Power consumption control method and system for a neural network system based on a complex architecture

Abstract

The application discloses a power consumption control method and system of a neural network system based on a composite framework, and relates to the technical field of neural network systems. The expert-level identification module performs integral operation of different dimensions on each data feature, and determines the corresponding matching integral in combination with a logic sniffing receptor. According to the analysis of the integral comparison result, a plurality of activation factors are determined, the plurality of activation factors are loaded to the corresponding composite framework, the corresponding expert-level identification module is dynamically activated, matching is performed along the plurality of mapping relationships recorded in the interaction matrix, and finally, the execution module is screened out. The corresponding power consumption distribution information is marked, the current load condition of the neural network system and the data processing amount of the target data are further combined to determine the power consumption adjustment logic, the power consumption control event of the neural network system is determined along the analysis of the power consumption adjustment logic, and the accuracy of the power consumption control event of the neural network system is improved.

Description

Technical Field

[0001] This invention relates to the technical field of neural network systems, and in particular to a power consumption control method and system for a neural network system based on a composite architecture. Background Technology

[0002] With the rapid development of artificial intelligence and edge computing, neural network systems have been widely used in demanding scenarios with high dynamics and low latency, such as autonomous driving chassis control and closed-loop scheduling of industrial robots. In these scenarios, neural network systems are usually deployed on edge chips with limited computing power and power consumption.

[0003] Existing expert systems typically rely on highly complex floating-point matrix multiplication or dense fully connected attention mechanisms for feature matching and routing when faced with input target data. The extensive use of multipliers in the pre-processing stage generates extremely high dynamic power consumption, which contradicts the original intention of low-power design at the edge. At the same time, the system is prone to blindly triggering the response of a large number of expert modules due to weak noise or non-critical data features. This single wake-up method means that the activated expert modules are often not the best nodes to process the current data, which cannot guarantee the accuracy of the activated expert-level recognition modules and affects the accuracy of power control events of the neural network system. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a power consumption control method and system for a neural network system based on a composite architecture.

[0005] This invention provides a power consumption control method for a neural network system based on a composite architecture, comprising: In a neural network system, the dynamic computation of the corresponding computing unit is triggered by the multi-directional broadcast of the target data, and the corresponding data features are marked during the computation process. The corresponding expert-level recognition module dynamically responds along each data feature. The expert-level recognition module performs integration operations on each data feature in different dimensions and determines the corresponding matching integral by combining with the logical sniffing receptor. The matching integral is compared with a preset matching integral threshold, and the corresponding integral comparison result is output. Based on the analysis of the integral comparison result, multiple activation factors are determined. These multiple activation factors are loaded into the corresponding composite architecture, and a sparse attention mechanism is introduced into the composite architecture to dynamically activate the corresponding expert-level recognition module, thereby determining the activated expert-level recognition module. There are multiple activated expert-level recognition modules. The multiple activated expert-level recognition modules load the corresponding interaction matrix and match along the multiple mapping relationships recorded in the interaction matrix to select the final execution module. The system monitors the data processing of the target data by the final execution module in real time and marks the corresponding power distribution information. It then combines the current load of the neural network system and the amount of data processed by the target data to determine the corresponding power adjustment logic. The power control event of the neural network system is determined by parsing the power adjustment logic.

[0006] This invention provides a power control system for a neural network system based on a composite architecture. The power control system for the neural network system based on the composite architecture is applied to the aforementioned power control method for a neural network system based on a composite architecture. The power control system for the neural network system based on the composite architecture includes: The integration module is used in the neural network system to trigger the dynamic calculation of the corresponding computing unit based on the multi-directional broadcast of the target data, and to mark the corresponding data features during the calculation process. It dynamically responds to the corresponding expert-level recognition module along each data feature. The expert-level recognition module performs integration calculations on each data feature in different dimensions and combines the logical sniffing receptor to determine the corresponding matching integral. The composite architecture module is used to compare the matching integral with the preset matching integral threshold and output the corresponding integral comparison result. Based on the analysis of the integral comparison result, multiple activation factors are determined, and the multiple activation factors are loaded into the corresponding composite architecture. A sparse attention mechanism is introduced into the composite architecture to dynamically activate the corresponding expert-level recognition module, thereby determining the activated expert-level recognition module. The filtering module is used to select the final execution module. There are multiple activated expert-level recognition modules. The multiple activated expert-level recognition modules load the corresponding interaction matrix and match along the multiple mapping relationships recorded in the interaction matrix. The power consumption control module is used to monitor the data processing of the target data by the final execution module in real time, mark the corresponding power consumption distribution information, and further determine the corresponding power consumption adjustment logic by combining the current load of the neural network system and the amount of data processing of the target data. The power consumption control event of the neural network system is determined by parsing the power consumption adjustment logic.

[0007] Compared with the prior art, the beneficial effects of the present invention are: (1) In the neural network system, the corresponding computing unit is dynamically calculated based on the multi-directional broadcast of the target data, and the corresponding data features are marked during the calculation process. The corresponding expert-level recognition module is dynamically responded to along each data feature. The expert-level recognition module performs integration operations on each data feature in different dimensions and determines the corresponding matching integral in combination with the logical sniffing receptor. The matching integral is compared with the preset matching integral threshold and the corresponding integral comparison result is output. Multiple activation factors are determined based on the analysis of the integral comparison result. The multiple activation factors are loaded into the corresponding composite architecture and a sparse attention mechanism is introduced into the composite architecture to dynamically activate the corresponding expert-level recognition module, thereby determining the activated expert-level recognition module. The integral comparison result is introduced to further control the multiple activation factors and the composite architecture, thereby improving the accuracy of the activated expert-level recognition module.

[0008] (2) There are multiple activated expert-level recognition modules. The multiple activated expert-level recognition modules load the corresponding interaction matrix and match along the multiple mapping relationships recorded in the interaction matrix to select the final execution module. The data processing of the target data by the final execution module is monitored in real time, and the corresponding power distribution information is marked. The corresponding power adjustment logic is determined by combining the current load of the neural network system and the data processing volume of the target data. The power control event of the neural network system is determined by parsing the power adjustment logic. The power distribution information, the current load of the neural network system and the data processing volume of the target data are fully considered, which improves the accuracy of the power control event of the neural network system. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the power consumption control method for a neural network system based on a composite architecture according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating step S11 in the power consumption control method for a neural network system based on a composite architecture according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating step S12 in the power consumption control method for a neural network system based on a composite architecture according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating step S13 in the power consumption control method for a neural network system based on a composite architecture according to an embodiment of the present invention. Figure 5 This is a flowchart illustrating step S14 in the power consumption control method for a neural network system based on a composite architecture according to an embodiment of the present invention. Figure 6 This is a schematic diagram of the power consumption control system of a neural network system based on a composite architecture in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0011] Please see Figures 1 to 6 A power consumption control method for a neural network system based on a composite architecture is proposed and applied to neural network system scenarios. The power consumption control method for a neural network system based on a composite architecture includes: Step S11: In the neural network system, the dynamic calculation of the corresponding computing unit is triggered by the multi-directional broadcast of the target data, and the corresponding data features are marked during the calculation process. The corresponding expert-level recognition module is dynamically responded to along each data feature. The expert-level recognition module performs integration operations on each data feature in different dimensions and determines the corresponding matching integral by combining the logical sniffing receptor. Step S12: Compare the matching integral with the preset matching integral threshold and output the corresponding integral comparison result. Based on the analysis of the integral comparison result, determine multiple activation factors, load the multiple activation factors into the corresponding composite architecture, and introduce a sparse attention mechanism into the composite architecture to dynamically activate the corresponding expert-level recognition module, thereby determining the activated expert-level recognition module. Step S13: There are multiple activated expert-level recognition modules. The multiple activated expert-level recognition modules load the corresponding interaction matrix and match along the multiple mapping relationships recorded in the interaction matrix to select the final execution module. Step S14: Monitor the data processing of the target data by the final execution module in real time, mark the corresponding power distribution information, and further combine the current load of the neural network system and the data processing volume of the target data to determine the corresponding power adjustment logic. Determine the power control event of the neural network system by parsing the power adjustment logic.

