Astronavigation-level artificial intelligence computing platform based on YuLong810A

Abstract

The invention relates to an astronavigation-level artificial intelligence computing platform based on YuLong810A. The astronavigation-level artificial intelligence computing platform comprises a DDR3 / 4 high-speed cache module for caching high-capacity computing data, an NAND FLASH high-capacity storage module for storing a user application program, an NOR FLASH fast executable storage module for storing a key program, a system clock module for performing clock driving on a functional module, a reset module for completing system reset, and a data storage module for storing data. The system power management unit is used for realizing reliable supply and management of a system power supply; the high-speed bus network is used for realizing high-speed data communication; the low-speed bus network is used for realizing low-speed data communication; and the computing power platform is used for realizing AI computing power support and computing power improvement of an aerospace-level artificial intelligence computing platform. A high-performance and low-power-consumption satellite-borne intelligent computing platform supporting mass data on-orbit processing and a matched intelligent algorithm are provided, and computing power support is provided for space intelligent sensing capability generation; a modularized, universal and extensible design framework is adopted, the system can be reused in spacecrafts of different models, the development cost is saved, and the development period is shortened.

Description

Technical Field

[0001] The present invention relates to the field of aerospace artificial intelligence technology. More specifically, the present invention relates to an aerospace-grade artificial intelligence computing platform based on YuLong810A. Background Art

[0002] In terms of satellite payload data, with the development of satellite payload technology, the output data volume of various high-resolution detection instruments has increased significantly. At present, the on-orbit information processing ability of the space-based information system is weak, and its operation mode is still to transmit data to the ground station for processing through data relay satellites or when the satellite passes overhead, and then redistribute it. The rapid expansion of remote sensing data volume causes huge pressure on data transmission and data processing; there are many invalid data on the satellite, wasting on-board storage resources and downlink channel resources; the data delivery cycle under manual processing is long, making it difficult to meet the timeliness requirements of products. Therefore, there is an urgent need to process remote sensing data on orbit, extract effective information and downlink it to improve the resource utilization rate of the satellite system. Summary of the Invention

[0003] Embodiments of the present application provide an aerospace-grade artificial intelligence computing platform based on YuLong810A. To provide a basic understanding of some aspects of the disclosed embodiments, a simple summary is given below. This summary part is not a general review, nor is it intended to identify key / important constituent elements or delineate the protection scope of these embodiments. Its sole purpose is to present some concepts in a simple form as a preface to the subsequent detailed description.

[0004] The present application provides an aerospace-grade artificial intelligence computing platform based on YuLong810A, which includes: DDR3 / 4 cache module, composed of a 72-bit DDR buffer memory VD3D16G72RB199SS2WH with ECC function and DDR interface signals, for caching large-capacity computing data of the aerospace-grade artificial intelligence computing platform; NAND FLASH large-capacity storage module, composed of an 8-bit NAND FLASH large-capacity memory VDNF32G08RS50MS4V25-II and NF interface signals, for storing user application programs of the aerospace-grade artificial intelligence computing platform; NOR FLASH fast-executable storage module, composed of a 16-bit NOR FLASH fast memory JFM29GL256RH and EMIF interface signals, for storing the BootLoader key program of the aerospace-grade artificial intelligence computing platform; The system clock module consists of a 24MHz active crystal oscillator, a 32.768MHz active crystal oscillator or a passive crystal, a 16MHz active crystal oscillator, a 25MHz active crystal oscillator, and a 62.5MH high-speed differential clock, and is used to drive the clocks of the functional modules of the aerospace-grade artificial intelligence computing platform; The reset module consists of JSR706RD and a system reset signal, and is used to complete the system reset of the aerospace-grade artificial intelligence computing platform; The system power management unit consists of RSHF4644ARH, RSW3301HRH, RSW1101HRH, and various power rail signals, and is used to realize the reliable supply and management of the system power of the aerospace-grade artificial intelligence computing platform; The high-speed bus network consists of SRIO, PCI-E, a 10 / 100 / 1000M adaptive network interface, and high-speed related signals, and is used to realize the high-speed data communication of the aerospace-grade artificial intelligence computing platform; The low-speed bus network consists of SPI, IIC, UART, a 1553B interface, and low-speed related signals, and is used to realize the low-speed data communication of the aerospace-grade artificial intelligence computing platform; The computing power platform consists of one or more aerospace-grade SOC embedded chips YuLong810A with AI computing capabilities, and is used to realize the AI computing power support and computing power improvement of the aerospace-grade artificial intelligence computing platform.

