Satellite-borne computing center system

Abstract

The invention provides a satellite-borne computing center system, in particular to a satellite-borne computing center system with high performance and high computing power, which comprises a data interaction module, a task management module, an artificial intelligence AI module M1 and an artificial intelligence AI module M2. The data interaction module is an input and output interface for high-speed data processing; the task management module serves as a task scheduling center and distributes calculation tasks to the AI module; the artificial intelligence AI module M1 and the artificial intelligence AI module M2 adopt domestic and imported AI chips, and the artificial intelligence AI module can work in an RC mode and an EP mode, provides computing power guarantee for satellite-borne high-computing-power computing tasks, and has the characteristics of flexible expansion and agile deployment. All the modules are interconnected through a serial port management plane, a gigabit network control plane, a PCIE data plane and a power supply plane, and control and data interaction between processors are achieved.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence hardware architecture of satellite information systems, and specifically relates to a satellite-borne computing center system, and specifically relates to a satellite-borne high-performance and high-computing power computing center system. Background Art

[0002] Artificial intelligence computing is developing rapidly, and breakthroughs in satellite-borne edge computing will improve the efficiency of satellite use and provide software and hardware services for satellite on-orbit applications. The computing power of traditional satellites can no longer meet the development needs of artificial intelligence in the new era. High-computing system design is the software and hardware guarantee for satellite-borne artificial intelligence computing.

[0003] The existing satellite-borne processing power is at the level of 20-50TPOS, which is still far behind the 278TOPS computing power of NVIDIA Orin. There is an urgent need to develop a domestic high-computing satellite-borne processor platform. Relying on Huawei Ascend's latest and most powerful processor Ascend310P, a 6U high-computing board is developed to achieve a peak 280TOPS high-computing satellite processor. The high-performance CPU features 8 cores, a main frequency of more than 2.0Ghz, and a memory space of more than 32GB.

[0004] Traditional satellite onboard integrated electronic systems provide remote control and telemetry for satellites. They are the brains of satellite control and information systems, and are characterized by high reliability, always-on, primary and standby redundant design, and no failure in orbit. Compared with traditional integrated electronics, the onboard high-performance computing center has the characteristics of high computing power, high storage, and high bandwidth. It is part of the payload data processing in satellite applications and works with the payload. When the payload is shut down, the high-performance computing center can be powered off and radiated to dissipate heat. The high-performance computing center can also serve as a data processing center for other satellite payloads if the energy of the satellite allows.

[0005] The patent document with application number 202311792926.4 discloses an on-board processing system based on a domestic AI chip, which includes a routing processing module, an AI processing module, a refresh injection module, a bus communication module, and a management and power distribution module. Among them, the routing processing module receives data acquired by the camera subsystem, the AI ​​processing module processes the received data in real time, the refresh injection module receives satellite platform data and performs injection program control, the bus communication module receives input data from the satellite management platform, and the management and power distribution module connects to the primary power supply of the satellite platform and provides power supply guarantee for the system of the present invention.

[0006] The patent document with application number 201910888320.8 discloses a satellite data processing system architecture design, including: primitive extraction step: extracting algorithm primitives from processing algorithms with different functions, and the algorithm primitives are linearly independent, and the algorithm primitives are non-related independent basic units that constitute each processing algorithm; rule establishment step: analyzing the functional dependencies and task algorithm structure of the processing algorithm, and establishing online reconstruction algorithm generation rules based on algorithm primitives; mapping platform design step: designing a containerized hardware resource autonomous mapping platform to realize the autonomous mapping capability of the target algorithm to the hardware platform resources.

[0007] The patent document with application number 202311699805.5 discloses a template construction system for universal hardware modules of a satellite integrated information system, wherein the universal modules include: a data processing module, a data storage module, a data exchange module, a management control module and a secondary power supply module; the secondary power supply module provides the required power for the data processing module, the data storage module, the data exchange module and the management control module; the data exchange module provides star-shaped data exchange for the data processing module, the data storage module and the management control module; the data processing module, the data storage module and the management control module respectively realize high-performance processing, large-capacity storage and high-reliability management functions; the universal modules are further decomposed, including a reconfigurable processing configuration template, a high-reliability processing configuration template, a power conversion configuration template, a system monitoring configuration template and an extensible template.