[0012] refer to Figure 2 In step S11, the specific steps are as follows: S111: Label the neural network system and determine the target data in the state to be processed based on the dynamic recognition of the neural network system. The target data is further processed in the neural network system and broadcast in multiple directions along the multi-directional broadcast path of the neural network, thereby triggering the response of each computing unit, so that each computing unit performs dynamic calculation on the same target data to label each calculation process. Based on the dynamic recognition of each calculation process, the corresponding data features are determined. The data features have multiple characteristics. S112: Input multiple data features into the recognition module database and trigger the dynamic response of multiple expert-level recognition modules in the recognition model database. In the response process, the integration operation of each data feature in different dimensions is triggered, and polymorphic logic operation is performed in combination with the logic sniffing receptor of the composite architecture. At this time, the logic sniffing receptor does not need to use a multiplier, but only completes the corresponding matching integral output through the underlying adder and subtractor.

[0013] In the embodiments of this application, a neural network system is labeled, and target data in a state to be processed is determined based on the dynamic recognition of the neural network system. The target data is further processed in the neural network system and broadcast in multiple directions along the multi-directional broadcast path of the neural network, thereby triggering the response of each computing unit. This enables each computing unit to perform dynamic calculations on the same target data to label each calculation process. Based on the dynamic recognition of each calculation process, the corresponding data features are determined. There are multiple data features, which are compatible with the overall consideration of the dynamic recognition of each calculation process, ensuring the accuracy of the corresponding data features.

[0014] At this time, the system scans the global control register of the neural network in real time through the underlying hardware state machine to mark the current steady-state nodes of the neural network. When a new data stream from the external sensor bus causes the gradient or potential of the hidden state of the neural network to jump out of bounds, the system accurately extracts and identifies the target data segment in the "to be processed" state from the continuous data stream based on this dynamic identification process, completing the first interception of data entering the computing domain.

[0015] The identified target data does not directly enter the high-power matrix operation unit, but enters the system's ultra-low-power preprocessing buffer for bit width truncation or alignment. The system bus controller loads the processed target data onto the global multi-directional broadcast path of the neural network, and broadcasts it indiscriminately to all downstream nodes along the physical connections of the network topology in the form of a single drive current.

[0016] Each basic computing unit on the broadcast path synchronously triggers a response after receiving the same broadcast signal. Each unit uses its local lightweight logic gates with extremely low bit width to perform concurrent dynamic calculations on the same target data. During this process, the system clock assigns an independent micro-timestamp tag to the running cycle of each computing unit and binds the tag to the intermediate state value generated by the calculation, thereby completing the explicit marking of each independent calculation process at the hardware level.

[0017] The system feature parser polls and dynamically identifies the intermediate states of each tagged computation process based on the micro-timestamp tags bound in the preceding sequence. By comparing the state flip rate or local extrema, it abstracts and solidifies the computation process results containing different physical meanings into discrete data features. Finally, it outputs multiple orthogonal or complementary data features for the same target data, such as edge mutation features and low-frequency smoothing features, providing multi-dimensional discrimination criteria for subsequent power consumption routing.

[0018] Specifically, in the chassis control system of autonomous vehicles, a continuous-time recurrent neural network is deployed to adjust the vehicle's suspension height in real time. This system operates continuously while the vehicle is in motion, making power consumption control crucial. When the vehicle's front-facing camera and radar detect a pothole on the road surface, the system control register detects a sharp jump in the error potential representing the "expected trajectory" and "actual feedback" within the network. Based on this, the system marks the current network state and identifies this frame, which contains a mixed data packet of the pothole image and the current suspension stress, as target data to be processed. Unlike traditional methods, this target data does not directly trigger the entire neural network's massive floating-point matrix to update cyclically. Instead, it is pushed to the system's global data bus and synchronously sent along a multi-directional broadcast path with extremely low latency to all the underlying computing units in the network responsible for different dimensions such as "road surface smoothness," "vehicle attitude," and "tire grip."

[0019] After receiving the same target data, these computing units simultaneously trigger responses and perform dynamic calculations. For example, the attitude computing unit extracts pitch angle data, and the smoothness computing unit extracts turbulence frequency data. During the calculation, the hardware clock marks the calculation process of these different modules.

[0020] The system's feature analysis logic dynamically identifies these markers, quickly extracting the data feature of "sharp front axle descent" from the attitude calculation process and the data feature of "high-frequency continuous vibration" from the smoothness calculation process. Finally, for this deep pit target data, it outputs multiple high-precision data features. Through this series of operations, the system degrades the complex physical signal into extremely low-bit discrete features in advance without activating any high-power multipliers. This lays an absolutely low-power foundation for subsequently determining which expert modules need to be powered on and which can be directly physically powered off (PowerGating).

[0021] Furthermore, multiple data features are input into the recognition module database, triggering dynamic responses from multiple expert-level recognition modules in the recognition model database. During the response process, integration operations of each data feature in different dimensions are triggered, and polymorphic logic operations are performed in conjunction with the composite architecture logic sniffing receptor. At this point, the logic sniffing receptor does not need to use a multiplier; it only uses the underlying adders and subtractors to complete the corresponding matching integral output. The underlying adders and subtractors are introduced to complete the corresponding matching integral output.

[0022] At this point, the system uses multiple data features as index keys and loads them in parallel into a pre-configured recognition module database. This database is mapped as a distributed storage array at the hardware level. When a specific feature key is detected to match the corresponding storage partition, the system directly triggers the dynamic response of multiple expert-level recognition modules bound to that partition through a hardware interrupt signal, switching them from the default standby listening state to the data ready state.

[0023] Upon receiving the corresponding data features, each awakened expert-level recognition module initiates its internally configured integral accumulation tree to perform integral operations on the data features across different dimensions. This process involves continuously sampling and linearly accumulating the state bits of the features within a fixed clock cycle, transforming discrete feature pulses into integral scalars with definite dimensions. These scalars are then used to evaluate the absolute strength of the feature's weight in the entire network state.