[0005] According to a preferred embodiment, the DDR3 / 4 high-speed cache module, the NAND FLASH large-capacity storage module, the NOR FLASH fast-executable storage module, the system clock module, the reset module, the system power management unit, the high-speed bus network, the low-speed bus network, and the computing power platform all use a custom-form bus interface and a high-speed connector HSIH4-160ZKV1-02B of aerospace electric appliances to effectively connect the interface module and the substrate.

[0006] According to a preferred embodiment, the computing power platform adopts a standardized interface.

[0007] According to a preferred embodiment, the computing power platform includes a GPU unit, a standard shader core, supports EVIS extended instructions, and supports 16 / 32 / 64-bit floating-point operations; the peak computing power is 64GFLOPS.

[0008] According to a preferred embodiment, the computing power platform includes a neural network acceleration unit, which is used to realize the platform acceleration ability of the aerospace-grade artificial intelligence computing platform.

[0009] According to a preferred embodiment, the neural network acceleration unit has matrix parallel convolution MACs, supports compression and pruning of multi-dimensional arrays for neural network processing, supports 8 / 16-bit fixed-point processing, and has a peak computing power of 12 TOPS.

[0010] According to a preferred embodiment, the computing power platform is used to improve the computing power of the aerospace-grade artificial intelligence computing platform by expanding the number of aerospace-grade SOC embedded chips YuLong810A.

[0011] The technical solution provided by this application may include the following beneficial effects: In this application, the aerospace-grade artificial intelligence computing platform based on YuLong810A uses a DDR3 / 4 cache module composed of a 72-bit DDR buffer memory VD3D16G72RB199SS2WH with ECC function and DDR interface signals to cache a large amount of computing data of the aerospace-grade artificial intelligence computing platform; a NAND Flash mass storage module composed of an 8-bit NAND Flash mass memory VDNF32G08RS50MS4V25-II and NF interface signals stores the user application programs of the aerospace-grade artificial intelligence computing platform; a NOR Flash fast executable storage module composed of a 16-bit NOR Flash fast memory JFM29GL256RH and EMIF interface signals stores the BootLoader key program of the aerospace-grade artificial intelligence computing platform; a system clock module composed of a 24MHz active crystal oscillator, a 32.768MHz active crystal oscillator or a passive crystal, a 16MHz active crystal oscillator, a 25MHz active crystal oscillator, and a 62.5MH high-speed differential clock drives the clock of the functional modules of the aerospace-grade artificial intelligence computing platform; a reset module composed of JSR706RD and a system reset signal completes the system reset of the aerospace-grade artificial intelligence computing platform; a system power management unit composed of RSHF4644ARH, RSW3301HRH, RSW1101HRH, and various power rail signals realizes reliable power supply and management of the system power of the aerospace-grade artificial intelligence computing platform; a high-speed bus network composed of SRIO, PCI-E, 10 / 100 / 1000M adaptive network interfaces, and high-speed related signals realizes high-speed data communication of the aerospace-grade artificial intelligence computing platform; a low-speed bus network composed of SPI, IIC, UART, 1553B interfaces, and low-speed related signals realizes low-speed data communication of the aerospace-grade artificial intelligence computing platform; a computing power platform composed of one or more aerospace-grade SOC embedded chips YuLong810A with AI computing capabilities is used to realize the AI computing power support and computing power improvement of the aerospace-grade artificial intelligence computing platform. The main purpose of this application is to provide a high-performance, low-power spaceborne artificial intelligence computing platform and supporting intelligent algorithms for supporting on-orbit processing of massive data, provide basic computing power support for the generation of space intelligent perception capabilities, and also provide a reference for the design of algorithm verification platforms in other high-requirement fields, and better serve the field of aerospace artificial intelligence technology. In terms of aerospace payload data in this application, the amount of remote sensing data is rapidly expanding. Through the artificial intelligence computing platform, relevant data can be processed on orbit, reducing the pressure on data transmission and data processing; the AI processing of invalid data on the satellite can be carried out through the artificial intelligence computing platform, making full use of the on-board storage resources and downlink channel resources; because the data is processed on orbit and the final useful results are fed back to the ground, there is no need for manual data processing, which can meet the timeliness requirements of products, improve the utilization rate of satellite system resources, and reduce the demand for data downlink bandwidth.In terms of aerospace intelligent processing algorithms, this application also provides a GPU unit and a standard shadercore, which support EVIS extended instructions and 16 / 32 / 64-bit floating-point operations; the peak computing power is 64 GFLOPS. The neural network acceleration unit has a matrix parallel convolution MAC; it supports compression and pruning of neural network multi-dimensional arrays; it supports 8 / 16-bit fixed-point processing; the peak computing power is 12 TOPS. The computing power can be improved by expanding the number of YuLong810A, ensuring the overall efficiency, reliability, and flexibility of the terminal artificial intelligence system. This application adopts a modular, generalized, and scalable platform design architecture, which can be reused in different models of spacecraft, greatly saving the development cost and shortening the development cycle.

