NPU-based intelligent coprocessor card
By using three NOR FLASH chips for redundant backup and a three-to-two voting system for the programmable logic control unit, combined with dual protection branches and domain-specific power conversion, the problem of firmware single-point failure and power branch monitoring of the NPU accelerator card in the space environment is solved, achieving highly reliable real-time intelligent processing capabilities.
Patent Information
- Application Number
- CN202611526693.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-29
- Publication Date
- 2026-11-17
AI Technical Summary
Existing NPU accelerator cards are prone to firmware errors due to single-event upsets in space environments, lack redundant backups and on-orbit repair capabilities, and lack independent monitoring and protection for power supply branches, making it difficult to achieve highly reliable real-time intelligent processing.
Three NOR FLASH chips are used to form a firmware redundancy backup. The firmware can be autonomously repaired by a three-to-two majority vote through a programmable logic control unit. Dual protection branches and domain power conversion are adopted to independently monitor each power supply branch. Combined with the status acquisition unit, real-time fault isolation and recovery are performed.
The system implements a multi-layered fault-tolerant mechanism for on-orbit firmware, reducing the risk of single-point failure, improving the reliability and fault isolation capability of the power system, and ensuring the continuity and maintainability of onboard intelligent computing tasks.
Smart Images

Figure CN123115032A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of embedded artificial intelligence computing hardware technology, and in particular to an intelligent coprocessor card based on an NPU. Background Technology
[0002] As satellite remote sensing, space target surveillance, on-orbit autonomous navigation, and deep space exploration missions increasingly demand real-time intelligent processing capabilities, spaceborne computing platforms need to perform large-scale image processing, multi-target detection, and deep learning model inference under limited power consumption and size. Traditional spaceborne computing platforms mostly use FPGAs or radiation-hardened DSPs, which, while having good environmental adaptability, have limited general-purpose artificial intelligence inference computing power, making it difficult to balance the scale of complex models with real-time processing latency.
[0003] Commercial or industrial-grade NPUs possess high parallel computing capabilities and mature software ecosystems, which can significantly improve the inference efficiency of intelligent payloads. However, existing NPU accelerator cards are typically designed for terrestrial data centers or industrial equipment conditions, lacking system-level hardening measures for space radiation environments, and cannot directly meet the high reliability requirements of spacecraft for long-term on-orbit operation.
[0004] High-energy particles in space can cause single-event upsets in registers, memories, or logic nodes, leading to firmware data errors, program loading failures, or abnormal logic operation. They can also induce single-event latch-up in power devices or chips, causing abnormally high currents in the corresponding power supply branches. If the faulty branch is not detected and disconnected in time, it may cause thermal damage to the device or even lead to the failure of the entire smart payload.
[0005] Existing NPU accelerator cards typically use a single non-volatile memory to store the boot firmware. When the firmware content in this memory is flipped or corrupted, the NPU cannot obtain a complete and correct boot program, and the card lacks autonomous judgment, backup boot, and on-orbit repair capabilities. As a result, the single firmware memory becomes a significant single point of failure.
[0006] Meanwhile, conventional accelerator cards mostly use commercial DC-DC power supplies and board-level total current acquisition methods, which not only lack input protection suitable for surge, short circuit, and abnormal latch-up currents, but also make it difficult to obtain the voltage and current status of power supply branches such as the NPU core, memory, input / output, and standby. When a local branch is abnormal, the total current data can easily mask the fault location, making it impossible to support rapid isolation and refined on-orbit telemetry.
[0007] In addition, the management logic of existing boards usually only completes basic reset or simple power-on control, without forming a unified closed loop for firmware redundancy arbitration, firmware update, domain power timing, health status monitoring and fault recovery.
[0008] Therefore, it is necessary to propose a highly reliable intelligent coprocessor card based on an NPU to reduce the impact of firmware corruption, power failures, and single points of failure on intelligent computing tasks in a space environment. Summary of the Invention
[0009] Purpose of the invention: To provide an NPU-based intelligent coprocessor card to solve the problems of single-point failure in firmware storage, lack of independent monitoring of power supply branches, difficulty in timely isolation of abnormal latches, and insufficient on-orbit recovery capability in existing NPU acceleration cards.