[0024] The logic sniffing receptor built into the composite architecture is directly cascaded with the output of each expert-level recognition module. After receiving the integral scalar, the logic sniffing receptor abandons the traditional floating-point weight matrix and instead polarizes the expert matching template pre-fixed in the register into a three-state vector. The system dynamically maps the integral scalar to "positive excitation state, denoted as 1", "negative excitation state, denoted as -1" or "irrelevant state, denoted as 0" in the three-state logic value according to the positive and negative polarity and amplitude range of the integral scalar, thus completing the conversion of data from the analog quantization domain to the pure digital logic domain.

[0025] During the final matching degree calculation, the system physically shields or bypasses the calculation clock of all floating-point multipliers within the hardware clock cycle; the logic sniffing receptor directly calls the most basic adder and subtractor in the arithmetic logic unit to perform pure accumulation and subtraction operations on the mapped three-state logic values, that is, it performs addition when encountering state 1, performs subtraction when encountering state -1, and remains unchanged when encountering state 0. Finally, it outputs the matching integral scalar representing the fit between the feature and the expert module within an extremely short gate delay.

[0026] Specifically, after the system extracts multiple data features such as "sharp front axle drop" and "high-frequency continuous vibration" in S111, it immediately inputs these features into the identification module database. This database contains expert-level identification modules responsible for different control strategies such as "suspension damping adjustment", "brake torque distribution", and "engine torque compensation". When the "high-frequency continuous vibration" feature enters the database, a hardware interrupt instantly triggers the dynamic response of a few relevant expert modules such as "suspension damping adjustment", while other unrelated modules continue to remain dormant.

[0027] The awakened "suspension damping adjustment" expert module then initiates integral operations in different dimensions for the feature of "high-frequency continuous vibration". For example, it continuously samples the vibration feature for 5 clock cycles in the time dimension to obtain an integral value representing extremely high vibration intensity. This integral value enters the logic sniffing receptor of the composite architecture for polymorphic logic operations. The "deep pit response standard template" pre-stored inside the sniffing receptor is polarized into three states: in the face of high-intensity vibration at this time, the receptor maps the corresponding dimension to the positive excitation state (1) and maps the non-vibration dimension to the irrelevant state (0).

[0028] Because the neural network is in a high-frequency closed-loop state, traditional floating-point multiplication consumes a huge amount of instantaneous current; but at this moment, the logic sniffing receptor does not need to use a multiplier. The hardware scheduler directly cuts off the power supply of the multiplier and only calls the extremely low-power adder and subtractor at the bottom layer; for the above three-state mapping results, the adder directly accumulates the value of the positive excitation state (1), and the subtractor removes irrelevant noise. With just a few pure integer addition and subtraction flips, the corresponding matching integral is output in nanoseconds. This extremely low matching integral not only accurately determines that the "suspension damping adjustment" module is the best candidate to deal with the current crisis, but more importantly, the core high-energy-consuming unit of the system is in a completely zero-power state during the long pre-judgment process from feature matching to integral output.

[0029] refer to Figure 3 In step S12, the specific steps are as follows: S121: Obtain a preset matching integration threshold and compare the matching integration with the threshold to determine the corresponding integration comparison result. The integration comparison result includes cases where the matching integration is greater than the preset matching integration threshold and cases where the matching integration is less than the preset matching integration threshold. S122: Based on the analysis of the integral result, determine the target content whose matching integral is greater than the preset matching integral threshold, and determine the corresponding activation factors by identifying the multi-dimensional aspects of the target content to obtain multiple activation factors; S123: Mark the corresponding composite architecture, determine the corresponding data space based on the dynamic detection of the composite architecture, input multiple activation factors into the same data space, and perform dynamic calculations in combination with the sparse attention mechanism of the data space to dynamically activate the corresponding expert-level recognition module, and then mark the activated expert-level recognition modules in sequence, while the remaining expert-level recognition modules are in a dormant state.

[0030] In the embodiments of this application, a preset matching score threshold is obtained, and the matching score is compared with the score to determine the corresponding score comparison result. The score comparison result covers the matching score being greater than the preset matching score threshold and the matching score being less than the preset matching score threshold. The matching score being greater than the preset matching score threshold and the matching score being less than the preset matching score threshold are introduced.

[0031] At this point, the system obtains the preset matching integration threshold under the current operating condition from the dynamic configuration register of the composite architecture. This threshold is not a fixed constant, but a scalar value generated by the power management unit after dynamic calibration based on the current global remaining power of the neural network, the chip junction temperature, and the historical average activation rate per unit time. After obtaining the threshold, the system preloads it into the hardware flip-flop dedicated to the comparison logic as the reference level for subsequent decisions.

[0032] The system directly inputs the matched integral output from the previous steps through addition and subtraction into the hardware comparator, and performs bit-to-bit parallel difference logic operation with the preset matched integral threshold locked in the flip-flop. This comparison process abandons the conditional branch judgment at the software level, and directly uses the carry flag and sign bit of the arithmetic logic unit to obtain the binary status bit representing the size relationship between the two in a single clock cycle, thus completing the integral comparison without delay.

[0033] The system parses the binary status bits output by the comparator, abstracts them into explicit discrete integral comparison results, and broadcasts them along the global bus. These comparison results are strictly defined as two mutually exclusive polarity states: when the status bit is high, it is explicitly defined as a "pass state" where "the matched integral is greater than the preset matched integral threshold"; when the status bit is low, it is explicitly defined as a "block state" where "the matched integral is less than the preset matched integral threshold". This serves as the sole Boolean logic input for subsequent hardware power gating actions.

[0034] Specifically, after the expert-level recognition modules such as "suspension damping adjustment" rapidly output the "matching integral" representing the feature fit through pure addition and subtraction, the system executes the acquisition of the preset matching integral threshold. Considering that the vehicle is in a complex continuous closed-loop control and the battery is not fully charged, in order to prevent excessive wake-up from causing computing power overload, the power management unit dynamically adjusts and calibrates the original threshold from the reference value "50" to "65", and loads this threshold "65", the preset matching integral threshold, to the reference end of the underlying hardware comparator.

[0035] The system sends the matching integral calculated by the "suspension damping adjustment" module, assuming it is "80", to the comparator and performs a hard-wired integral comparison with the threshold "65". Since there is no software judgment jump overhead, the comparator obtains the size relationship through the borrow and carry flags of the subtractor.

[0036] The system parses the flag and determines the corresponding integral comparison result. Because "80" is numerically dominant, the parsing logic directly outputs a high-level signal, clearly defining the current integral comparison result as "the matching integral is greater than the preset matching integral threshold". This extremely clear binary result not only announces the result of the "suspension damping adjustment" module in the sniffing phase, but also, at the same time, the matching integral calculated by other expert modules responsible for "engine torque compensation" is only "30", and the system synchronously parses its comparison result as "the matching integral is less than the preset matching integral threshold".

[0037] Furthermore, based on the analysis of the integration result, the target content whose matching integral is greater than the preset matching integral threshold is determined. The corresponding activation factors are determined by multi-dimensional identification of the target content to obtain multiple activation factors. This approach incorporates the overall consideration of the integration result analysis and ensures the accuracy of the target content whose matching integral is greater than the preset matching integral threshold.