[0012] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The drawings herein are incorporated into the specification and form a part of the specification, showing embodiments consistent with the present invention and used together with the specification to explain the principles of the present invention.

[0014] Figure 1 It is a schematic diagram of a single YuLong810A aerospace-grade artificial intelligence computing platform based on YuLong810A provided by an embodiment of this application; Figure 2 It is a schematic diagram of a dual YuLong810A aerospace-grade artificial intelligence computing platform based on YuLong810A provided by an embodiment of this application; Figure 3 It is a schematic diagram of the power supply timing of a single YuLong810A aerospace-grade artificial intelligence computing platform based on YuLong810A provided by an embodiment of this application; Figure 4 It is a schematic diagram of the storage part function of a single YuLong810A aerospace-grade artificial intelligence computing platform based on YuLong810A provided by an embodiment of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0015] The following description and the drawings fully illustrate the specific embodiments of the present invention, enabling those skilled in the art to practice them.

[0016] It should be clear that the described embodiments are only a part of the embodiments of the present invention, not all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0017] When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. On the contrary, they are merely examples of systems and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0018] In the description of the present invention, it should be understood that the terms "first", "second", etc. are only used for descriptive purposes and cannot be construed as indicating or implying relative importance. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. In addition, in the description of the present invention, unless otherwise specified, "a plurality" means two or more. "And / or" describes the association relationship of associated objects and indicates that three relationships can exist. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally represents an "or" relationship between the associated objects before and after.

[0019] The following will be combined with the attached Figure 1 - attached Figure 4 , to provide a detailed introduction to an aerospace-grade artificial intelligence computing platform based on YuLong810A provided by the embodiments of the present application.

[0020] Please refer to Figures 1 - 4 , an aerospace-grade artificial intelligence computing platform based on YuLong810A in the embodiments of the present application may include: In terms of on-orbit intelligent processing algorithms, deep neural networks have two characteristics: feature learning and deep structure, which are beneficial to improving the classification and recognition accuracy of remote sensing images. The classification accuracy of common networks has reached a very high level at present, even exceeding the 96.4% level that humans can achieve. Although artificial intelligence technology based on deep neural networks has extremely high application value in the processing of space-based information in complex scenarios, deep neural network models are huge in scale and often have the characteristics of computing-intensive and memory-intensive. Directly applying existing deep learning methods to on-orbit information processing cannot meet the real-time requirements of tasks. As can be seen from the data in Table 1, the number of single-frame forward inference operations of common networks has reached the level of billions, and the number of weights is above the level of millions. Therefore, it is necessary to adopt a software-hardware co-optimization method to optimize and compress the network from the software dimension, reduce the calculation amount and storage amount of the deep neural network model itself while ensuring the inference accuracy, and at the same time select a suitable hardware platform from the hardware dimension, conduct targeted high-energy efficiency application domain customized structure design, provide convenient deployment support from the upper-layer algorithm to the underlying hardware on the compilation tool chain, and provide support for fault tolerance and reconfigurability functions at the system-level software level, so as to ensure the overall high efficiency, reliability, and flexibility of the terminal artificial intelligence system.

