25results about How to "Reduce computing latency" patented technology

The invention discloses a large model reasoning hardware accelerator based on data flow execution, and belongs to the technical field of large model reasoning hardware acceleration. Each node comprises a linear calculation kernel used for sequentially completing query / key / value projection and attention output projection in multi-head attention calculation, dimension raising projection and dimension reduction projection calculation in a feedforward neural network layer, and inverse quantization operation after each projection is output; the multi-head attention calculation kernel is used for completing loading and quantization of query / key / value vectors, multi-head division and cache management, attention score calculation, Softmax normalization, context vector generation and KV cache updating; the RLA kernel is used for completing residual addition, layer normalization, quantization / inverse quantization and nonlinear activation calculation of each layer; the three cores in the same node are located in the same SLR area, and data transmission is carried out through a hardware queue and an HLS blocking mechanism. According to the invention, high hardware utilization rate and low-delay reasoning can be realized.

The invention discloses a shift change information intelligent complementation and error early warning system based on a large language model, which belongs to the technical field of medical information processing and comprises a terminal acquisition module, an intelligent distribution complementation module and an early warning control module. Structure and semantic fingerprints are generated for the data table, a shift exchange asset meta model is constructed through fusion similarity clustering merging, and a semantic enhancement model is generated in combination with medical shift exchange specifications to complete information field completion; the intelligent distribution and completion module configures an optimal path with the minimum comprehensive overhead for the acquisition module by relying on a regional dynamic shift information network formed by double sub-networks; the early warning control module is combined with a large language model semantic analysis and rule engine, monitors an acquisition track and identifies information abnormity to execute error early warning, and an early warning result is fed back to the intelligent distribution complementation module, so that dynamic adjustment of an optimal path is realized, and the integrity, accuracy and acquisition efficiency of medical shift information are improved.

The application discloses a kind of robot dexterous operation systems, methods, equipment and media for transparent object, belong to robot dexterous operation and three-dimensional perception field, system includes: point cloud data construction module, can be constructed and noisy to obtain noisy hand-object interaction point cloud by physical simulation;Point cloud feature extraction module can extract high-dimensional latent features from noisy hand-object interaction point cloud;Query prediction module can infer object shape and pose by query point generation and decoder decoding;Perception coding module can extract features from global point cloud, hand-object interaction point cloud and tactile information to obtain multi-source perception features;Multi-modal feature fusion module can realize feature integration through self-attention mechanism and cross-attention mechanism calculation, and output fusion features;Action generation strategy module can generate differentiated action instructions for robot arm and dexterous hand based on fusion features.The system can achieve higher adaptability and stronger generalization capability in transparent object operation task.

The application relates to the technical field of laser precision machining, in particular to an optical coherence tomography real-time feedback control system based on neuromorphic computing, which comprises an optical coherence tomography device and a processing module; the optical coherence tomography device comprises a collecting module and an executing module; the collecting module is used for collecting an analog voltage signal corresponding to an original optical interference signal of a biological tissue; the processing module is used for converting the analog voltage signal into a digital signal stream, inputting the digital signal stream into a neural network model, outputting a target depth position index from the neural network model, and generating a control signal according to the target depth position index; and the executing module is used for controlling the energy output of a surgical laser guided by the optical coherence tomography according to the control signal, simulating a reflex arc mechanism of a biological nervous system, constructing a bionic optical reflex arc architecture, and being low in calculation delay and small in computing power consumption. Therefore, the problems that related technologies cannot match microsecond laser pulses and cannot realize precise control are solved.

The present application relates to a kind of marine diffusible matter concentration field prediction method, system and equipment, belong to marine environment monitoring technical field.It includes: respectively deploying acoustic sensor node cluster to each sub-region of diffusible matter distribution, and constructs multiple NG-RC module;The acoustic sensor node in each acoustic sensor node cluster is divided into prediction and observation sensor node, and corresponding original observation sequence is obtained;Input vector is obtained according to original observation sequence, the vector is input into NG-RC module, and the local concentration prediction result of each sub-region is obtained;Cross-regional association information fusion is carried out to each NG-RC module, and according to local concentration prediction result, the global prediction result of each sub-region is obtained;According to the spatial position of each sub-region, the global prediction result of each sub-region is spliced, and the prediction concentration value of corresponding time whole region is obtained.The present application improves timing prediction ability and improves prediction precision, and enhances the adaptability to complex environment.