[0010] Technical solution: An NPU-based intelligent coprocessor card, comprising: an NPU intelligent processor, a programmable logic control unit, a firmware storage unit, a power supply unit, a status acquisition unit, and an external connection interface; the firmware storage unit comprises three NOR FLASH chips respectively connected to the programmable logic control unit, and the three NOR FLASH chips store the same NPU firmware boot program; The programmable logic control unit is communicatively connected to the NPU intelligent processor and the three NOR FLASH chips respectively. During the process of reading the NPU firmware boot program, it synchronously acquires bit data from the three NOR FLASH chips at the same address offset for each address offset of the NPU firmware boot program. It performs a two-out-of-three majority vote on the three bit data acquired each time and determines at least two identical bit data as the correct bit data at the corresponding address offset, so as to form the voted firmware data for the NPU intelligent processor to load. The power supply unit includes an input protection circuit and multiple power conversion modules connected to the input protection circuit. The multiple power conversion modules respectively provide corresponding operating power to the NPU intelligent processor and the programmable logic control unit. The status acquisition unit is connected to the power supply branches of the multiple power conversion modules and the programmable logic control unit, respectively, and is used to acquire the operating status of each power supply branch. The programmable logic control unit is also used to control the power-on sequence of multiple power conversion modules, and to perform alarm, reset and / or power-off protection when the operating state is abnormal; The NPU intelligent processor and the programmable logic control unit are respectively connected to external devices through the external connection interface.
[0011] Preferably, the three NOR FLASH chips are connected to the programmable logic control unit through independent data reading channels; The programmable logic control unit is configured to send read instructions pointing to the same address offset to the three NOR FLASH chips, and synchronously receive bit data returned by the three NOR FLASH chips via the three data read channels; the three NOR FLASH chips also each have an independent chip select signal terminal connected to the programmable logic control unit.
[0012] Preferably, the programmable logic control unit is configured to: read the three NOR FLASH chips in ascending order of address offset, and for each bit in the address offset, perform consistency comparisons on the corresponding bit data of the first NOR FLASH and the second NOR FLASH, the first NOR FLASH and the third NOR FLASH, and the second NOR FLASH and the third NOR FLASH respectively, take the bit value corresponding to at least one set of mutually consistent bit data as the correct bit data of the corresponding bit, and combine the correct bit data in the address offset order to generate the voted firmware data.
[0013] Preferably, the programmable logic control unit is further configured to: compare the bit data returned by the three NOR FLASH chips with the corresponding correct bit data, record the NOR FLASH chips with inconsistent bit data and their address offsets, and after the voted firmware data passes the integrity check, use the correct bit data to write back and repair the recorded address offsets. When the voted firmware data fails the integrity check, the NPU intelligent processor remains in a reset state, generates a firmware upload request, receives the upload data packet input via the RS422 interface, and updates the firmware content in the three NOR FLASH chips according to the upload data packet.
[0014] Preferably, the input protection circuit includes a first protection branch and a second protection branch connected in parallel between the external power input terminal and the power conversion input terminal. The first protection branch is provided with a first fuse, and the second protection branch is provided with a second fuse and a current-limiting resistor connected in series with the second fuse. The rated current of the first fuse and the second fuse is the same, and the DC resistance of the current-limiting resistor is greater than 20 times the DC resistance of the second fuse.
[0015] Preferably, the external power input terminal is used to connect to a 12V DC power supply, and the power conversion input terminal is connected to at least one filter capacitor; the multiple power conversion modules are set independently according to the load functional domain, and at least form an NPU core power supply branch, a memory power supply branch, an input / output power supply branch, a standby power supply branch, and a programmable logic control unit power supply branch.
[0016] Preferably, the status acquisition unit includes a voltage detection circuit and a current detection circuit that are configured one-to-one with each of the power supply branches, as well as a temperature detection circuit for detecting the temperature of the board; the output terminals of each voltage detection circuit, each current detection circuit and the temperature detection circuit are all connected to the programmable logic control unit, and each power conversion module is provided with an enable terminal controlled by the programmable logic control unit.
[0017] Preferably, the programmable logic control unit is configured to collect and record at least one of the following: device health status, driver health status, computing core utilization, core operating frequency, memory utilization, number of ECC errors, number of PCIe software retransmissions, and number of over-temperature frequency reductions and power outages of the NPU intelligent processor, and output the collection results as on-orbit telemetry information through the external connection interface.
[0018] Preferably, the external connection interface includes an FMC connector, and the NPU intelligent processor is connected to the FMC connector via at least one of a PCIe bus, a UART interface, and a GPIO interface; The programmable logic control unit is connected to the FMC connector via at least one of the UART interface, GPIO interface, reset interface and JTAG interface; the clock unit provides clock signals to the NPU intelligent processor and the programmable logic control unit respectively.
[0019] Preferably, the programmable logic control unit includes a power-on timing control module, a firmware arbitration module, a firmware update module, and a health management module; the health management module is used to compare the operating status with a preset threshold, and perform at least one fault handling operation according to the anomaly level, including telemetry alarm, frequency reduction control, power outage of the corresponding power supply branch, and power-on recovery. Compared with the prior art, this application has the following beneficial effects: 1. This application uses three physically independent NOR FLASH chips to form redundant copies of the same boot firmware. The programmable logic control unit performs a three-out-of-two majority vote on the three storage bits with the same address offset and the same bit position in real time. This allows the correct bit data to be output without switching the entire boot source when a single NOR FLASH chip experiences a single-event flip, thus eliminating the single point of failure caused by a single firmware memory.