[0021] Table 1 Accuracy and Model Scale of Common Classification Neural Network Models model Top-5 accuracy Number of forward inference operations Number of parameters (M) AlexNet 83.6% 1.44GOPs 61 VGG-16 92.6% 31.0GOPs 138 GoogLeNet 93.3% 2.86GOPs 7 ResNet-152 96.4% 22.6GOPs 60.2 In summary, it is necessary to develop a high-performance, low-power spaceborne intelligent heterogeneous computing platform and supporting intelligent algorithms that can support on-orbit processing of massive data, so as to provide basic computing power support for the generation of space intelligent perception capabilities. Therefore, the spaceborne artificial intelligence computing platform based on YuLong810A described in this application is of great significance for improving the work efficiency of astronauts. The platform can be divided into, for example Figure 1 the spaceborne artificial intelligence computing platform with a single YuLong810A as shown in Figure 2 and the spaceborne artificial intelligence computing platform with two YuLong810As as shown in Figure 1 More computing power can be achieved by expanding the number of spaceborne SOC embedded chips YuLong810A; the above platforms all adopt standardized interfaces and can be reused between different models of spacecraft, reducing the development cost and improving the product reliability. Among them Figure 2 the spaceborne artificial intelligence computing platform with two YuLong810As consists of two CPUs, namely YuLong810A0 module and YuLong810A1 module, to form a standard platform, and relevant signals are led out through a high-speed connector; both YuLong810A0 module and YuLong810A1 module consist of a minimum system composed of a NAND Flash large-capacity memory, a DDR3 buffer memory, and a Nor FLASH fast memory, and expand low-speed bus interfaces such as 1553B, USB, IIS, and IIC, and high-speed bus interfaces such as MIPI, PCIE, and SRIO. The YuLong810A0 module and the YuLong810A1 module monitor each other's status through the UART interface.

[0022] Both the YuLong810A0 module and the YuLong810A1 module are spaceborne SOC embedded chips YuLong810A.

[0023] The aerospace-grade artificial intelligence computing platform based on YuLong810A described in this application adopts a modular design concept and integrates high-performance computing units, data storage units, FPGA interface management units, system power management units, algorithm deployment and testing frameworks, and data analysis and visualization systems. Through the construction of the platform, a comprehensive verification of artificial intelligence algorithms under extreme conditions is achieved. At the same time, the platform supports the rapid deployment and iterative testing of multiple algorithms, accelerating the transformation process of algorithms from research and development to application. It not only provides strong support for the development of artificial intelligence algorithms in the aerospace field but also provides a reference for the design of algorithm verification platforms in other high-demand fields. The aerospace-grade artificial intelligence computing platform based on YuLong810A includes: DDR3 / 4 cache modules, NAND FLASH mass storage modules, NOR FLASH fast executable storage modules, system clock modules, reset modules, and system power management units. At the same time, it also provides high-speed bus networks, low-speed bus networks, and computing power platforms.

[0024] The high-performance computing units, algorithm deployment and testing frameworks, and data analysis and visualization systems correspond to the computing power platforms, system clock modules, reset modules, high-speed bus networks, and low-speed bus networks of this application; the data storage units correspond to the DDR3 / 4 cache modules, NAND FLASH mass storage modules, and NOR FLASH fast executable storage modules of this application; the FPGA interface management unit corresponds to the effective connection between the bus interface, interface module, and substrate of this application.

[0025] The DDR3 / 4 cache module consists of a 72-bit DDR buffer memory VD3D16G72RB199SS2WH with ECC function and DDR interface signals, and is used to quickly cache the large-capacity computing data of the aerospace-grade artificial intelligence computing platform using a DDR3 controller.

[0026] The NAND FLASH mass storage module consists of an 8-bit NAND FLASH mass memory VDNF32G08RS50MS4V25-II and NF interface signals, and is used to store the user application programs of the aerospace-grade artificial intelligence computing platform using a NAND controller.

[0027] The NOR FLASH fast executable storage module consists of a 16-bit NOR FLASH fast memory JFM29GL256RH and EMIF interface signals, and is used to quickly and reliably store the key BootLoader program of the aerospace-grade artificial intelligence computing platform using an EMIF controller.

[0028] Figure 4The memory in it includes: a 72-bit DDR buffer memory VD3D16G72RB199SS2WH with ECC function, an 8-bit NAND FLASH mass memory VDNF32G08RS50MS4V25-II, and a 16-bit NOR FLASH fast memory JFM29GL256RH; the chip is an aerospace-grade SOC embedded chip YuLong810A. Figure 4 The main control signals between the separate YuLong810A0 module and the memory are composed of signals such as #CE[3:0], CLE, ALE, #WE[3:0], #RE[3:0], #WP, #RB[3:0], DQ[7:0], etc., which are chip select, read-write, write protection and other signals, to realize the write and read control of the memory.

[0029] The system clock module consists of a 24MHz active crystal oscillator, a 32.768MHz active crystal oscillator or a passive crystal, a 16MHz active crystal oscillator, a 25MHz active crystal oscillator, and a 62.5MH high-speed differential clock, and is used to clock-drive each functional module of the system of the aerospace-grade artificial intelligence computing platform.