[0008] In addition, the AI ​​module in patent document CN118227318A only has RC mode, not EP mode, and each AI module is relatively independent. Compared with the AI ​​module in EP mode, patent document CN118227318A has the disadvantage of no computing power balance and distributed collaborative computing capabilities in terms of fast access to memory data. Compared with patent document CN118227318A, the present invention has both RC mode and EP mode, which can be flexibly switched according to application requirements in on-orbit applications.

[0009] The bus interconnection mechanism used in patent document CN118227318A is SRIO, and the bus interconnection mechanism used in the present invention is PCIE. In computing power pooling and collaborative scheduling, the EP mode under the PCIE mechanism can unify the application layer APP into one CPU, which has the advantage of achieving computing power balance and pooling. However, the patent document CN118227318A uses the SRIO bus, and the application layer APP must be allocated to each separate AI module for operation, which affects computing power pooling and balance.

[0010] Patent document CN118227318A lacks EP mode. The EP mode of the present invention is similar to the host graphics card operation. In terms of reliability, the failure of an AI module in the EP mode does not affect the task operation. In the RC mode, each AI module executes the task independently, and the failure of this module corresponds to the failure of the task.

[0011] The AI ​​module actually used in patent document CN118227318A is Yulong810A, with a computing power of 20TOPS. The AI ​​module used in the present invention is Huawei Ascend310p, with a computing power of 140TOPS, which achieves cross-generation in terms of computing power level. In addition, Ascend310P uses a purely domestic CPU, while the CPU used in Yulong810A in patent document CN118227318A is an imported CPU. The present invention has advantages in performance and localization.

[0012] In summary, the present invention implements the reconstruction switching of RC and EP modes on a set of AI hardware, and performs mode switching according to the on-orbit application tasks. In the onboard AI operation, the steady-state RC mode will be switched to the emergency high-computing EP mode according to the high-computing power task. In terms of reliability, the failure of a single AI module after the computing power is pooled in the EP mode does not affect the normal execution of the task. Summary of the invention

[0013] In view of the defects in the prior art, an object of the present invention is to provide a satellite-borne computing center system.

[0014] A satellite-borne computing center system provided according to the present invention includes: a data interaction module, a task management module, and one or more artificial intelligence AI modules;

[0015] The data interaction module includes FPGA, which realizes the functions of data reception and processing result output;

[0016] The task management module includes a CPU, which implements task management, data transmission and reception, and scheduling of the artificial intelligence (AI) module;

[0017] The artificial intelligence (AI) module includes an NPU, which provides high computing power for onboard computing;

[0018] Core processors such as FPGA, CPU, and NPU are interconnected through the serial bus, gigabit network bus, and PCIE bus in the bus interconnection mechanism.

[0019] Preferably, the serial bus realizes telemetry and remote control, parameter upload and modification, and telemetry parameter download between modules, receives control instructions from the integrated electronics of the entire satellite, and feeds back the telemetry status of each computing stage of the high-performance computer;

[0020] The Gigabit Ethernet bus realizes task control, target slicing result transmission, containerized task deployment, and identifies the core processor based on the IP address to achieve task-level interconnection;

[0021] The PCIE bus realizes the remote sensing data interaction function; when the artificial intelligence AI module is in RC mode, the data exchange module acts as the EP end, receives the payload remote sensing data and transparently transmits it to the artificial intelligence AI module for calculation; when the artificial intelligence AI module is in EP mode, the data exchange module first caches the data in the task control module memory, and then the CPU allocates the remote sensing calculation task to the artificial intelligence AI module through the PCIE bus.

[0022] Preferably, the power supply bus in the bus interconnection mechanism independently supplies power to the task control module, the data exchange module, and the artificial intelligence AI module; the power supply bus receives ground remote control instructions to power on and off each module according to task requirements; the power supply bus switches between full power and full computing power mode and low power consumption and low computing power mode according to ground task requirements.

[0023] Preferably, the multi-channel PCIE interface of the task control module is interconnected with the artificial intelligence AI module, and at the same time, the remote sensing camera image data of the FPGA of the data exchange module is received through the PCIE interface or the optical port;

[0024] The AI ​​module is interconnected with the CPU of the task management module and the FPGA of the data exchange module through the PCIE interface. The AI ​​module has RC mode and EP mode, and the two modes can be switched freely.