[0020] 2. This application can automatically overwrite erroneous copies with correct firmware copies, and request the ground or master control equipment to perform an update when all three firmware copies are erroneous, thereby forming a multi-layered fault tolerance mechanism that combines backup startup, autonomous repair and external recovery.
[0021] 3. This application adopts a dual protection branch and a domain-specific power conversion structure to achieve surge, overcurrent and short-circuit protection at the input end, and to implement relatively independent power supply for different loads such as the NPU core, memory, input / output, standby and control logic, which can reduce the risk of a single power supply failure spreading to other functional domains.
[0022] 4. This application collects the voltage, current and board temperature of each power supply branch in real time, and can perform alarm, reset, frequency reduction, power off or power on according to the abnormal branch and abnormal level, so that overcurrent faults such as single-event latch-up can be detected and isolated in a timely manner.
[0023] 5. This application integrates firmware arbitration, power-on timing, status acquisition, telemetry reporting, and fault recovery into the programmable logic control unit, retaining basic monitoring and protection capabilities without relying on the normal operation of the NPU, which is suitable for improving the continuity and maintainability of onboard intelligent computing tasks. Attached Figure Description
[0024] Figure 1 This is a schematic diagram of the overall structure of an NPU-based intelligent coprocessor card according to one embodiment of the present invention.
[0025] Figure 2 This is a schematic diagram of a triple-redundant firmware storage and chip select connection according to one embodiment of the present invention.
[0026] Figure 3 This is a block diagram of a 12V main power input protection circuit according to one embodiment of the present invention.
[0027] Figure 4 This is an electrical schematic diagram of a 12V main power input protection circuit according to one embodiment of the present invention.
[0028] Figure 5 This is a schematic diagram of a domain-based power supply tree according to one embodiment of the present invention.
[0029] Figure 6 This is a power tree for an NPU-based smart coprocessor card according to one embodiment of the present invention. Detailed Implementation
[0030] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.
[0031] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0032] Furthermore, the terms "installation," "setup," "equipped with," "connection," "linking," and "socketing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral structure; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium, or an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of these terms in this invention based on the specific circumstances.
[0033] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0034] like Figures 1-6 As shown, this application relates to an NPU-based intelligent coprocessor card. Figure 1 As shown, this NPU-based intelligent coprocessor card includes: an NPU intelligent processor, a programmable logic control unit, a firmware storage unit, a power supply unit, a status acquisition unit, a clock unit, and an external connection interface; The firmware storage unit includes three NOR FLASH chips, each connected to the programmable logic control unit, and the three NOR FLASH chips store the same NPU firmware boot program. The programmable logic control unit is communicatively connected to the NPU intelligent processor and the three NOR FLASH chips respectively. During the process of reading the NPU firmware boot program, it synchronously acquires bit data from the three NOR FLASH chips at the same address offset for each address offset of the NPU firmware boot program. It performs a two-out-of-three majority vote on the three bit data acquired each time and determines at least two identical bit data as the correct bit data at the corresponding address offset, so as to form the voted firmware data for the NPU intelligent processor to load. The power supply unit includes an input protection circuit and multiple power conversion modules connected to the input protection circuit. The multiple power conversion modules respectively provide corresponding operating power to the NPU intelligent processor and the programmable logic control unit. The status acquisition unit is connected to the power supply branches of the multiple power conversion modules and the programmable logic control unit, respectively, and is used to acquire the operating status of each power supply branch. The programmable logic control unit is also used to control the power-on sequence of multiple power conversion modules, and to perform alarm, reset and / or power-off protection when the operating state is abnormal; The NPU intelligent processor and the programmable logic control unit are respectively connected to external devices through the external connection interface.
[0035] Specifically, such as Figure 1 As shown, the high-reliability intelligent coprocessor card in this embodiment mainly includes an NPU intelligent processor, a programmable logic control unit (CPLD), three NOR FLASH chips, multiple power conversion modules, a status acquisition unit, a clock unit, and an FMC external connection interface. All components can be mounted on the same printed circuit board to form a coprocessor card that can be plugged into onboard main control equipment or payload processing equipment.
[0036] The NPU (Neural Processing Unit) is used to perform inference calculations for neural network models, image or video encoding / decoding, and related data preprocessing. As a non-limiting example, the NPU could be the domestically produced Tianshu Zhikai 100 chip, which features 32GB of HBM2E memory and a peak INT8 computing power of 384 TOPS. In practical applications, other NPU devices with similar interface capabilities and computational functions can also be selected.