[0030] The reset module consists of JSR706RD and a system reset signal, and is used to complete the reliable reset of the system of the aerospace-grade artificial intelligence computing platform.

[0031] The system power management unit consists of RSHF4644ARH, RSW3301HRH, RSW1101HRH and each power rail signal, and is used to realize the reliable power supply and management of the system power of the aerospace-grade artificial intelligence computing platform.

[0032] The high-speed bus network consists of SRIO, PCI-E, a 10 / 100 / 1000M adaptive network interface and high-speed related signals, and is used to realize the high-speed data communication of the aerospace-grade artificial intelligence computing platform.

[0033] The low-speed bus network consists of SPI, IIC, UART, a 1553B interface and low-speed related signals, and is used to realize the low-speed data communication of the aerospace-grade artificial intelligence computing platform.

[0034] The computing power platform consists of one or more aerospace-grade SOC embedded chips YuLong810A with AI computing power, and is used to realize the AI computing power support and computing power improvement of the aerospace-grade artificial intelligence computing platform.

[0035] In the embodiments of the present application, the main controller of the computing power platform selects YuLong810A, the first domestic aerospace-grade SOC embedded chip with AI computing capabilities, which serves as the computing center. Currently, there is no artificial intelligence chip on spacecraft. By introducing this platform, the gap in the aerospace artificial intelligence platform is filled, and AI computing power support and computing power improvement are achieved.

[0036] The different numbers of aerospace-grade SOC embedded chips YuLong810A determine the different computing powers of the computing power platform. According to different computing powers, it can be divided into an aerospace artificial intelligence computing platform with a single YuLong810A and an aerospace artificial intelligence computing platform with two YuLong810A.

[0037] According to a preferred embodiment, the computing power platform adopts a standardized interface; in the embodiments of the present application, the computing power is verified through artificial intelligence algorithms. The computing power platform is equipped with a neural network acceleration unit, which is represented by NNA and is used to achieve the platform acceleration ability of the aerospace artificial intelligence computing platform.

[0038] The present application provides a GPU unit and a standard shader core inside the computing power platform, which supports EVIS extended instructions and 16 / 32 / 64-bit floating-point operations; the peak computing power is 64 GFLOPS. The neural network acceleration unit has a matrix parallel convolution MAC; it supports compression and pruning of neural network multi-dimensional arrays; it supports 8 / 16-bit fixed-point processing; the peak computing power is 12 TOPS. The computing power can be improved by expanding the number of aerospace-grade SOC embedded chips YuLong810A, ensuring the overall high efficiency, reliability, and flexibility of the terminal artificial intelligence system.

[0039] The neural network acceleration unit in the embodiments of the present application has the platform acceleration ability for deep learning and neural network algorithms, making the computing power platform have certain advantages in front-end image processing, front-end signal processing, and intelligent control. The above platforms in the embodiments of the present application all adopt standardized interfaces, can be reused between different models of spacecraft, reduce the development cost, and improve the product reliability.

[0040] According to a preferred embodiment, the DDR3 / 4 high-speed cache module, the NAND FLASH large-capacity storage module, the NOR FLASH fast-executable storage module, the system clock module, the reset module, the system power management unit, the high-speed bus network, the low-speed bus network, and the computing power platform are each designed as a separate module and are all connected effectively between the interface module and the substrate and between the platform and external devices through a custom-form bus interface using the HSIH4-160ZKV1-02B high-speed connector of aerospace electrical appliances to achieve the application support of the platform. The power supply timing is as Figure 3As shown, the power-on sequence of the single YuLong810A0 module is: 810A0_VDD08_SOC → 810A0_VDD08_CPU → 810A0_VDD18 → 810A0_VDD33 → 810A0_VDDQ1V5 → 810A0_VTTDDR → 810A0_VTTREF → 810A0_VIP_0.9V.

[0041] The aerospace-grade artificial intelligence computing platform based on YuLong810A described in this application mainly aims to provide a high-performance and low-power spaceborne artificial intelligence computing platform and supporting intelligent algorithms for supporting on-orbit processing of massive data, provide basic computing power support for the generation of space intelligent perception capabilities, and also provide a reference for the design of algorithm verification platforms in other high-demand fields, and better serve the field of aerospace artificial intelligence technology. In terms of aerospace payload data in this application, the amount of remote sensing data is rapidly expanding. Through the artificial intelligence computing platform, relevant data can be processed on orbit, reducing the pressure on data transmission and data processing; the AI processing of invalid data on the satellite can be carried out through the artificial intelligence computing platform, making full use of the on-board storage resources and downlink channel resources; because the data is processed on orbit and the final useful results are fed back to the ground, there is no need for manual data processing, which can meet the timeliness requirements of the product, improve the utilization rate of satellite system resources, and reduce the demand for data downlink bandwidth. This application adopts a modular, general-purpose, and scalable platform design architecture, which can be reused in different models of spacecraft, greatly saving the development cost and shortening the development cycle.