[0038] At this point, the system captures the high-level state bit of the output of the preceding comparator through the edge triggering mechanism, performs polarity analysis on the integral comparison result, accurately eliminates invalid interception signals that present a low-level state, and only maps the valid pass signal that represents "the matching integral is greater than the preset matching integral threshold" into a concrete target content. This target content is encapsulated at the hardware level as a data packet containing the physical identifier of the winning expert module, the timestamp of the feature response, and the corresponding integral overflow level, thereby uniquely locking the candidate object with high-priority calculation authority in the huge network cluster.

[0039] The system initiates multi-dimensional feature decoupling logic for the locked target content, instead of treating it as a single pass instruction, it performs a dimensionality reduction scan along the data topology structure contained within the target content; the system extracts the target content's feature response latency rate in the time dimension, feature distribution sparsity in the spatial dimension, and node activation frequency in the historical dimension, respectively. By dispersing the single integration result into multiple orthogonal evaluation dimensions, a multi-dimensional evaluation vector is constructed for subsequent refined power consumption allocation.

[0040] The system inputs the decoupled multidimensional evaluation vector into a preset nonlinear mapping lookup table, and concretizes the abstract dimensional parameters into control instructions that can be directly recognized by the execution unit. Each dimension parameter mapping generates an independent micro-control signal, which includes local voltage boost requests, clock frequency scaling factors, and data path gating signals. The system defines these micro-control signals with actual hardware driving capabilities as activation factors, and packages and queues multiple activation factors derived from the same target content to complete the batch acquisition of activation factors.

[0041] Specifically, after the system obtains the comparison result that the matching integral (80) of the "suspension damping adjustment" module is greater than the threshold (65), the system performs polarity analysis on this integral comparison result, directly filters out the low-level invalid signals such as "engine torque compensation (integral 30)" generated at the same time, and converts the valid information corresponding to the high-level signal into specific target content; at this time, the target content is no longer an abstract fraction, but is precisely locked by the system into a set of high-priority data packets containing "expert module ID=suspension damping_03" and "response delay=2 clock cycles".

[0042] After locking onto the target content, the system did not rush to power it on, but instead performed multi-dimensional identification along the target content. The system analysis found that: in the time dimension, the feature responded extremely quickly; in the spatial dimension, other control requests on the current bus were very sparse, and the vehicle was in a straight-line state; in the historical dimension, the "suspension damping_03" module had never been activated in the past second. The scanning results of these three dimensions constituted a three-dimensional portrait of the target content.

[0043] The system maps this 3D profile to determine the corresponding activation factors and acquire multiple activation factors. For "extremely fast response", the system generates a signal for "activation factor A: multiply the local clock frequency to the highest frequency". For "sparse bus", the system generates a signal for "activation factor B: fully open the data bus gating gate of this module". For "historical cold start", the system generates a signal for "activation factor C: precharge its internal registers to eliminate startup delay". Through this series of operations, the system not only knows "who to wake up", but also prepares in advance multiple specific hardware control methods for "how to wake it up in the most extreme way and in line with the current closed-loop physical laws". Thus, while acquiring these multiple activation factors, it avoids the transient current impact and useless power consumption caused by blind power-on.

[0044] Therefore, the corresponding composite architecture is labeled, and the corresponding data space is determined based on the dynamic detection of the composite architecture. Multiple activation factors are input into the same data space, and dynamic calculations are performed using the sparse attention mechanism of the data space to dynamically activate the corresponding expert-level recognition modules. Then, the activated expert-level recognition modules are labeled sequentially, while the remaining expert-level recognition modules are in a dormant state. This approach takes into account the overall consideration of dynamic detection of composite architectures, ensuring the accuracy of the corresponding data space. At the same time, integral comparison results are introduced to further control multiple activation factors and composite architectures, thereby improving the accuracy of the activated expert-level recognition modules.

[0045] At this point, the system sends a status set instruction to the specific composite architecture carrying the current computing task through the bus arbiter to complete the exclusive marking of the composite architecture; the system dynamically detects the cache queue and cross switch matrix inside the marked composite architecture, and dynamically delineates a temporary isolation area with fixed physical boundaries in the hardware topology based on the physical address of the currently free storage block and the available interconnect bandwidth. This isolation area is defined and mapped as the exclusive data space for the current round to prevent bus congestion or crosstalk during the transmission of activation signaling.

[0046] The system injects multiple activation factors, including clock, voltage, and gating dimensions, obtained from the previous steps, into the defined data space via parallel data lines. Inside the data space, the system calls the sparse attention mechanism computation kernel embedded in SRAM. This mechanism does not perform fully connected dense attention matrix operations, but instead uses the injected activation factors as query keys to quickly locate a very small number of relevant nodes in the data space through index routing. It assigns high weight coefficients to these hit nodes and forces the weights of the remaining unhit nodes to zero, generating an extremely sparse attention mask.

[0047] The system uses the generated sparse attention mask as a hardware-level enable signal and distributes it along the output channel of the data space. Only when the physical address of the expert-level recognition module precisely corresponds to the non-zero high-weight coefficient in the mask is the power gating inside the module released, achieving precise dynamic activation. At the same time, the system assigns an incrementing sequence number to each module that successfully unlocks from hibernation and burns the sequence number into the status register of the corresponding module to complete the sequential marking, thereby establishing the absolute execution priority of the activated module in the subsequent data pipeline.

[0048] While the mask is being distributed, the system sends a low-level sustain signal to all expert-level recognition modules in the data space that have not been assigned non-zero weights, forcibly shutting down the clock trees of these modules and keeping their power gating switches in an open-locked state. This ensures that the remaining expert-level recognition modules are completely isolated from the computation topology during the current control cycle of the neural network, maintaining an absolute dormant state with zero leakage current.

[0049] Specifically, the system has packaged multiple activation factors for the "Suspension Damping_03" module, such as frequency multiplication and pre-charging. Next, it enters the substantive hardware resource scheduling and wake-up phase. The system marks the corresponding composite architecture in the underlying FPGA or ASIC chip, that is, the computing power cluster area responsible for chassis suspension control. Through dynamic detection, it finds that a certain high-speed SRAM in this area is in an idle state. Therefore, it immediately defines this SRAM and its surrounding interconnect bus at the physical level to construct a closed data space to ensure that the subsequent wake-up command will not be interfered with by other parallel autonomous driving tasks.

[0050] The system inputs multiple activation factors, including "frequency multiplication" and "pre-charging", into the same data space. At this time, the sparse attention mechanism embedded in the data space starts to operate. Instead of traversing all modules in the cluster, it directly uses these activation factors as clues to perform precise indexing, determine that only the "suspension damping_03" module is highly relevant to the current crisis handling, and generate a sparse mask with a weight of 1 for it, while the attention weights of modules such as "vehicle stability_07" that are also in the same cluster are directly counted as 0.