[0037] The programmable logic control unit (PLC) serves as a board-level management core independent of the NPU intelligent processor. It performs tasks such as power-on timing control, reset control, verification and arbitration of three firmware copies, firmware updates, power supply status acquisition, and health management. The PLC can utilize CPLD devices with flight experience or high reliability ratings, such as the A3PE3000-1FGG484I, to ensure that basic protection logic remains operational even when the NPU is not started or malfunctions.
[0038] A QSPI or SPI communication link, a reset signal link, a UART communication link, and a GPIO signal link are established between the NPU intelligent processor and the CPLD. The CPLD can control the NPU intelligent processor to maintain or de-reset via the reset signal, and exchange startup status, fault status, and management commands via UART or GPIO.
[0039] The clock unit may include a first clock source providing a reference clock to the NPU intelligent processor and a second clock source providing a reference clock to the CPLD. In one embodiment, the clock source frequency is 25MHz. The clock unit coordinates with the power supply timing to ensure that the CPLD obtains a stable clock first, and then performs subsequent firmware verification and NPU boot control.
[0040] The FMC connector forms a unified connection interface between the board and external devices. The NPU intelligent processor can exchange high-bandwidth inference data with the external host via the PCIe 3.0 x16 bus, and can connect to the FMC connector via PCIe clock, UART and other signals; the CPLD can connect to the FMC connector through multiple GPIO, multiple UART, reset and JTAG interfaces to realize status telemetry, maintenance and debugging, firmware uploading and board-level control.
[0041] Furthermore, the three NOR FLASH chips are connected to the programmable logic control unit through independent data reading channels; The programmable logic control unit is configured to send read instructions pointing to the same address offset to the three NOR FLASH chips, and synchronously receive bit data returned by the three NOR FLASH chips via the three data read channels; the three NOR FLASH chips also each have an independent chip select signal terminal connected to the programmable logic control unit.
[0042] Furthermore, the programmable logic control unit is configured to: read the three NOR FLASH chips in ascending order of address offset, and for each bit in the address offset, perform consistency comparisons on the corresponding bit data of the first NOR FLASH and the second NOR FLASH, the first NOR FLASH and the third NOR FLASH, and the second NOR FLASH and the third NOR FLASH respectively, take the bit value corresponding to at least one set of mutually consistent bit data as the correct bit data of the corresponding bit, and combine the correct bit data in the address offset order to generate the voted firmware data.
[0043] Furthermore, the programmable logic control unit is also configured to: compare the bit data returned by the three NOR FLASH chips with the corresponding correct bit data, record the NOR FLASH chips with inconsistent bit data and their address offsets, and after the voted firmware data passes the integrity check, use the correct bit data to write back and repair the recorded address offsets. When the voted firmware data fails the integrity check, the NPU intelligent processor remains in a reset state, generates a firmware upload request, receives the upload data packet input via the RS422 interface, and updates the firmware content in the three NOR FLASH chips according to the upload data packet.
[0044] Specifically, such as Figure 2 As shown, the three physically independent NOR flash chips are designated as the first NOR flash, the second NOR flash, and the third NOR flash, respectively. All three NOR flash chips store the same version of the NPU firmware boot program. The three NOR flash chips can share clock and data input signals, and simultaneously return read data to the CPLD through the first, second, and third data output channels, respectively. The CPLD also controls the corresponding NOR flash chip using three independent chip select signals: CS0, CS1, and CS2. Therefore, the CPLD can synchronously send read commands with the same address offset to all three NOR flash chips, and can also individually select any one of the NOR flash chips to perform erase, write, or readback verification.
[0045] After the board is powered on, the CPLD first keeps the NPU intelligent processor in a reset state, initializes the read address to the firmware start address, and synchronously sends read commands for the same address offset to the three NOR FLASH chips. The three NOR FLASH chips return bit data through their respective data output channels. The CPLD latches the three return bits within the same read cycle and continues to advance the read process according to the address offset order of bytes, words, or preset data blocks.
[0046] For any bit within the same address offset, the CPLD records the bit data returned from the first NOR FLASH, second NOR FLASH, and third NOR FLASH as D1, D2, and D3, respectively, and performs a two-out-of-three majority vote according to Y = (D1 and D2) or (D1 and D3) or (D2 and D3). When at least two of D1, D2, and D3 are the same, the majority bit value is used as the correct bit data Y for the corresponding address offset and corresponding bit, and Y is sequentially written to the voting output buffer or directly output to the NPU firmware loading channel.