[0042] The above-disclosed are only the preferred embodiments of this application. Of course, the scope of rights of this application cannot be limited by this. Therefore, equivalent changes made according to the claims of this application still fall within the scope covered by this application.

Claims

1. A space-grade artificial intelligence computing platform based on YuLong810A, characterized in that: include: DDR3 / 4 cache module, consisting of 72-bit DDR buffer memory VD3D16G72RB199SS2WH with ECC function and DDR interface signals, used to cache large-capacity computing data of aerospace-grade artificial intelligence computing platforms; A NAND FLASH large-capacity storage module, composed of an 8-bit NAND FLASH large-capacity memory VDNF32G08RS50MS4V25-II and NF interface signals, is used to store user applications of the aerospace-grade artificial intelligence computing platform; NOR FLASH fast executable storage module, composed of 16-bit NOR FLASH fast memory JFM29GL256RH and EMIF interface signal, used to store the BootLoader key program of the aerospace-grade artificial intelligence computing platform; A system clock module, composed of a 24MHz active crystal oscillator, a 32.768MHz active crystal oscillator or a passive crystal, a 16MHz active crystal oscillator, a 25MHz active crystal oscillator, and a 62.5MH high-speed differential clock, for clock driving the functional modules of the aerospace-grade artificial intelligence computing platform; A reset module, composed of JSR706RD and a system reset signal, for completing the system reset of the aerospace-grade artificial intelligence computing platform; A system power management unit, which is composed of RSHF4644ARH, RSW3301HRH, RSW1101HRH and various power rail signals, and is used to realize the reliable supply and management of system power for the aerospace-grade artificial intelligence computing platform; A high-speed bus network, consisting of SRIO, PCI-E, 10 / 100 / 1000M adaptive network interface and high-speed related signals, for realizing high-speed data communication of the aerospace-grade artificial intelligence computing platform; A low-speed bus network, consisting of SPI, IIC, UART, 1553B interfaces and low-speed related signals, is used to realize low-speed data communication of the aerospace-grade artificial intelligence computing platform; The computing power platform is composed of one or more aerospace-grade SOC embedded chips YuLong810A with AI computing capabilities, which are used to realize AI computing power support and computing power improvement of the aerospace-grade artificial intelligence computing platform.

2. The aerospace-grade artificial intelligence computing platform based on YuLong810A according to claim 1, characterized in that: The DDR3 / 4 cache module, the NAND FLASH large-capacity storage module, the NOR FLASH fast executable storage module, the system clock module, the reset module, the system power management unit, the high-speed bus network, the low-speed bus network and the computing power platform all use a customized bus interface and the HSIH4-160ZKV1-02B high-speed connector of Aerospace Electrical to achieve effective connection between the interface module and the substrate.

3. The aerospace-grade artificial intelligence computing platform based on YuLong810A according to claim 1, characterized in that: The computing power platform adopts a standardized interface.

4. The aerospace-grade artificial intelligence computing platform based on YuLong810A according to claim 1, characterized in that: The computing power platform includes a GPU unit, a standard shader core, supports EVIS extended instructions, supports 16 / 32 / 64-bit floating-point operations; and has a peak computing power of 64 GFLOPS.

5. The aerospace-grade artificial intelligence computing platform based on YuLong810A according to claim 1, characterized in that: The computing power platform includes a neural network acceleration unit, which is used to realize the platform acceleration capability of the aerospace-grade artificial intelligence computing platform.

6. The aerospace-grade artificial intelligence computing platform based on YuLong810A according to claim 5, characterized in that: The neural network acceleration unit has a matrix parallel convolution MAC, supports compression and pruning of multi-dimensional array processing of neural networks, supports 8 / 16-bit fixed-point processing, and has a peak computing power of 12TOPS.

7. The aerospace-grade artificial intelligence computing platform based on YuLong810A according to claim 1, characterized in that: The computing power platform is used to improve the computing power of the aerospace-grade artificial intelligence computing platform by expanding the number of the aerospace-grade SOC embedded chip YuLong810A.

Images