[0051] Based on this sparse mask, the system begins to dynamically activate the corresponding expert-level recognition modules; the high-weight signals in the mask directly reach the power control unit of "Suspension Damping_03", releasing its physical power-off state, and the system immediately marks it with the sequential mark "Sequence Number 001", meaning that it has the highest priority and immediately takes over the suspension bus; at the same time, since the sparse attention mechanism determines that the weight of modules such as "Vehicle Stability_07" is 0, the system does not generate any wake-up level, and these remaining expert-level recognition modules continue to be in a deep sleep state.

[0052] refer to Figure 4 In step S13, the specific steps are as follows: S131: Obtain multiple activated expert-level recognition modules, load the multiple activated expert-level recognition modules into the same matching space, and perform calculations in combination with the interaction matrix recorded in the matching space, thereby performing dynamic matching along multiple mapping relationships in the interaction matrix. At this time, the multiple mapping relationships include enhancement mapping relationships, suppression mapping relationships, and complementary mapping relationships. S132: During the dynamic matching process of multiple activated expert-level recognition modules, the matching information of each activated expert-level recognition module is marked, thereby determining the final execution module by filtering multiple matching information.

[0053] In the embodiments of this application, multiple activated expert-level recognition modules are obtained, and the multiple activated expert-level recognition modules are loaded into the same matching space. The interaction matrix recorded in the matching space is used for calculation, thereby performing dynamic matching along multiple mapping relationships in the interaction matrix. At this time, the multiple mapping relationships include enhancement mapping relationships, suppression mapping relationships and complementary mapping relationships. Enhancement mapping relationships, suppression mapping relationships and complementary mapping relationships are introduced.

[0054] At this point, the system reads the status registers marked sequentially in the previous steps to globally obtain the physical identifiers of all expert-level identification modules that have been de-gated by power. The system sends a space allocation micro-instruction to the underlying memory controller to allocate a mutually exclusive dedicated storage area in the shared cache. Through address mapping, the output cache pointers of these expert-level identification modules scattered on different computing cores are uniformly redirected to this area, completing the logical co-address loading, thereby constructing a matching space isolated from external background noise.

[0055] The system retrieves a pre-configured interaction matrix from a non-volatile storage medium and loads it into the lookup table array of the matching space. This interaction matrix discards dense floating-point weights and is compressed into a sparse Boolean graph containing only the connection topology between expert modules or an integer indicator vector with very low bits. Before the matching operation starts, the system eliminates invalid paths with absolute weight values ​​of zero in the matrix through masking operations, retaining only the topology connections with non-zero indicator values ​​to eliminate bus flip power consumption caused by invalid addressing.

[0056] The system drives the routing logic within the matching space, performing parallel table lookups and dynamic matching along multiple non-zero topological connections retained in the preprocessed interaction matrix. During this process, the system strictly decouples multiple mapping relationships into three mutually exclusive hardware control logics based on the sign and direction bits of each indicator vector in the interaction matrix: when the indicator vector represents positive superposition, it is defined as an enhancement mapping relationship, triggering arithmetic addition merging of the corresponding module's output values; when the indicator vector represents negative cancellation, it is defined as a suppression mapping relationship, triggering arithmetic subtraction truncation of the corresponding module's output values; when the indicator vector represents heterogeneous feature splicing, it is defined as a complementary mapping relationship, triggering bit-by-bit splicing operation of the corresponding module's output features in the bit-width dimension.

[0057] Specifically, after a few expert-level recognition modules, such as "Suspension Damping_03", are successfully awakened and marked by the sparse attention mechanism, the system acquires multiple activated expert-level recognition modules. Assuming that in a short period of time, due to continuous vibration caused by the deep pit, in addition to "Suspension Damping_03", the "Suspension Height_01" module also exceeds the threshold and is activated, the system immediately redirects the calculation output pointers of these two modules and loads them into the same matching space to ensure that their data exchange does not go through the external bus, thus minimizing communication power consumption.

[0058] The system retrieves a unique interaction matrix specific to the chassis control system. This matrix records the physical rules governing which suspension modules should follow and which to defend against when dealing with complex road conditions. To save power, the system only extracts non-zero associated paths from the matrix, discarding a large number of irrelevant and invalid calculations. The system performs dynamic matching along the interaction matrix and precisely executes multiple mapping relationships based on the built-in symbol indicators within the matrix. First, the system detects that "suspension damping_03" and "suspension height_01" need to exert force simultaneously when dealing with deep pits, so it triggers the enhanced mapping relationship and uses an adder to superimpose the output characteristics of the two as a control signal.

[0059] Second, the system simultaneously detects a conflict between the current force direction of "suspension damping_03" and a weak wake-up signal responsible for "comfort smoothing" in the matrix. In order to prevent the control oscillation from wasting power, the system immediately triggers the suppression mapping relationship and forcibly subtracts and cancels the "comfort smoothing" signal from the current calculation path through the subtractor.

[0060] Third, the system determines that "suspension height_03" provides shock absorber oil pressure data, while "suspension damping_03" provides valve opening data. Since they are different physical quantities, a complementary mapping relationship is triggered. Without adding any multiplication calculations, the 1-bit features of the two are directly concatenated bit by bit at the register bit width level. Through these three extremely lightweight logic mappings, the system completes complex multi-module collaborative decision-making that can only be achieved by multiplying traditional large floating-point matrices at a microwatt power consumption level.

[0061] Furthermore, during the dynamic matching process of multiple activated expert-level recognition modules, the matching information of each activated expert-level recognition module is marked, thereby determining the final execution module by filtering multiple matching information. This approach takes into account the overall consideration of multiple matching information and ensures the accuracy of the final execution module.

[0062] At this time, during the concurrent table lookup and logical operation of multiple activated expert-level recognition modules along the interaction matrix, the system is configured with a lightweight micro-operation tracking register to capture and mark the behavioral data of each module in the matching process in real time. The marking action specifically records the enhanced cumulative peak value, suppression cancellation amplitude, and complementary splicing bit width utilization rate generated by each expert-level recognition module, and encapsulates these heterogeneous dynamic data into matching information with timestamps and source node identifiers, which is then fixed in the cache line of the matching space.

[0063] The system inputs multiple matching information into a preset arbitration sorting logic, abandoning the traditional screening method that uses absolute numerical value as the sole criterion. Instead, it introduces a joint evaluation function of "feature focus" and "expected power consumption". The arbitration logic analyzes the enhanced peak value and bit width utilization in the matching information one by one, calculates the focus concentration of the feature represented by each information in the current control cycle, and removes matching information that, although the value is high, will cause subsequent large bit width operations. In this way, it quickly converges the feature pointer with the highest energy efficiency ratio among multiple matching information.

[0064] Based on the source node identifier of the unique or very few core matching information retained after filtering, the system traces back to the physical computing unit that generated the information along the reverse pointer of the matching space; the system sends the final strobe latch signal to the physical computing unit, forcibly closes the output buffer of other activated modules, and absolutely converges the control authority of the entire neural network system to the single node at the current moment, formally establishing its logical identity as the only final execution module with bus write authority.