[0025] Among them, the reconstruction switching between RC mode and EP mode is based on a set of AI hardware, and the mode switching is performed according to the on-orbit application tasks; in the on-board AI calculations, the steady-state RC mode is switched to the emergency high-computing power EP mode according to the set high-computing power tasks; in the EP mode, after the computing power is pooled, the failure of a single AI module will not affect the normal execution of the task.

[0026] Preferably, in RC mode, the artificial intelligence AI module works independently to achieve fault isolation, and is driven by the FPGA data of the data interaction module.

[0027] Preferably, when the artificial intelligence AI module runs in RC mode, the FPGA of the data interaction module runs in EP mode, the remote sensing data stream enters the artificial intelligence AI module for calculation, and the calculation results are returned to the CPU. After the CPU packages and frames, it is sent out by the FPGA through the interconnection bus; the AI ​​framework, NPU driver, AI engine, and AI running APP of the artificial intelligence AI module all run in the ARM processor of the artificial intelligence AI module.

[0028] Preferably, in the EP mode, the artificial intelligence AI module is uniformly managed and controlled by the CPU of the task management module, which can realize the pooling of computing power resources and achieve computing power expansion.

[0029] Preferably, when the artificial intelligence AI module runs in EP mode, the CPU of the task management module runs in RC mode, and the remote sensing data stream is sent to the memory of the task management module CPU through FPGA for cache, and the CPU is used as the main scheduler to assign tasks to the artificial intelligence AI module to realize remote sensing data calculation; the CPU calculation results are packaged by the CPU and sent out through FPGA; the AI ​​framework, NPU driver, AI engine, and AI running APP of the artificial intelligence AI module all run in the task management module CPU, and the CPU side can recognize multiple artificial intelligence AI modules, and the computing power pooling is expandable.

[0030] Preferably, the data interaction module, task management module, artificial intelligence AI module and other modules are used as a unified VPX board, one or more modules are configured according to task requirements, and each module is integrated into a unified VPX backplane.

[0031] A satellite provided according to the present invention comprises the onboard computing center system.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. The present invention starts from the satellite application performance evaluation and designs a satellite high-performance computing system.

[0034] 2. The high-performance CPU of the task control module in the present invention is used for satellite remote sensing task management, and has the functions of remote sensing target characteristic management, target library construction management, geographic information management, etc., which can automatically complete the historical information of remote sensing targets and improve the efficiency of satellite use.

[0035] 3. The artificial intelligence AI module in the present invention reaches the 100 TOPS level, which meets the AI ​​operation under large model parameters and has both EP mode and RC mode. The AI ​​module can be expanded arbitrarily, has the ability of computing power pooling, agile deployment of network algorithms, and has stronger adaptability in dealing with large model parameter processing. In RC mode, the artificial intelligence AI module is driven by the data exchange FPGA data flow to work independently, with high reliability and strong anti-damage capability.

[0036] 4. The data exchange FPGA in the present invention adopts SRAM type FPGA, receives the payload remote sensing data and returns the task results of the high performance computing center to the data transmission to transmit to the user.

[0037] 5. The independent modules in the present invention are interconnected through the serial port management plane, the network port control plane, the PCIE data plane and the power supply plane, are visible to each other, tasks are flexibly deployed, and algorithms are reconstructed and migrated. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0039] Figure 1 This is the system block diagram of the high-performance onboard computing center;

[0040] Figure 2 This is the block diagram of the high computing power board system;

[0041] Figure 3 This is the system workflow diagram under the EP mode of the artificial intelligence AI module;

[0042] Figure 4 This is the system working information flow diagram in RC mode of artificial intelligence AI module. DETAILED DESCRIPTION

[0043] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0044] According to the present invention, a satellite-borne high-performance computing center is provided, including: a data interaction module, a task management module, an artificial intelligence AI module M1, and an artificial intelligence AI module M2; the data interaction module, the task management module, the artificial intelligence AI module M1, and the artificial intelligence AI module M2 are used as a unified VPX board, and one or more modules are configured according to task requirements, and each module is integrated into a unified VPX backplane to realize the scalability of computing resources.

[0045] The core processor of the data interaction module is FPGA, which realizes the functions of data reception and processing result output; the core processor of the task management module is CPU, which realizes task management scheduling and is the scheduling center of data reception and transmission and artificial intelligence AI module, with high memory and large-capacity storage characteristics; the core processor of the artificial intelligence AI module is ARM NPU, which provides high computing power for onboard computing.

