interface, the M.2M Key interface is defined with pins corresponding to multiple functions, which can realize functions including high-speed data transmission and device identification, and the processor adopts an AI processor with model Hailo-8. The high-performance AI algorithm power card adopts a M.2M Key interface form, relies on an Hailo-8 TM AI processor to build core computing power, has a powerful algorithm of 26Tops, and is equipped with a complete PCIe Gen 3x4 channel interface. This design not only can run stably in a wide working temperature range of -40° to 85℃ to adapt to various complex environments, but also can achieve high-throughput data transmission, effectively supporting real-time and low-latency neural network inference tasks. In the artificial intelligence application scenario, it significantly improves the computing efficiency and response speed, and greatly reduces the development difficulty, providing a solid guarantee for quickly pushing products to the market. The high-performance AI algorithm power card has excellent developability and can meet the development needs of users in multiple scenarios and the functional expansion needs in specific scenarios. For example, the high-performance AI algorithm power card can be used as a key algorithm module of an automatic driving system, an intelligent security monitoring system, etc. BRIEF DESCRIPTION OF DRAWINGS

[0014] The accompanying drawings are included to provide a further understanding of the present application, and are incorporated in and constitute a part of the specification, illustrate embodiments of the present application and serve to explain the present application, and do not constitute a limitation of the present application. In the drawings:

[0015] Figure 1 is a front and back structure schematic diagram of an embodiment of the high-performance AI computing card of the present application, wherein Figure 1 The left side is a front structure schematic diagram, and the right side is a back structure schematic diagram.

[0016] Figure 2 is a mechanical dimension drawing of an embodiment of the high-performance AI computing card of the present application.

[0017] Figure 3 is a system block diagram of an embodiment of the high-performance AI computing card of the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0019] In one embodiment, a high-performance AI computing card is provided, please refer to Figures 1 to 3 .

[0020] As shown in Figures 1-3 , the high-performance AI computing card includes a PCB board 1, and a processor, a power management module and a memory are arranged on the PCB board 1, wherein the processor adopts a processor with a model of Hailo-8, and the memory includes a cache module and a large-capacity storage module to meet different data processing requirements. Specifically, the cache module in the memory is a memory module, and the large-capacity storage module is an external storage module. The memory module adopts a high-performance LPDDR4 memory chip to provide fast data read-write support for the processor, so as to meet the efficient processing requirements of data in AI calculation. The external storage module is an eMMC flash memory, which is used to store system software, AI models and related data, etc., to ensure that the computing card can quickly call the required data in different application scenarios. As shown in Figures 1-3As shown, the LPDDR4 memory chip and the eMMC flash are respectively electrically connected to the processor through internal circuits of the PCB, and the power management module provides stable working power for the LPDDR4 memory chip, the eMMC flash and the processor. The processor can read and write data from the LPDDR4 memory chip and the eMMC flash, and the PCB is also provided with peripheral circuits such as clock circuits, which guarantee the normal operation of the Hailo-8 TM processor, the power management module, the LPDDR4 memory chip and the eMMC flash can work normally.

[0021] In Figure 3 , 25MHz OSC is the clock signal input to the processor by the peripheral clock circuit, Hailo-8 is the processor model, PCIe 3.0 8GT / s x 4 refers to the third generation PCIe communication technology, M.2Connector M Key is an interface for supporting PCIe 4X transmission mode, TPs are integrated ports, which integrate JTAG interface and UART interface, I2C is the interface of I2C protocol, 64kb EEPROM memory is connected to the I2C interface, Overcurrent Indicator is the current overload detection module, M.2Connector M Key interface provides 3.3V voltage for it, and outputs 0.8V and 1.8V DC voltage to supply other components, that is, 0.8v Supply and 1.8v Supply.

[0022] As Figures 1-3 shown, the high-performance AI computing card also includes an interface module for connecting with external devices, the processor is arranged at the center of the PCB, the power management module and the memory are arranged at the side edge of the PCB, and the interface module is arranged at the other side edge of the PCB. This layout is reasonable, which facilitates the connection of the high-performance AI computing card with external devices and reduces signal interference. The interface module of the high-performance AI computing card includes an M.2 M Key interface (M.2Connector M Key) and a PCIe Gen 3x4 channel interface (PCIe 3.0 8GT / s x 4). The M.2 M Key interface adopts a high-quality connector to ensure the stability of the connection and the reliability of the data transmission. The M.2 M Key interface defines a plurality of pins corresponding to data transmission and function implementation. The function pins defined on the M.2 M Key interface are used to implement a plurality of different functions, such as high-speed data transmission, device identification, etc. In this embodiment, the high-performance AI computing card has 75 pins, that is, 75 pins. The interface module also defines a debugging interface for hardware debugging, which is convenient for development and troubleshooting. The debugging interface here is an I2C interface.