The application discloses a kind of adaptive optics real-time controller hardware architecture based on multi-core CPU and GPU.Shack-Hartmann wavefront sensor sub-aperture slope calculation, wavefront recovery and wavefront control are completed in parallel by multi-core CPU and GPU using the real-time controller hardware architecture.The application uses general multi-core CPU and GPU as computing platform, after receiving image data in multi-core CPU, part of sub-aperture image data is transmitted to GPU by PCIE bus using a CPU core, and wavefront calculation task of the part of sub-aperture is completed by GPU;At the same time, the remaining CPU core is used to complete the remaining sub-aperture wavefront calculation task in parallel, to reduce the purpose of real-time controller calculation delay.The application uses multi-core CPU and GPU in parallel to complete wavefront calculation task, to reduce the effect of adaptive optics real-time controller calculation delay, to meet the real-time requirement of adaptive optics system to real-time controller, especially suitable for next-generation large-aperture ground-based telescope adaptive optics system real-time control.

Embodiments of the present application provide a request processing method and device, electronic equipment and computer readable storage medium, and relate to the technical field of computers. The method comprises: determining a response end for responding to the display request according to the number of target electronic labels requested in the display request and / or the estimated response time of the central cloud server responding to the display request; and responding to the display request based on the response end. In this way, the display requirement of the electronic labels in the actual scene and the data processing capability of the central cloud server can be flexibly determined to respond to the display request by the central cloud server and / or the edge cloud server, thereby relieving the processing pressure of the central cloud server, reducing the network bandwidth occupation, and reducing the calculation delay.

The invention relates to the technical field of hydrological monitoring and computer vision, in particular to a hydrological station flow monitoring system and method based on deep learning, and the method comprises the steps: synchronously collecting the water level time sequence data of a river section and a video key frame flow containing water surface texture; converting the water level time sequence data into a corresponding water passing area numerical value and a hydraulic radius numerical value; outputting a flow state feature vector representing the dispersion degree of the water surface flow velocity; outputting a physical parameter vector changing along with time; calculating a moving average value of the river course roughness coefficient in a preset time window; judging whether the moving average value is greater than a preset siltation alarm threshold value or not; if the judgment result is yes, generating a river deposition early warning data packet and sending the data packet to a remote monitoring terminal; otherwise, generating a state log of normal operation of the river channel and storing the state log into the local database; the limit that a traditional flowmeter can only output numerical values is broken through, and a key diagnosis basis is provided for operation and maintenance and flood control decision making of a hydrological station.

The invention discloses an auto-regression and mask diffusion double-normal-form fusion language model hardware acceleration system based on an FPGA (Field Programmable Gate Array), and relates to the field of FPGA and machine learning. The invention provides a model core calculation process-oriented hardware acceleration architecture by aiming at a language model fusing double normal forms of an autoregression model and a mask diffusion model and adopting the characteristics of revising an attention mechanism and parallel generation of a KV cache. The method disclosed by the invention is implemented by taking an Eso-LMs (Esolic Language Models) model as an example. The system comprises a layer normalization and adaptive layer normalization modulation module, a QKV projection module, a rotation position coding application module, a KV cache management module, a multi-head attention calculation module, an output projection and residual connection module and a multi-layer perceptron module. By optimizing the calculation sequence and the data flow of each module in the model calculation process and adopting the pipeline parallel and resource reuse technology, the efficient reasoning acceleration of the language model fusing the autoregression and mask diffusion double normal forms on the FPGA platform is realized, and the model reasoning speed is obviously improved. According to the system, the advantage of FPGA customizable hardware acceleration is fully exerted, the utilization efficiency of hardware resources is improved, data pipeline blockage is eliminated, calculation delay and storage overhead are reduced by reconstructing the data flow direction and constructing a whole-process pipeline processing architecture, and the system is suitable for the deployment requirement of an edge calculation scene.