[0046] In the satellite-borne high-performance computing center, the core processors are interconnected through the serial bus, the gigabit network bus, and the PCIE bus. The serial bus realizes low-speed control such as telemetry and remote control, parameter annotation and modification between modules, and telemetry parameter transmission, receives control instructions from the integrated electronics of the entire satellite, and feedbacks the telemetry status of each computing stage of the high-performance computer. The gigabit network bus realizes functions such as task control, target slicing result transmission, and containerized task deployment. According to the IP address identification processor, it realizes task-level interconnection, containerized rapid deployment, task execution result feedback, and software on-orbit programming deployment. In the hardware product development stage of the satellite-borne high-performance supercomputing center, it also serves as a debugging login interface to realize algorithm debugging and updating, software deployment and other functions. The PCIE bus realizes the remote sensing high-speed data interaction function; realizes high-bandwidth transmission of remote sensing data. When the artificial intelligence AI module is in RC mode, the data exchange module acts as the EP end, receives the payload remote sensing data and transparently transmits it to the artificial intelligence AI module for high-computing power calculation; when the artificial intelligence AI module is in EP mode, the data exchange module first caches the data in the task control module memory, and then the CPU allocates the remote sensing computing task to the artificial intelligence AI module through the PCIE high-speed bus.

[0047] The power supply bus in the bus interconnection mechanism can independently power the mission control module, data exchange module, artificial intelligence AI module M1, and artificial intelligence AI module M2. The power supply bus can receive ground remote control commands to power on and off each module according to mission requirements. The power supply bus can switch to different computing power levels such as full power and full computing power mode, low power consumption and low computing power mode according to ground mission requirements.

[0048] The task control module with high-performance CPU has large-capacity DDR memory, rich FLASH storage, multiple PCIE3.0 interfaces, network interfaces, and UART interfaces. The multiple PCIE interfaces are interconnected with the artificial intelligence AI module, and the remote sensing camera image data of the data exchange module FPGA is received through the PCIE interface or optical port. The high-performance CPU is installed with basic libraries such as the operating system, NPU driver, CANN library or CUDA library, and AI image computing framework, and the tasks are deployed through containerization.

[0049] The AI ​​modules M1 and M2 are interconnected with the CPU of the task management module and the FPGA of the data exchange module through the PCIE high-speed interface. The AI ​​modules have RC mode and EP mode, and the two modes can be switched freely. In the RC mode, the AI ​​module works independently to achieve fault isolation, and is driven by the FPGA data of the data interaction module. In the EP mode, the AI ​​module is uniformly managed and controlled by the CPU of the task management module, which can realize the pooling of computing resources and the expansion of computing power.

[0050] When the artificial intelligence AI module runs in RC mode, the FPGA of the data interaction module runs in EP mode, and the remote sensing data stream enters the artificial intelligence AI module for high-speed computing. The calculation results are returned to the CPU, which is packaged and framed by the CPU and sent out by the FPGA through the interconnection bus. The AI ​​framework, NPU driver, AI engine (CUDA or CANN), and AI running APP of the artificial intelligence AI module all run in the ARM processor of the artificial intelligence AI module.

[0051] When the artificial intelligence AI module runs in EP mode, the CPU of the task management module runs in RC mode, and the remote sensing data stream is sent to the memory of the task management module CPU through FPGA for cache. The CPU is used as the main scheduler to assign tasks to the artificial intelligence AI module to realize high-speed calculation of remote sensing data. The CPU calculation results are packaged by the CPU and sent through FPGA. The AI ​​framework, NPU driver, AI engine (CUDA or CANN), and AI running APP of the artificial intelligence AI module all run in the task management module CPU. Multiple artificial intelligence AI modules can be seen on the CPU side, and the computing power pooling is expandable.

[0052] In the present invention, the data interaction module is an input and output interface for high-speed data processing; the task management module, as the task scheduling center, allocates computing tasks to the artificial intelligence AI module; the artificial intelligence AI module M1 and the artificial intelligence AI module M2 use two kinds of AI chips, domestic and imported, and the artificial intelligence AI module can work in RC and EP modes, providing computing power guarantee for satellite-borne high-computing computing tasks, with the characteristics of flexible expansion and agile deployment. The modules are interconnected through the serial port management plane, the Gigabit network control plane, the PCIE data plane, and the power supply plane to realize the control and data interaction between processors.