[0047] The three NOR flash chips are packaged with independent memory chips, and the memory bit flip event caused by a single-event event can be treated as an independent event. Let p be the probability that a bit flip occurs at a specific address offset in any NOR flash chip within a given task statistical period. Then the probability that at least two NOR flash chips simultaneously experience bit flips at the same address offset and the same bit flip is P = C(3, 2)p. 2(1-p)+p 3 =3p 2 -2p 3 When considering the mission orbit radiation environment, device radiation test data, and statistical periods, it is determined that p is no higher than approximately 5.77 × 10⁻⁶. -8 When P is not higher than 1×10 -14 That is, the theoretical probability of outputting incorrect data at this position through bit-by-bit voting reaches 10. -14 The magnitude or lower. In practical engineering applications, p and the corresponding threshold can be adjusted according to the device batch, shielding conditions, and task cycle.
[0048] The CPLD combines the correct bit data obtained from each vote according to the address offset order to generate voted firmware data, and performs cyclic redundancy check, hash check, or other integrity check on the voted firmware data. After the voted firmware data passes the integrity check, the CPLD forwards it to the NPU intelligent processor or allows the NPU intelligent processor to load it via the CPLD's voting output channel; after the firmware loading conditions are met, the CPLD releases the NPU intelligent processor from the reset, causing the NPU intelligent processor to execute the boot program.
[0049] During the voting process, the CPLD compares the raw bit data from the three NOR FLASH chips with the corresponding correct bit data. When the bit data returned by a NOR FLASH chip is inconsistent with the correct bit data, the chip identifier, address offset, and bit position of that NOR FLASH chip are recorded. After the NPU intelligent processor successfully boots up, the CPLD can perform background repair based on the recorded information, writing the bytes, words, or data blocks containing the correct bit data back to the corresponding NOR FLASH chip, and performing readback voting and integrity verification after the write operation is completed. During the repair process, at least two other NOR FLASH chips are kept from being erased or written simultaneously to avoid weakening redundancy during the repair operation.
[0050] If the firmware data still fails the integrity check after voting, it indicates that there may be simultaneous flipping of two or more NOR FLASH chips with the same address offset or the same bit position, inconsistent firmware versions, or corrupted data blocks. In this case, the CPLD does not release the NPU intelligent processor from its reset, but generates a firmware upload request and sends the fault status to the onboard master control device via the RS422 interface or an external connection interface. The master control device sends the firmware upload data packet to the CPLD. After performing an integrity check on the upload data packet, the CPLD updates the three NOR FLASH chips sequentially, performing a readback verification after each update.
[0051] After the firmware update is complete and the uploaded firmware data passes the integrity check, the CPLD re-executes the synchronous reading of the three NOR FLASH chips, the bit-by-bit 3-to-2 majority voting, and the NPU boot process. If the uploaded data packet check fails or the written firmware data still fails the integrity check after voting, the CPLD maintains a protected state and continuously reports fault information to avoid loading incomplete or incorrect firmware.
[0052] Furthermore, the input protection circuit includes a first protection branch and a second protection branch connected in parallel between the external power input terminal and the power conversion input terminal. The first protection branch is equipped with a first fuse, and the second protection branch is equipped with a second fuse and a current-limiting resistor connected in series with the second fuse. The rated current of the first fuse and the second fuse is the same, and the DC resistance of the current-limiting resistor is greater than 20 times the DC resistance of the second fuse.
[0053] Furthermore, the external power input terminal is used to connect to a 12V DC power supply, and the power conversion input terminal is connected to at least one filter capacitor; the multiple power conversion modules are set independently according to the load functional domain, and at least form an NPU core power supply branch, a memory power supply branch, an input / output power supply branch, a standby power supply branch, and a programmable logic control unit power supply branch.
[0054] Specifically, such as Figure 3 As shown, the board is powered by an external 12V DC power supply. The input protection circuit includes a first protection branch and a second protection branch connected in parallel. A first fuse is connected in series in the first protection branch, and a second fuse and a current-limiting resistor are connected in series in the second protection branch. The two protection branches converge at the power conversion input terminal.
[0055] The first and second fuses can have the same rated current. The DC resistance of the current-limiting resistor is more than 20 times that of the second fuse, ensuring that most of the operating current flows through the first protection branch under normal conditions, while the second protection branch remains under current-limited shunt. When the first protection branch blows due to surge, overcurrent, or short circuit, the second protection branch suppresses subsequent fault current through the current-limiting resistor, providing a buffer for fault detection and protection.
[0056] A filter capacitor is incorporated at the power conversion input to absorb transient fluctuations and reduce switching noise. The input protection circuit, in conjunction with the subsequent aerospace-grade power conversion module, isolates surges, overcurrents, and short circuits from external connectors.
[0057] like Figure 5 As shown, a dual-fuse protection circuit structure is adopted: fuses with the same rated current are allowed to be used in parallel, and a current-limiting resistor is connected in series on one of the branches. The DC resistance of the current-limiting resistor should be more than 20 times the DC resistance of the fuse.
