237 results about "Hyperscale" patented technology

An end-side-cloud three-in-one active chat robot system for high-emotional merchants comprises an end side, a local edge server and a cloud end, and the end side is used for collecting multi-modal data of a user and performing local lightweight real-time processing and response execution; the local edge server is used for receiving and fusing the multi-modal features and the context information from the end side, and carrying out sentiment calculation, dialogue management and active trigger decision making with medium complexity; the cloud end is used for operating a super-large-scale model and providing global knowledge management, long-term user portrait storage and model training optimization; and the end side, the local edge server and the cloud end carry out cooperative communication through an encrypted channel to form a distributed intelligent processing architecture. According to the method, global optimization is realized by integrating end-side lightweight sensing, edge multi-modal fusion and cloud long-term memory. A composite finite state machine (FSM) active questioning mechanism is combined with sentiment calculation, and is different from traditional rule type triggering. Off-line and on-line fusion scheduling and multi-agent role playing are combined to be applied to a high-emotional-quotient interaction scene, and the simulation is improved. Multi-modal emotion perception circulation is introduced, and the problem of single text emotion misjudgment is solved.

Technologies for providing energy efficiency technology in extreme mMIMO systems in a cellular network cellular network (e.g., 5G wireless network, 6G wireless network) are described. The method collects data representing conditions for potential energy saving (ES) modes, comprising morphology data and traffic pattern data. The method determines, using the collected data, one or more energy saving (ES) modes for one or more components of a cellular network in at least one of a time (T) domain, a frequency (F) domain, or a space (S) domain, wherein the one or more ES modes cause at least one adjustment to the one or more components in the at least one of the time (T) domain, the frequency (F) domain, or the space (S) domain.

The invention discloses a robot hierarchical path planning method and device in a complex environment, a medium and a product, and the method comprises the steps: responding to a path planning request of a robot in an airport terminal, dividing the airport terminal into a plurality of dynamic Voronoi regions according to an underlying environment grid map and people flow density information, a minimum path cost algorithm with people flow density weighting is adopted to construct a global path anchor point sequence, the global path anchor point sequence is further split into a plurality of local path planning requests, a double-heuristic cost function is constructed, an improved algorithm is adopted to generate local sub-paths, and according to surrounding environment information collected by a robot in real time, the local sub-paths are calculated. And optimizing each local sub-path by adopting a reinforcement learning algorithm, generating motion control instructions for the robot at different control time points, and realizing dynamic obstacle avoidance of the robot. The path planning efficiency and the motion safety of the mobile robot in a super-large-scale dynamic environment are remarkably improved, and the technical problems that a traditional method is high in calculation complexity and poor in real-time performance are effectively solved.

A simulation method, apparatus and terminal for long-time-scale hierarchical operation of a power grid, and a medium. In the technical solution provided in the present application, on the basis of a hierarchical subordination relationship, simulation of a large-scale power grid is divided into three levels: a national level, a provincial level, and a prefectural level; by means of a hierarchical decoupling method, operation simulation of an ultra-large-scale and complex power grid is simplified into three stages: national-level power grid operation simulation, provincial-level power grid operation simulation, and prefectural-level power grid operation simulation; a lower-level simulation model only needs to perform operation simulation on the basis of key parameters output by an upper-level simulation model, and other than a limited number of key parameter interactions, the operation processes of simulation models at all levels are independent of each other; thus, the computational complexity of operation simulation of the power grid is simplified, and the association between upper-level and lower-level power grids is also taken into consideration, so that when facing massive scenario computation requirements, the operation time can be effectively reduced, and the operation efficiency can be improved.

The invention relates to the technical field of large language models, in particular to a high-concurrency large language model high-speed reasoning deployment method, which comprises the following steps of: extracting a weight coefficient corresponding to a user identity identifier, an emergency degree numerical value mapped by a request type and a resource occupation value converted by an estimated calculation amount according to the user identity, the request type and the estimated calculation amount; according to the method, the dynamic priority is generated by performing comprehensive operation on the user identity, the request type and the estimated calculation amount, so that differentiated services for reasoning tasks are realized, interactive requests with high urgency degree can bypass batch processing tasks with long time consumption, the response delay and delay jitter of key services are reduced, and the user experience is improved. Meanwhile, according to the complexity of the request text and the geographic coordinates of the user, the task is intelligently routed to the model node with the most suitable scale in the distributed network, so that the wide area network transmission delay is greatly reduced through edge processing, and the computing power waste caused by using a super-large-scale model to process the simple task is also avoided.