The application discloses an intelligent agent real-time decision-making method and system based on Riemannian manifold and a medium, and relates to the technical field of artificial intelligence. The method is executed by an intelligent decision-making system, and comprises the following steps: constructing a behavior transition graph according to a historical trajectory data set; mapping an attention matrix generated by a large language model in a decision-making task to a symmetric positive definite matrix manifold for representation; embedding the behavior transition graph into a hyperbolic Riemannian space for representation; learning a geometric alignment mapping from the symmetric positive definite matrix manifold and the hyperbolic Riemannian space to a common metric space by optimizing an objective loss function; in the reasoning stage, obtaining current environment observation data, inputting the large language model, obtaining corresponding attention representation, and mapping the attention representation to the common metric space to obtain a query vector; and taking an action corresponding to an action node with a geometric distance lower than a preset threshold as a control instruction. In this way, the decision-making calculation overhead can be significantly reduced, and the cross-scene generalization capability and decision-making interpretability can be improved.

The application discloses an SNN hardware accelerator, belongs to the technical field of integrated circuits, and provides a hardware architecture for realizing an STDP+DFA algorithm, and realizes the calculation of each network layer of an SNN network based on a LIF neuron array; a plurality of activity counters corresponding to each neuron in the SNN network are arranged in the SNN hardware accelerator; the number of pulse firings of the corresponding neuron is recorded through the activity counter; when a network training request is received, the neuron with the number of pulse firings greater than a preset activity threshold is taken as an active neuron, the active neuron firing a pulse at a current time step t is taken as a neuron to be updated, and when the corresponding error signal is calculated, only the column vector under the column index corresponding to the neuron to be updated in the preset sparse matrix is multiplied by the corresponding error loss gradient vector, so that the SNN hardware accelerator with high precision can be realized with lower calculation resource consumption.

The invention relates to the technical field of electronic test and measurement, and particularly discloses a complex signal multi-objective optimization system and method based on a PXIe bus, the system comprises a system controller, a PXIe bus backboard and a functional hardware module, the functional hardware module comprises a signal acquisition module, a core processing module, a signal output module and an auxiliary module; the signal acquisition module is used for converting a complex analog signal into a complex digital signal; the core processing module is used for processing the complex digital signals according to the current configuration parameters and obtaining optimized configuration parameters through a multi-objective optimization algorithm; and the signal output module is used for converting the optimized configuration parameters into actual physical signals and then outputting the actual physical signals so as to drive subsequent execution equipment or serve as feedback signals. According to the invention, the high bandwidth, low delay and modularization characteristics of the PXIe bus are utilized, and a multi-objective optimization algorithm is combined to realize efficient acquisition, processing and optimization of complex signals, so that the real-time performance, precision and flexibility of signal processing are improved.

This invention discloses a method for parallelizing and pipelining a nested multiplier, relating to the field of intelligent warehousing technology. Based on the nested multiplication concept, the multiplication operation is decomposed into a formula to obtain the squaring operation of the high-bit part and the difference of squares plus shift operation of the low-bit part. Based on the characteristic of 2ab in the multiplication formula, the intermediate values ​​required for high-bit calculation are obtained by querying a preset low-bit square value table through the low-bit results. For batch data, multiple sets of data are processed simultaneously through a parallel architecture. The overall calculation process is divided into multiple pipeline stages, and continuous calculation is achieved through stage cascading. The pipeline output results of each parallel computing unit are collected, verified, and integrated to output a complete set of multiplication results. This invention achieves breakthroughs in hardware resources, computational latency, and batch throughput through the collaborative design of nested recursive simplification, pipeline reuse, and parallel architecture, providing an efficient solution for high real-time and high-parallelism digital computing scenarios.

This invention provides a method for accelerating the compression of intelligent models on edge terminals, relating to the field of edge intelligence. Its key feature is that it includes: acquiring real-time operating status data of the edge terminal and feature information of the current input data, wherein the real-time operating status data includes computing resource load, memory occupancy, battery level, and device temperature. The advantages of this invention are: by sensing the device status and input complexity in real time, and utilizing reinforcement learning to dynamically generate an optimal strategy combination including pruning, hybrid quantization, and early termination mechanisms, this method overcomes the limitations of traditional static compression, achieving a dynamic balance between computing power, accuracy, and energy consumption. It not only significantly reduces inference latency and memory usage but also adaptively adjusts based on battery level and temperature, effectively solving the problems of difficult deployment, high heat generation, and short battery life of large models on resource-constrained edge devices, and greatly improving the real-time performance and stability of intelligent applications.