[0023] As Figure 2 shown, inFigure 2 In particular, the width of the PCB is 22 mm, the total length is 80 mm, the PCB is divided into two parts, the left part is 42 mm long, and the right part is 38 mm long, the electronic components such as the processor and the memory are arranged in the left part, the contact array 3 is arranged in the right part, the gold finger 2 is arranged at the left end of the PCB, and the contact array 3 and the gold finger 2 are independent signal transmission lines. The M.2M Key interface and the PCIe Gen 3x4 channel interface can be independently arranged, so that the M.2M Key interface and the PCIe Gen 3x4 channel interface are designed by independent signal transmission lines respectively, so as to reduce signal interference.

[0024] The data interface is defined based on the corresponding industry standard and application requirement, each interface is composed of a specific number of pins, each pin has its unique role in an interface, and these data interfaces include but are not limited to the following and their functions: the M.2M Key interface is used to connect external storage devices or other expansion modules, such as connecting a large-capacity solid-state disk through this interface to meet the application scenarios with higher requirements for data storage capacity and read-write speed. The PCIe Gen 3x4 channel interface is a high-speed data transmission channel, which can realize high-throughput data interaction between the host or other devices, such as connecting the AI computing card with the PCIe slot of the server to quickly transmit a large amount of image, video and other data for AI inference operation.

[0025] Specifically, the correspondence between the terminals of the above interfaces and the PCIe interface line sequence of the high-performance AI computing card is shown in Table 1 below. In Table 1, the numbers from 1 to 75 in the PIN Number(s) column represent the pin numbers on the PCIe interface, and the middle column Signal represents the signal name. For example, “CONFIG 0(GND)”, “CONFIG_1(NC)”, etc. represent the signal type or function corresponding to the pin. The last column Type represents the signal type, i.e. whether the pin is an input (I), output (O) or input / output (I / O) type.

[0026] The functions of all pins of the AI computing card are specified as “default function” in Table 1 below. For example, in Table 1, the number 21 represents the pin number 21 on the PCIe interface; “GND” in the middle column represents ground, and “NC” represents no connection; the type of the “PERST#” signal in the Type column is “I”, indicating that it is an input signal; the type of the “PET[n / p]0” is “O”, indicating that it is an output signal.

[0027] Table 1: PCIe interface line sequence description table

[0028]

[0029]

[0030] The high-performance AI computing card adopts an M.2 M Key interface form, relies on Hailo-8 TM AI processor to build core operation ability, has 26Tops strong computing power, and is equipped with complete PCIe Gen 3 x4 channel interface.This design not only can be stably operated in the wide working temperature range of-40° to 85℃ to adapt to various complex environments, but also can reach high throughput data transmission, effectively supports real-time and low-delay neural network inference tasks, in artificial intelligence application scene, it significantly improves the computing efficiency and response speed, and greatly reduces the development difficulty, provides solid guarantee for the product to be quickly pushed to the market.

[0031] The high-performance AI computing card in the embodiment has excellent developable performance, can meet the development requirements of users in multiple scenes and the functional expansion requirements in specific scenes, for example, the high-performance AI computing card can be used as a key computing module of an automatic driving system, an intelligent security monitoring system and the like.

[0032] Although the embodiments of the utility model have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirits of the utility model, the scope of the utility model is defined by the appended claims and their equivalents.

Claims

1. A high-performance AI computing card, characterized in that: The application relates to a PCB board provided with a processor, a power management module and a memory, and an interface module for connecting with external equipment, wherein the interface module comprises an M.2 M Key interface and a PCIe Gen 3x4 channel interface, the M.2 M Key interface is defined with pins corresponding to multiple functions, and the functions including high-speed data transmission and equipment identification can be realized, and the processor adopts an AI processor with a model number of Hailo-8.

2. The high-performance AI computing card of claim 1, wherein: The processor, the power management module and the memory are centrally arranged in a middle area of the PCB board, and the interface module is located at an edge of the PCB board.

3. The high-performance AI computing card of claim 1, wherein: The M.2 M Key interface and the PCIe Gen 3x4 channel interface are respectively designed with independent signal transmission lines to reduce signal interference.

4. The high-performance AI computing card of claim 2, wherein: The interface module further comprises a debugging interface for hardware debugging, which is convenient for development and fault troubleshooting.

5. The high-performance AI computing card of claim 2, wherein: The memory comprises a cache module and a large-capacity storage module to meet different data processing requirements.

Images