[0058] The high-reliability intelligent coprocessor card is powered by an external 12V power supply. After passing through a protection circuit, the external 12V is converted into 3.3V, 1.8V, 1.2V, etc. via an aerospace-grade power supply to power the NPU.
[0059] Furthermore, the status acquisition unit includes a voltage detection circuit and a current detection circuit that are configured one-to-one with each of the power supply branches, as well as a temperature detection circuit for detecting the temperature of the board; the output terminals of each voltage detection circuit, each current detection circuit and the temperature detection circuit are all connected to the programmable logic control unit, and each power conversion module is provided with an enable terminal controlled by the programmable logic control unit.
[0060] Furthermore, the programmable logic control unit is configured to collect and record at least one of the following: device health status, driver health status, computing core utilization, core operating frequency, memory utilization, number of ECC errors, number of PCIe software retransmissions, and number of over-temperature frequency reductions and power outages of the NPU intelligent processor, and output the collection results as on-orbit telemetry information through the external connection interface.
[0061] Furthermore, the external connection interface includes an FMC connector, and the NPU intelligent processor is connected to the FMC connector via at least one of a PCIe bus, a UART interface, and a GPIO interface; The programmable logic control unit is connected to the FMC connector via at least one of the UART interface, GPIO interface, reset interface and JTAG interface; the clock unit provides clock signals to the NPU intelligent processor and the programmable logic control unit respectively.
[0062] Furthermore, the programmable logic control unit includes a power-on timing control module, a firmware arbitration module, a firmware update module, and a health management module; the health management module is used to compare the operating status with a preset threshold, and perform at least one fault handling operation in sequence according to the abnormality level, including telemetry alarm, frequency reduction control, power outage of the corresponding power supply branch, and power-on recovery.
[0063] Specifically, such as Figure 4 and 6 As shown, the protected 12V power supply is input to multiple power conversion modules. These modules are configured according to load functional domains, including a core power conversion module, a storage power conversion module, an input / output power conversion module, a standby power conversion module, and a control power conversion module. Each power conversion module can output voltages of 0.75V, 0.84V, 1.2V, 1.8V, 2.5V, and 3.3V, matching the specific device requirements.
[0064] The core power conversion module supplies power to the NPU computing core, the storage power conversion module supplies power to the on-chip or packaged memory, the input / output power conversion module supplies power to the NPU interface circuitry, the standby power conversion module supplies power to the standby domain, and the control power conversion module supplies power to the CPLD and its related logic. These different power supply branches are relatively independent, allowing for the individual location and control of localized faults.
[0065] Each power conversion module has an enable terminal connected to the CPLD. Upon power-up, the CPLD first enables the power supplies related to its own control domain and status acquisition. After confirming that the control domain voltage is stable, it sequentially enables the standby power supply branch, input / output power supply branch, memory power supply branch, and NPU core power supply branch. After each power supply meets the preset stability conditions, the CPLD then releases the NPU intelligent processor reset.
[0066] When power is lost, the CPLD can shut down each power supply branch in the reverse order of power-on or in the order determined by the NPU device requirements. By having the CPLD uniformly control the power enable and reset signals, it is possible to avoid the impact caused by multiple power rails transitioning simultaneously, and reduce the probability of abnormal startup caused by power timing not meeting device requirements.
[0067] The status acquisition unit includes voltage and current detection circuits located in each power supply branch, as well as temperature detection circuits located in key heat-generating areas of the board. The voltage and current detection circuits can output analog sample values or digital status values, respectively, and the CPLD reads the detection results according to a preset cycle.
[0068] The CPLD sets corresponding normal ranges, alarm thresholds, and protection thresholds for different power supply branches. When the voltage or current of a power supply branch exceeds the protection threshold, the CPLD can locate the abnormal load domain based on the branch identifier and shut down the enable terminal of the corresponding power conversion module. When the abnormality involves the NPU core or multiple power supply branches, the CPLD can maintain the NPU reset and perform a controlled power-off of the entire board.
[0069] Single-event latch-up typically manifests as a sudden increase in current in the corresponding power supply branch. Because this implementation sets up independent current detection for each functional domain, the CPLD can identify local overcurrents before the total input current reaches the board-wide protection threshold, reducing the possibility of continuous latch-up current damaging the NPU intelligent processor or power supply devices.
[0070] The temperature detection circuit is used to obtain the temperature near the NPU, the power conversion module, or other critical locations on the board. When the temperature reaches the first threshold, the CPLD sends a frequency reduction command to the NPU intelligent processor or external host via UART or GPIO; when the temperature reaches a higher second threshold, the CPLD executes power-off of the corresponding power supply branch or board-wide protection.
[0071] In addition to hardware power status, the board also collects operational information such as device health status, driver health status, compute core utilization, core operating frequency, memory utilization, number of ECC errors, number of PCIe software retransmissions, number of module over-temperature frequency reductions, and number of power outages. The relevant software status can be sent from the NPU-side management program to the CPLD or external controller via UART, PCIe management channel, or shared registers.