[0065] Specifically, while modules such as "Suspension Damping_03" and "Suspension Height_01" are dynamically mapped in the matching space through enhancement, suppression, and complementation, the system must decide which module will make the final decision in a very short time to stop meaningless power consumption from collaborative calculations. The system uses low-level tracking registers to mark the matching information of each activated expert-level recognition module in real time. For example, the system records that "Suspension Damping_03" accumulates an extremely high peak signal of "valve needs to be closed instantly" during enhancement mapping, and at the same time records that "Suspension Height_01" generates a normal bit width information of "airbag needs to maintain its current state" during complementary splicing. These records, which contain signal strength and data type information, constitute their respective matching information.

[0066] The system screened multiple matching information. Although the cumulative peak value of "suspension damping_03" was extremely high, the system evaluation found that if the calculation continued down from its peak value, it might be necessary to wake up the huge and power-consuming continuous-time differential calculation network in the backend. In contrast, although the value of "suspension height_01" was flat, its feature focus was extremely high, directly pointing to the most critical physical support requirement. Moreover, the bit width required for subsequent calculations was extremely narrow, requiring only a simple 1-bit switch output. Based on this joint evaluation of "feature focus and expected power consumption", the system decisively selected the matching information corresponding to "suspension height_01" and eliminated "suspension damping_03".

[0067] The system traces back to the source address of the matched information selected and sends a strobe latch signal to the "suspension height_01" module. At the same time, it cuts off the data transmission path from "suspension damping_03" to the chassis bus, thus determining that the final execution module is "suspension height_01". Through this rapid convergence from "multi-module dynamic negotiation" to "single-module dictatorship", the system avoids the waste of computing power and power caused by multiple modules sending conflicting commands to the chassis actuator at the same time.

[0068] refer to Figure 5 In step S14, the specific steps are as follows: S141: The final execution module processes the target data and monitors the data processing process in real time. It synchronously marks multiple power consumption regions of the target data during the data processing process and determines the corresponding power consumption distribution information based on the reverse tracing of each power consumption region. S142: Mark the current load of the neural network system, construct the corresponding functional adjustment framework based on the multi-factor fusion of the current load of the neural network system and the corresponding power distribution information, further determine the corresponding power adjustment logic by combining the data processing volume of the target data, perform dynamic analysis on the determined power adjustment logic, and determine the power control event of the neural network system by combining the corresponding weight allocation information during the analysis process.

[0069] In the embodiments of this application, the final execution module processes the target data and monitors the data processing process in real time. It synchronously marks multiple power consumption regions of the target data during the data processing process and determines the corresponding power consumption distribution information based on the reverse tracing of each power consumption region. This approach is compatible with the overall consideration of reverse tracing of each power consumption region and ensures the accuracy of the corresponding power consumption distribution information.

[0070] At this point, the locked final execution module receives the target data and fully activates its internal bit-linear layer computing unit to perform actual data processing. During this process, the system strictly shields high-precision floating-point operations. The final execution module only calls extremely low bit weights of 1-bit or 1.58-bit and performs layer-by-layer nonlinear deduction and feature mapping on the target data through pure integer adders and shifters, completing the data inference calculation under the physical constraint of ensuring extremely low dynamic power consumption.

[0071] Simultaneously with the bit-based derivation performed by the final execution module, the system activates a hardware performance counter array distributed in various logic clusters within the final execution module. This array collects instantaneous current fluctuations, arithmetic logic unit toggles, and register file read / write frequencies in real time during each calculation cycle at a frequency synchronized with the data processing clock, forming a fine-grained power consumption timing trajectory that accompanies the movement of the data stream.

[0072] The system divides the data processing process into regions based on the synchronously acquired power consumption timing trajectory and the layout coordinates of the physical logic units inside the final execution module. The system spatiotemporally binds the calculation cycles that generate high flip-flop rates and high current spikes with specific hardware macrocells. When the power consumption sampling value of a certain region exceeds the micro threshold in a continuous clock cycle, the system immediately adds a spatial label, thereby dynamically dividing multiple discrete power consumption regions on the entire data processing time axis.

[0073] For the marked multiple power consumption areas, the system retrieves the global integrated routing diagram of the composite architecture, starts from the coordinates of each high-power physical node, and performs logical backtracking along the reverse path of the data flow. The system associates and packages the specific network layer parameters, activation sparsity of specific weight groups, and bandwidth utilization of local buses associated with the backtracking path, strips out unrelated background noise, and finally aggregates to generate a power distribution information that is accurately mapped to the specific algorithm operators and hardware microarchitecture level.

[0074] Specifically, the "suspension height_01" module begins to process the target data. After receiving the feature data of the deep pit, it does not use any traditional floating-point multiplication, but instead calls the internal ultra-low power 1.58-bit BitLinear structure. Relying purely on the underlying adders and shifters, it quickly calculates the integer control command that "the front axle airbag needs to be inflated and raised by 20 millimeters instantly". This is the only part of the entire neural network that truly consumes energy for inference.

[0075] At the same instant this reasoning occurred, the system initiated real-time monitoring; the hardware performance counters deployed within the module recorded synchronously with nanosecond-level resolution, without the slightest delay: during the clock cycles that calculated the "airbag inflation volume," a sudden spike appeared in the current of the specific adder group responsible for accumulating the height difference.

[0076] After the system detects this spike, it immediately marks the power consumption region. The system determines that the current spike originates from a local integer accumulation array with physical coordinates (X:4, Y:12). Therefore, the system marks this moment in the spatiotemporal record of data processing, clearly dividing the calculation process at this moment into a "high power consumption region". The other register regions that only perform data pass-through or zero value skipping are marked as "extremely low power consumption regions". At this time, multiple power consumption regions of different levels are marked in a scattered pattern along the complete path of data processing.

[0077] For the "high-power region" (X:4, Y:12) just marked, the system doesn't look at its current current consumption, but instead traces its origins back to its roots. By examining the structured wiring diagram, it discovers that this region consumes power because a specific weight matrix in the third layer of the network failed to produce zero-value sparsity when processing high-frequency signals from deep pits, causing a large number of logic gates to flip simultaneously. The system extracts this discovery and combines it with the tracing results of other low-power regions to generate extremely detailed power distribution information. This information clearly tells the system which specific network layer and which specific set of weights caused the computational overhead when dealing with deep pits, providing a surgical-level basis for the next step of system-level self-adjustment.

[0078] Furthermore, the current load of the neural network system is marked, and a corresponding functional adjustment framework is constructed based on the fusion of multiple factors, including the current load and the corresponding power distribution information. The corresponding power adjustment logic is then determined by combining the data processing volume of the target data. This determined power adjustment logic is dynamically analyzed, and during the analysis process, the power control event of the neural network system is determined by combining the corresponding weight allocation information. This approach takes into account the overall data processing volume of the target data, ensuring the accuracy of the corresponding power adjustment logic. Simultaneously, by fully considering the power distribution information, the current load of the neural network system, and the data processing volume of the target data, the accuracy of the power control event of the neural network system is improved.