[0053] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A satellite-borne computing center system, characterized in that: include: Data interaction module, task management module, and one or more artificial intelligence AI modules; The data interaction module includes FPGA, which realizes the functions of data reception and processing result output; The task management module includes a CPU, which implements task management, data transmission and reception, and scheduling of the artificial intelligence (AI) module; The artificial intelligence (AI) module includes an NPU, which provides high computing power for onboard computing; Core processors such as FPGA, CPU, and NPU are interconnected through the serial bus, gigabit network bus, and PCIE bus in the bus interconnection mechanism.

2. The satellite-based computing center system according to claim 1, characterized in that: The serial bus realizes telemetry and remote control between modules, parameter upload and modification, and telemetry parameter download, receives control instructions from the integrated electronics of the whole satellite, and feeds back the telemetry status of each computing stage of the high-performance computer; The Gigabit Ethernet bus realizes task control, target slicing result transmission, containerized task deployment, and identifies the core processor based on the IP address to achieve task-level interconnection; The PCIE bus realizes the remote sensing data interaction function; when the artificial intelligence AI module is in RC mode, the data exchange module acts as the EP end, receives the payload remote sensing data and transparently transmits it to the artificial intelligence AI module for calculation; when the artificial intelligence AI module is in EP mode, the data exchange module first caches the data in the task control module memory, and then the CPU allocates the remote sensing calculation task to the artificial intelligence AI module through the PCIE bus.

3. The satellite-based computing center system according to claim 1, characterized in that: The power supply bus in the bus interconnection mechanism independently supplies power to the mission control module, data exchange module, and artificial intelligence AI module; the power supply bus receives ground remote control commands to power on and off each module according to mission requirements; the power supply bus switches between full-power and full-computing-power mode and low-power and low-computing-power mode according to ground mission requirements.

4. The satellite-based computing center system according to claim 1, characterized in that: The multi-channel PCIE interface of the task control module is interconnected with the artificial intelligence AI module, and at the same time, the remote sensing camera image data of the FPGA of the data exchange module is received through the PCIE interface or optical port; The AI ​​module is interconnected with the CPU of the task management module and the FPGA of the data exchange module through the PCIE interface. The AI ​​module has RC mode and EP mode, and the two modes can be switched freely. Among them, the reconstruction switching between RC mode and EP mode is based on a set of AI hardware, and the mode switching is performed according to the on-orbit application tasks; in the on-board AI calculations, the steady-state RC mode is switched to the emergency high-computing power EP mode according to the set high-computing power tasks; in the EP mode, after the computing power is pooled, the failure of a single AI module will not affect the normal execution of the task.

5. The satellite-based computing center system according to claim 4, characterized in that: In RC mode, the artificial intelligence AI module works independently to achieve fault isolation, and is driven by the FPGA data of the data interaction module.

6. The satellite-borne computing center system according to claim 5, characterized in that: When the artificial intelligence AI module runs in RC mode, the FPGA of the data interaction module runs in EP mode, the remote sensing data stream enters the artificial intelligence AI module for calculation, and the calculation results are returned to the CPU. After the CPU packages and frames, it is sent out by the FPGA through the interconnection bus; the AI ​​framework, NPU driver, AI engine, and AI running APP of the artificial intelligence AI module all run in the ARM processor of the artificial intelligence AI module.

7. The satellite-borne computing center system according to claim 4, characterized in that: In EP mode, the artificial intelligence AI module is uniformly managed and controlled by the CPU of the task management module, which can realize the pooling of computing resources and achieve computing power expansion.

8. The satellite-borne computing center system according to claim 7, characterized in that: When the artificial intelligence AI module runs in EP mode, the CPU of the task management module runs in RC mode. The remote sensing data stream is sent to the memory of the task management module CPU through FPGA for cache. The CPU acts as the main scheduler to assign tasks to the artificial intelligence AI module to realize remote sensing data calculation. The CPU calculation results are packaged by the CPU and sent out through FPGA. The AI ​​framework, NPU driver, AI engine, and AI running APP of the artificial intelligence AI module all run in the task management module CPU. The CPU side can recognize multiple artificial intelligence AI modules, and the computing power pooling is expandable.

9. The satellite-borne computing center system according to claim 1, characterized in that: The data interaction module, task management module, and artificial intelligence AI module are unified VPX boards. One or more modules are configured according to task requirements, and each module is integrated into a unified VPX backplane.

10. A satellite, characterized in that: A satellite-based computing center system comprising any one of claims 1 to 9.

Images