[0072] The CPLD records hardware sampling results and software operation information according to a unified time base and outputs them as on-orbit telemetry data via the FMC connector or RS422 interface. The telemetry data can at least include the status category, abnormal power supply branch, sampled value, threshold, handling action, and cumulative number of times, which helps the ground system to judge the health trend of the board.
[0073] In one fault classification method, when the detected value reaches the alarm threshold but not the protection threshold, the health management module only records the event and sends a telemetry alarm; when the detected value is continuously abnormal or reaches the protection threshold, the NPU is reset or the frequency is reduced; when severe overcurrent, overtemperature or voltage instability is detected, the corresponding power supply branch is shut down.
[0074] After a power supply branch is de-energized, the CPLD can wait for a preset power dissipation and cooling time before re-energizing the branch according to the normal power-on sequence and re-performing firmware verification and NPU startup. If the same anomaly occurs repeatedly within a limited number of times, the CPLD will keep the faulty branch off and request external maintenance to avoid cumulative damage caused by cyclic power-on.
[0075] Therefore, the board can achieve fault detection within seconds and fault recovery within minutes. The fault detection time and recovery time can be configured according to the task level, telemetry cycle, power stabilization time and NPU startup time, and are not limited to fixed values.
[0076] A complete startup process of this embodiment includes: an external 12V power supply is input through an input protection circuit; the control power conversion module is first enabled; the CPLD completes its own initialization and status acquisition unit check; the CPLD synchronously sends read instructions with the same address offset to three NOR FLASH chips, performs a three-to-two majority vote on the returned bit data bit by bit, and forms voted firmware data; after the voted firmware data passes integrity verification, each NPU power supply branch is enabled in a preset order; the NPU intelligent processor loads the voted firmware data; the CPLD is de-reset; the NPU intelligent processor starts up and establishes PCIe communication with the external host.
[0077] One abnormal recovery process in this embodiment includes: the status acquisition unit detects abnormal voltage, current, temperature, or operating status; the health management module compares the abnormal value with the threshold and determines the abnormal level; the CPLD records and reports the abnormal event; according to the abnormal level, it performs frequency reduction, reset, branch power-off, or whole board power-off; after the protection waiting time ends, it re-powers on; it re-executes the synchronous reading of the three NOR FLASH chips, bit-by-bit majority voting, integrity verification, and NPU startup; after startup, it continues to monitor the corresponding fault item.
[0078] When an abnormal event is a bit flip of a single NOR FLASH chip, the firmware arbitration module outputs the correct bit data in real time through a three-to-two majority vote. The firmware update module then repairs the corresponding address offset based on the inconsistency record. If the firmware data still fails the integrity check after voting, the board enters the uploading recovery process. When an abnormal event is a damage to the intelligent computing software stack while the hardware power supply is normal, the software stack can be reloaded or the NPU intelligent processor can be reset through an external main controller.
[0079] This invention does not limit the specific models of the NPU intelligent processor, CPLD, NOR FLASH, power conversion module, or connectors, nor does it limit the number of power supply branches or the output voltage. The three NOR FLASH chips can be connected via parallel read interfaces, independent serial read interfaces, or other interfaces that enable the CPLD to obtain three sets of bit data corresponding to the same address offset. As long as a combination of three firmware redundancies, bit-by-bit majority voting, independent management and control, domain-specific power supply monitoring, and anomaly protection is adopted, the device models, number of interfaces, thresholds, and timing parameters can be adjusted according to the specific task.
[0080] In addition to its use in satellite on-orbit intelligent identification, this invention can also be used for launch vehicle image analysis, autonomous perception of space robots, real-time reasoning of deep space probes, and other embedded devices that have high requirements for computing power and fault tolerance.
[0081] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.
Claims
1. An NPU-based intelligent coprocessor card, characterized in that, include: The system comprises an NPU intelligent processor, a programmable logic control unit, a firmware storage unit, a power supply unit, a status acquisition unit, and an external connection interface; the firmware storage unit includes three NOR FLASH chips, each connected to the programmable logic control unit, and the three NOR FLASH chips store the same NPU firmware boot program. The programmable logic control unit is communicatively connected to the NPU intelligent processor and the three NOR FLASH chips respectively. During the process of reading the NPU firmware boot program, it synchronously acquires bit data from the three NOR FLASH chips at the same address offset for each address offset of the NPU firmware boot program. It performs a two-out-of-three majority vote on the three bit data acquired each time and determines at least two identical bit data as the correct bit data at the corresponding address offset, so as to form the voted firmware data for the NPU intelligent processor to load. The power supply unit includes an input protection circuit and multiple power conversion modules connected to the input protection circuit. The multiple power conversion modules respectively provide corresponding operating power to the NPU intelligent processor and the programmable logic control unit. The status acquisition unit is connected to the power supply branches of the multiple power conversion modules and the programmable logic control unit, respectively, and is used to acquire the operating status of each power supply branch. The programmable logic control unit is also used to control the power-on sequence of multiple power conversion modules, and to perform alarm, reset and / or power-off protection when the operating state is abnormal; The NPU intelligent processor and the programmable logic control unit are respectively connected to external devices through the external connection interface.