[0079] At this point, the system moves away from a single microscopic execution perspective and scans all modules within the neural network system that are not physically powered off through a global bus monitor. It then performs real-time statistics on the number of data streams occupying the cross-switch matrix in parallel within the current cycle, the congestion depth of the shared cache queue, and the bandwidth utilization of the external memory. The system normalizes these multi-dimensional macroscopic operating indicators and maps them into a continuous state vector that characterizes the overall system pressure, thereby accurately marking the current load status of the neural network system.

[0080] The system inputs the micro-level power consumption distribution information obtained in the previous steps and the currently labeled macro-level load state vector into a preset joint decision network. Instead of performing traditional linear superposition, this decision network constructs a two-dimensional tensor power consumption regulation framework, performs cross-attention operations on the micro-level "where power is consumed the most" and the macro-level "whether the system is congested", and generates a joint response surface with spatial location constraints and global time constraints, which serves as the physical boundary for subsequent regulation actions.

[0081] The system obtains the absolute data processing volume of the target data being processed by the current final execution module, such as the time sequence length or feature map resolution, and injects it as a penalty factor or scaling factor into the aforementioned power regulation framework. If the data processing volume is large, the framework tends to trigger vertical frequency reduction extension logic; if the processing volume is small, it triggers horizontal acceleration breakthrough logic. The system decodes and dynamically parses the instantiated regulation logic line by line, transforming it from abstract policy semantics into specific microarchitecture control parameters, such as frequency division coefficient and voltage island level.

[0082] After parsing out the specific control parameters, the system retrieves the weight allocation information of each sub-network in the composite architecture, that is, the importance ratio of different control modules in the overall chassis control. The system performs safety verification and clamping operations on the control parameters and weight allocation information, suppressing extreme voltage drop actions for core paths with extremely high weights, and amplifying the voltage drop and power-off amplitude for non-core paths with extremely low weights. These microarchitecture control parameters, which have been corrected by safety clamping, are packaged and transformed into enable level toggling signals that can be directly executed by the underlying hardware, and established as the final power consumption control event of the neural network system.

[0083] Specifically, the "suspension height_01" module is frantically calculating the airbag inflation volume, and the system has found that the high power consumption is caused by the lack of sparse weights in the third layer. At this point, the system must immediately decide how to adjust the computing power distribution of the entire vehicle and mark the current load status. The system scans the entire vehicle's neural network and finds that at the same time, the "LiDAR point cloud processing" network for autonomous driving is also receiving a massive influx of data, and the shared cache is extremely congested. The system's current macroscopic load is marked as "extremely high pressure and intense resource competition".

[0084] The system combines the microscopic discovery of "high power consumption of the third layer weight" and the macroscopic discovery of "extreme cache congestion" to construct an adjustment framework. The conclusion of this framework is that it is absolutely impossible to accelerate deep pit calculation by increasing voltage frequency (DVFS), otherwise it will compete with the radar network for power and cause the system to thermally crash. The strategy of "reducing bypass computing power in exchange for smooth operation of individual units" must be adopted.

[0085] The system determines that the current deep pit data belongs to high-frequency continuous vibration and the data processing volume is huge. Combining the previous framework, the system instantiates and dynamically parses a specific power consumption adjustment logic: "In order to ensure that the long sequence calculation of 'suspension height_01' does not lag, it is necessary to immediately forcibly clear the non-core calculation queue around it and appropriately reduce the running frequency of its own third-level addition tree."

[0086] The system consulted the safety dictionary of the chassis composite architecture and found that "suspension height_01" is directly related to vehicle bottoming out. Its weight allocation information belongs to "highest safety level, weight ratio 90%", and it is absolutely not allowed to perform any frequency reduction action on itself that may cause calculation delay. Therefore, the system modified the parsing result, and transferred all frequency reduction actions to the "chassis comfort filter" module with a very low weight, accounting for 5%. The system sent a series of enable level signals to the underlying hardware to determine the power consumption control event of the neural network system as: "instantly physically cut off the power supply of the 'chassis comfort filter' module, inject all the electrical energy released by it into the shared buffer to relieve congestion, and at the same time maintain the full frequency operation of 'suspension height_01'".

[0087] Please see Figure 6 , Figure 6 This is a schematic diagram of the structural composition of the power control system for a neural network system based on a composite architecture according to an embodiment of the present invention; the power control system for the neural network system based on the composite architecture is applied to the power control method for the aforementioned neural network system based on a composite architecture; the power control system for the neural network system based on the composite architecture includes: The integral operation module 21 is used in the neural network system to trigger the dynamic calculation of the corresponding computing unit according to the multi-directional broadcast of the target data, and to mark the corresponding data features during the calculation process. It dynamically responds to the corresponding expert-level recognition module along each data feature. The expert-level recognition module performs integral operations on each data feature in different dimensions and determines the corresponding matching integral by combining the logical sniffing receptor. The composite architecture module 22 is used to compare the matching integral with the preset matching integral threshold and output the corresponding integral comparison result. Based on the analysis of the integral comparison result, multiple activation factors are determined, and the multiple activation factors are loaded into the corresponding composite architecture. A sparse attention mechanism is introduced into the composite architecture to dynamically activate the corresponding expert-level recognition module, thereby determining the activated expert-level recognition module. The filtering module 23 is used to select the final execution module by having multiple activated expert-level recognition modules. The multiple activated expert-level recognition modules load the corresponding interaction matrix and match along the multiple mapping relationships recorded in the interaction matrix. The power consumption control module 24 is used to monitor the data processing of the target data by the final execution module in real time, mark the corresponding power consumption distribution information, and further determine the corresponding power consumption adjustment logic by combining the current load of the neural network system and the amount of data processing of the target data. The power consumption control event of the neural network system is determined by parsing the power consumption adjustment logic.

[0088] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A power consumption control method for a neural network system based on a composite architecture, characterized in that, include: In a neural network system, the dynamic computation of the corresponding computing unit is triggered by the multi-directional broadcast of the target data, and the corresponding data features are marked during the computation process. The corresponding expert-level recognition module dynamically responds along each data feature. The expert-level recognition module performs integration operations on each data feature in different dimensions and determines the corresponding matching integral by combining with the logical sniffing receptor. The matching integral is compared with a preset matching integral threshold, and the corresponding integral comparison result is output. Based on the analysis of the integral comparison result, multiple activation factors are determined. These multiple activation factors are loaded into the corresponding composite architecture, and a sparse attention mechanism is introduced into the composite architecture to dynamically activate the corresponding expert-level recognition module, thereby determining the activated expert-level recognition module. There are multiple activated expert-level recognition modules. The multiple activated expert-level recognition modules load the corresponding interaction matrix and match along the multiple mapping relationships recorded in the interaction matrix to select the final execution module. The system monitors the data processing of the target data by the final execution module in real time and marks the corresponding power distribution information. It then combines the current load of the neural network system and the amount of data processed by the target data to determine the corresponding power adjustment logic. The power control event of the neural network system is determined by parsing the power adjustment logic.