The application is specifically an optimization method of an air unmanned aerial vehicle communication network based on AI, relates to the technical field of communication networks, and comprises intelligent sensing and data processing, an AI decision and optimization core, communication and task collaborative management, and resource and energy consumption optimization.In the application, a laser radar scans a canyon terrain in real time, a cutoff frequency is calculated in combination with a rectangular waveguide model, a transverse electric wave transmission mode is dynamically adjusted, terrain changes are predicted in advance and mode switching is triggered, a buffer interval and a rollback mechanism are matched, and signal interruption caused by a terrain-induced multipath trap can be avoided; when a canyon width suddenly changes, mode switching is started in advance, parameter updating is completed before the unmanned aerial vehicle reaches a critical area, the risk of communication interruption is reduced, the rollback mechanism can correct erroneous switching, and link continuity is ensured.

This invention belongs to the field of high-precision position detection technology, and particularly relates to an absolute grating position calculation method based on an adaptive error-correcting sliding window. A photodetector array, in conjunction with an error-correcting sliding window, collects photoelectric signals of symbols, converting them into a symbol amplitude sequence. Abnormal erased symbols are identified through symbol confidence levels. If erased symbols are found, the window length is gradually expanded, and the process is repeated iteratively until a qualified target error-correcting sliding window is obtained. After removing erased symbols within the target error-correcting sliding window, grating position calculation is performed. After detecting a single acquisition position, the process is repeated for subsequent acquisition positions until the task is completed. This invention relies on adaptive iteration of the window to achieve signal error correction, reducing the interference of bad symbols on the calculation results and effectively improving the overall performance of grating position detection.

This invention discloses an FPV unmanned aerial vehicle (UAV) identification method and apparatus based on analog image transmission signals. The method includes: S1, setting acquisition parameter information, including acquisition center frequency and acquisition bandwidth; S2, acquiring spatial electromagnetic wave signals based on the acquisition parameter information to obtain an acquisition signal; S3, performing identification processing on the acquisition signal to obtain an identification signal.

The application relates to the technical field of artificial intelligence security, and provides an agent runtime security control method and device, which comprises the following steps: maintaining a structured security state digest containing a historical interaction compressed representation and not containing complete plaintext during the running of an agent; in response to a current interaction request, performing security reflection based on the digest and a loaded dynamic defense strategy, and outputting a security disposal decision; when a preset security anomaly triggering condition is met, marking an accident candidate session; extracting a generalization risk feature based on the interaction track of the accident candidate session, performing desensitization processing, and generating a dynamic defense strategy for subsequent session loading. The application realizes the structured precipitation and lightweight runtime security reflection of defense knowledge, and supports the closed-loop self-evolution of security capability.

The invention relates to the technical field of computing resource allocation, and discloses a satellite-assisted multi-unmanned aerial vehicle computing unloading method and device and a program product, and the method comprises the steps: obtaining a state sample set generated by interaction between each unmanned aerial vehicle and a satellite mobile edge computing network and stored in an experience playback buffer area; judging whether the number of state samples in each state sample set is greater than or equal to a preset batch size; and when the number of the state samples in the state sample set is greater than or equal to a preset batch size, updating the initial calculation unloading strategy of each unmanned aerial vehicle by using a near-end strategy optimization algorithm until the satellite mobile edge computing network is in a stable state, and obtaining an optimal calculation unloading strategy of the unmanned aerial vehicle, according to the invention, low energy consumption and low calculation delay of the unmanned aerial vehicles are ensured in a high-dynamic network environment, and conflicts among the unmanned aerial vehicles when the unmanned aerial vehicles acquire calculation resources are reduced.

This invention relates to the field of integrated circuit design technology, and particularly to an access system, memory system, and access method for three-dimensional data access. The access system includes: a configuration register unit, an address generation unit, a data assembly and splitting unit, and a dual-pipeline controller. The address generation unit, connected to the configuration register unit and the dual-pipeline controller, calculates a one-dimensional physical address based on a target index value input by the user and sends it to the dual-pipeline controller. The dual-pipeline controller, connected to the memory interface, independently manages read and write flows based on user read / write requests, enabling read and write operations in the memory based on the one-dimensional physical address. The data assembly and splitting unit, connected to the dual-pipeline controller, assembles and splits data during the read / write flow before sending it to the dual-pipeline controller. This solution uses hardware to implement three-dimensional data indexing, which can improve data processing efficiency and system flexibility.