2. The NPU-based intelligent coprocessor card according to claim 1, characterized in that, The three NOR FLASH chips are connected to the programmable logic control unit through independent data reading channels; The programmable logic control unit is configured to send read instructions pointing to the same address offset to the three NOR FLASH chips, and synchronously receive bit data returned by the three NOR FLASH chips via the three data read channels; the three NOR FLASH chips also each have an independent chip select signal terminal connected to the programmable logic control unit.
3. The NPU-based intelligent coprocessor card according to claim 2, characterized in that, The programmable logic control unit is configured to: read the three NOR FLASH chips in ascending order of address offset, and for each bit in the address offset, perform consistency comparisons on the corresponding bit data of the first NOR FLASH and the second NOR FLASH, the first NOR FLASH and the third NOR FLASH, and the second NOR FLASH and the third NOR FLASH respectively, take the bit value corresponding to at least one set of mutually consistent bit data as the correct bit data of the corresponding bit, and combine the correct bit data in the address offset order to generate the voted firmware data.
4. The NPU-based intelligent coprocessor card according to claim 3, characterized in that, The programmable logic control unit is further configured to: compare the bit data returned by the three NOR FLASH chips with the corresponding correct bit data, record the NOR FLASH chips with inconsistent bit data and their address offsets, and, after the voted firmware data passes the integrity check, use the correct bit data to write back and repair the recorded address offsets. When the voted firmware data fails the integrity check, the NPU intelligent processor remains in a reset state, generates a firmware upload request, receives the upload data packet input via the RS422 interface, and updates the firmware content in the three NOR FLASH chips according to the upload data packet.
5. The NPU-based intelligent coprocessor card according to claim 1, characterized in that, The input protection circuit includes a first protection branch and a second protection branch connected in parallel between the external power input terminal and the power conversion input terminal. The first protection branch is equipped with a first fuse, and the second protection branch is equipped with a second fuse and a current-limiting resistor connected in series with the second fuse. The rated current of the first fuse and the second fuse is the same, and the DC resistance of the current-limiting resistor is greater than 20 times the DC resistance of the second fuse.
6. The NPU-based intelligent coprocessor card according to claim 5, characterized in that, The external power input terminal is used to connect to a 12V DC power supply, and the power conversion input terminal is connected to at least one filter capacitor; the multiple power conversion modules are set independently according to the load functional domain, and at least form an NPU core power supply branch, a memory power supply branch, an input / output power supply branch, a standby power supply branch, and a programmable logic control unit power supply branch.
7. The NPU-based intelligent coprocessor card according to claim 6, characterized in that, The status acquisition unit includes a voltage detection circuit and a current detection circuit, each corresponding to one of the power supply branches, as well as a temperature detection circuit for detecting the temperature of the board. The output terminals of each voltage detection circuit, each current detection circuit, and each temperature detection circuit are connected to the programmable logic control unit. Each power conversion module is provided with an enable terminal controlled by the programmable logic control unit.
8. The NPU-based intelligent coprocessor card according to claim 7, characterized in that, The programmable logic control unit is configured to collect and record at least one of the following: device health status, driver health status, computing core utilization, core operating frequency, memory utilization, number of ECC errors, number of PCIe software retransmissions, and number of over-temperature frequency reductions and power outages of the NPU intelligent processor, and output the collection results as on-orbit telemetry information through the external connection interface.
9. The NPU-based intelligent coprocessor card according to claim 1, characterized in that, The external connection interface includes an FMC connector, and the NPU intelligent processor is connected to the FMC connector via at least one of PCIe bus, UART interface and GPIO interface; The programmable logic control unit is connected to the FMC connector via at least one of the UART interface, GPIO interface, reset interface and JTAG interface; the clock unit provides clock signals to the NPU intelligent processor and the programmable logic control unit respectively.
10. The NPU-based intelligent coprocessor card according to claim 7, characterized in that, The programmable logic control unit includes a power-on timing control module, a firmware arbitration module, a firmware update module, and a health management module. The health management module is used to compare the operating status with a preset threshold and, according to the abnormality level, sequentially execute at least one fault handling operation among telemetry alarm, frequency reduction control, power outage of the corresponding power supply branch, and power-on recovery.