2. The power consumption control method for a neural network system based on a composite architecture according to claim 1, characterized in that, In the neural network system, dynamic computation of corresponding computing units is triggered based on the multi-directional broadcast of target data. During the computation process, corresponding data features are marked, and expert-level recognition modules dynamically respond along each data feature. These expert-level recognition modules perform integration operations on each data feature in different dimensions and determine the corresponding matching integral by combining with a logical sniffing receptor, including: The neural network system is labeled, and the target data in the state to be processed is determined based on the dynamic recognition of the neural network system. The target data is further processed in the neural network system and broadcast in multiple directions along the multi-directional broadcast path of the neural network, thereby triggering the response of each computing unit, so that each computing unit performs dynamic calculation on the same target data to label each calculation process. Based on the dynamic recognition of each calculation process, the corresponding data features are determined, and there are multiple data features.

3. The power consumption control method for a neural network system based on a composite architecture according to claim 2, characterized in that, In the neural network system, dynamic computation of corresponding computing units is triggered based on the multi-directional broadcast of target data. During the computation process, corresponding data features are marked, and expert-level recognition modules dynamically respond along each data feature. These expert-level recognition modules perform integration operations on each data feature in different dimensions and determine the corresponding matching integral by combining with a logical sniffing receptor. The system also includes: Multiple data features are input into the recognition module database, triggering dynamic responses from multiple expert-level recognition modules in the recognition model database. During the response process, integration operations of each data feature in different dimensions are triggered, and polymorphic logic operations are performed in conjunction with the logic sniffing receptor of the composite architecture. At this time, the logic sniffing receptor does not need to use a multiplier, but only uses the underlying adders and subtractors to complete the corresponding matching integral output.

4. The power consumption control method for a neural network system based on a composite architecture according to claim 1, characterized in that, The process involves comparing the matching integral with a preset matching integral threshold and outputting the corresponding integral comparison result. Based on the analysis of this result, multiple activation factors are determined. These activation factors are then loaded into the corresponding composite architecture, and a sparse attention mechanism is introduced into the composite architecture to dynamically activate the corresponding expert-level recognition module. This process determines the activated expert-level recognition module, including: Obtain a preset matching score threshold and compare the scores with the preset matching score to determine the corresponding score comparison result. The score comparison result covers the cases where the matching score is greater than the preset matching score threshold and the matching score is less than the preset matching score threshold. Based on the analysis of the integral result, the target content whose matching integral is greater than the preset matching integral threshold is determined. The corresponding activation factors are determined by multi-dimensional identification of the target content to obtain multiple activation factors.

5. The power consumption control method for a neural network system based on a composite architecture according to claim 4, characterized in that, The process of comparing the matching integral with a preset matching integral threshold and outputting the corresponding integral comparison result, determining multiple activation factors based on the analysis of the integral comparison result, loading the multiple activation factors into the corresponding composite architecture, and introducing a sparse attention mechanism into the composite architecture to dynamically activate the corresponding expert-level recognition module, thereby determining the activated expert-level recognition module, also includes: The corresponding composite architecture is labeled, and the corresponding data space is determined based on the dynamic detection of the composite architecture. Multiple activation factors are input into the same data space, and dynamic calculation is performed in combination with the sparse attention mechanism of the data space to dynamically activate the corresponding expert-level recognition module. Then, the activated expert-level recognition modules are labeled in sequence, while the remaining expert-level recognition modules are in a dormant state.

6. The power consumption control method for a neural network system based on a composite architecture according to claim 1, characterized in that, The activated expert-level recognition module has multiple modules. Each activated expert-level recognition module loads a corresponding interaction matrix and performs matching along multiple mapping relationships recorded in the interaction matrix to filter out the final execution module, including: Multiple activated expert-level recognition modules are obtained, loaded into the same matching space, and combined with the interaction matrix recorded in the matching space for calculation, thereby performing dynamic matching along multiple mapping relationships in the interaction matrix. At this time, the multiple mapping relationships include enhancement mapping relationships, suppression mapping relationships, and complementary mapping relationships.

7. The power consumption control method for a neural network system based on a composite architecture according to claim 6, characterized in that, The activated expert-level recognition module has multiple modules. Each activated expert-level recognition module loads a corresponding interaction matrix and performs matching along multiple mapping relationships recorded in the interaction matrix to filter out the final execution module. The module also includes: During the dynamic matching process of multiple activated expert-level recognition modules, the matching information of each activated expert-level recognition module is marked, thereby determining the final execution module by filtering multiple matching information.

8. The power consumption control method for a neural network system based on a composite architecture according to claim 1, characterized in that, The real-time monitoring final execution module processes the target data and marks the corresponding power distribution information. It further combines the current load of the neural network system and the amount of data processed for the target data to determine the corresponding power adjustment logic. Following the parsing of this power adjustment logic, it determines the power control events of the neural network system, including: The final execution module processes the target data and monitors the data processing process in real time. It simultaneously marks multiple power consumption regions of the target data during the data processing and determines the corresponding power consumption distribution information based on the reverse tracing of each power consumption region.

9. The power consumption control method for a neural network system based on a composite architecture according to claim 8, characterized in that, The real-time monitoring final execution module processes the target data and marks the corresponding power distribution information. It further combines the current load of the neural network system and the amount of data processed for the target data to determine the corresponding power adjustment logic. Following the parsing of this power adjustment logic, it determines the power control event of the neural network system. The module also includes: The current load of the neural network system is marked. Based on the current load of the neural network system and the corresponding power distribution information, a corresponding functional adjustment framework is constructed by multi-factor fusion. The corresponding power adjustment logic is further determined by combining the data processing volume of the target data. The corresponding power adjustment logic is dynamically analyzed, and the power control event of the neural network system is determined by combining the corresponding weight allocation information during the analysis process.

10. A power consumption control system for a neural network system based on a composite architecture, characterized in that, The power control system of the neural network system based on the composite architecture is applied to the power control method of the neural network system based on the composite architecture as described in any one of claims 1-9; The power consumption control system of the neural network system based on the composite architecture includes: The integration module is used in the neural network system to trigger the dynamic calculation of the corresponding computing unit based on the multi-directional broadcast of the target data, and to mark the corresponding data features during the calculation process. It dynamically responds to the corresponding expert-level recognition module along each data feature. The expert-level recognition module performs integration calculations on each data feature in different dimensions and combines the logical sniffing receptor to determine the corresponding matching integral. The composite architecture module is used to compare the matching integral with the preset matching integral threshold and output the corresponding integral comparison result. Based on the analysis of the integral comparison result, multiple activation factors are determined, and the multiple activation factors are loaded into the corresponding composite architecture. A sparse attention mechanism is introduced into the composite architecture to dynamically activate the corresponding expert-level recognition module, thereby determining the activated expert-level recognition module. The filtering module is used to select the final execution module. There are multiple activated expert-level recognition modules. The multiple activated expert-level recognition modules load the corresponding interaction matrix and match along the multiple mapping relationships recorded in the interaction matrix. The power consumption control module is used to monitor the data processing of the target data by the final execution module in real time, mark the corresponding power consumption distribution information, and further determine the corresponding power consumption adjustment logic by combining the current load of the neural network system and the amount of data processing of the target data. The power consumption control event of the neural network system is determined by parsing the power consumption adjustment logic.

Images