A processor and electronic equipment are applied to the field of data processing. The processor includes a scaling factor processing module including a smoothing pre-processing unit configured to receive the first tensor, the second tensor, and a smoothing control parameter, and a processing unit configured to process the first tensor and the second tensor by a plurality of arithmetic logic units and special function units in the processor. Executing smoothing operation of the first tensor and the second tensor to obtain a third tensor and a fourth tensor; the processing unit is configured to determine scaling factors of the third tensor and the fourth tensor and execute matrix multiplication and accumulation operation using the scaling factors based on the scaling factors of the third tensor and the fourth tensor to obtain a matrix multiplication operation result. The processor provides efficient smoothing preprocessing hardware, the processing does not need to be executed in special hardware devices such as a vector core, hardware-level efficient execution of smoothing preprocessing is achieved, hardware resource scheduling overhead and calculation delay are effectively reduced, and hardware execution efficiency is improved.

The invention provides a convolution structure optimization method for Q, K and V feature generation in a Transform operator, and belongs to the technical field of deep learning. Comprising the following steps: directly and independently generating a query feature, a key feature and a value feature by using three parallel feature generation paths; wherein each path sequentially comprises a first type of convolutional layer and a second type of convolutional layer; wherein the input features of the three paths are the same, and the last-stage outputs of the three paths are respectively used as the final query feature, the key feature and the value feature. Through the convolution splitting strategy, the grouping convolution reconstruction strategy, the elimination mechanism of the channel splitting operation and the storage optimization scheme of the intermediate tensor, the memory consumption can be remarkably reduced, one-time large-scale memory access and movement are reduced, higher execution efficiency is brought to the model, the reasoning acceleration effect is improved, and the method has the advantages of being simple in structure and convenient to operate. The method is especially suitable for equipment sensitive to bandwidth.

This application discloses a gas concentration field reconstruction and short-term prediction method and system based on sparse sensing, relating to the fields of gas leak monitoring and intelligent safety early warning technology. It can solve the problems of physical accuracy and near real-time computational response speed in global gas concentration field reconstruction. The scheme includes: acquiring concentration data from multiple gas sensors deployed in a target space; mapping each concentration data to a target planar grid according to the spatial location of each sensor, generating a two-dimensional concentration mapping image, and generating a mask image; inputting the two-dimensional concentration mapping image and the mask image into a pre-trained LeakU-Net reconstruction neural network model to obtain a two-dimensional continuous concentration field reconstruction map at the current moment; inputting a sequence composed of the two-dimensional continuous concentration field reconstruction map and two-dimensional continuous concentration field reconstruction maps from previous moments into a pre-trained LeakU-Net prediction neural network model to obtain a concentration field prediction map for the next moment; and monitoring gas safety in the target space based on the two-dimensional continuous concentration field reconstruction map and the concentration field prediction map.

This invention discloses an ultra-high-speed pipelined five-point median filtering method based on fully parallel hardware logic, comprising the following steps: S1, buffering the input one-dimensional signed number sequence x(n) into five points and outputting the data in two's complement form in parallel according to a specific timing sequence; S2, preprocessing the parallel data output in S1; S3, performing parallel pairwise comparisons on the preprocessed data in S2 to obtain the size relationship between the data, and storing the comparison results in the corresponding size relationship registers; S4, calculating the redundancy output median index number based on the preprocessed parallel data in S2 and the value of the size relationship register in S3; S5, using the median index number calculated in S4, selecting the index value of the corresponding index number in the original parallel data for output using a MUX. This method outputs the median filtering result with fewer comparisons, less clock delay, less hardware logic resources, and lower computational latency.

The invention provides a large language model dynamic cue word optimization method based on uncertainty perception and application thereof, and belongs to the technical field of natural language processing. The method comprises the following steps: constructing an initial cue word by using a reference example screened based on uncertainty indexes to perform preliminary reasoning; calculating a logarithmic focus uncertainty index based on the probability distribution entropy of the reasoning first word and the category prior probability; if the index is lower than a preset threshold value, a result is directly output; if the reference samples are higher than the threshold value, a retrieval enhancement mechanism is triggered, and enhanced prompt words (Prompt) containing reflection guidance and similar reference samples are constructed for secondary reasoning. According to the method, reasoning paths are dynamically allocated through real-time sensing of model confidence, excessive confidence of the model is effectively inhibited, calculation overhead is reduced, and task robustness in a complex scene is improved.