The invention relates to a power network structure weakness detection method based on multi-diagonal-block matrix decomposition, and belongs to the technical field of super-large-scale integrated circuits. According to the method, an original large-scale sparse matrix is converted into a band edge diagonal block structure and divided into a plurality of sub-matrixes capable of being solved independently by constructing a layering and blocking strategy for eliminating tree drive, and redundancy of full-matrix calculation is avoided. Meanwhile, a local approximate inverse algorithm of column norm truncation is designed, and target elements are calculated on the premise that a preset error threshold value is met. According to the method, the parallel computing architecture of the multi-core processor is fully utilized, the computing complexity is reduced by 2-3 orders of magnitude, the computing efficiency is effectively improved, the memory occupation is remarkably reduced, and a high-precision and high-efficiency solution is provided for detecting the weakness of the power network structure of the super-large-scale integrated circuit.

The invention provides an MCM-GPU-oriented resource storage optimization method, device and equipment, and the method comprises the steps: carrying out the static analysis of an operator-level data flow diagram of a calculation task in a compiling stage, and extracting the data dependence relation and life cycle characteristics of each operator; generating a storage optimization scheme based on the data dependency relationship and life cycle characteristics of each operator; in the operation stage, task and data collaborative allocation is carried out according to a data prefetching strategy and a data replacement strategy, task and data collaborative allocation and the calculation task are executed in parallel, and a data flow diagram and an operator dependency relationship of the task are analyzed by utilizing a compiling period; in combination with the data locality characteristics of the MCM-GPU multi-level storage architecture and the concurrency requirements of different modules for data access, an appropriate GPU module is selected to carry out data optimization management, so that frequent data exchange between a GPU and a CPU is reduced, and the performance potential of the MCM-GPU architecture in a super-large-scale task is brought into full play.

The invention relates to a multi-mode multi-agent RAG-based power dispatching method and system, and the method comprises the steps: generating a dispatching plan through the cooperative work of a plurality of heterogeneous agents with different division of labor, and receiving query input containing a power dispatching task; the planner agent decomposes the power dispatching task into a plurality of sub-tasks which are mutually dependent and generates a corresponding execution plan; according to the execution plan, the searcher agent sequentially executes two stages of fine-grained multi-modal retrieval on each sub-task from the multi-modal power knowledge base, coarse-grained whole-page retrieval is carried out in the first stage, fine-grained area fine reading and modal expansion are carried out in the second stage, and evidence related to the current sub-task is obtained and added into an evidence pool; and according to all evidences in the evidence pool, the responder agent answers the initially input query to generate a scheduling plan. According to the method, the problems of timeliness and multi-scene adaptability in super-large-scale power dispatching are effectively solved, and an innovative path is provided for guaranteeing dynamic safety and stability of a power grid.

The invention relates to the technical field of full-link monitoring, and discloses a Skwalking-based micro-service system full-link monitoring method, which comprises the following steps of: 1, automatically collecting full-link tracking data without invasion, loading an Agent probe through a bytecode enhancement technology when a micro-service instance is started, and loading the Agent probe through a bytecode enhancement technology; and the Agent probe automatically intercepts service calling under the condition of not modifying a service code. According to the technical scheme, a Kafka cluster is introduced between a data acquisition end and a processing platform to serve as a buffer channel, and an L1 and L2 distributed OAP processing architecture capable of being horizontally expanded is designed, so that the technical effect of guaranteeing the integrity and real-time performance of mass monitoring data in a super-large-scale micro-service environment is achieved; compared with a transmission architecture in which a monitoring probe is directly connected with a central server in the prior art, the transmission architecture solves the problem that the accuracy of a monitoring result is damaged due to the fact that data congestion, delay and even packet loss are easy to occur during data peak flood due to the lack of an effective buffering and peak clipping mechanism.

The invention relates to a low earth orbit satellite network measurement method, device and equipment, a medium and a product. The method comprises the steps that a topological graph of a low earth orbit satellite network is acquired, an Euler loop is determined according to the topological graph, and the topological graph comprises a plurality of satellite vertexes and edges between the adjacent satellite vertexes; determining a detection path set according to the Euler loop, and determining a report path set according to the detection path set; and collecting network state information of each satellite vertex based on the detection path set, and transmitting the network state information of each satellite vertex to a ground station for network measurement based on the report path set. By adopting the method, the whole-network state measurement of the super-large-scale low-orbit satellite network can be realized.

The invention discloses an enhanced high-emotional commerce interaction system based on end-side cloud collaboration. The enhanced high-emotional commerce interaction system comprises end-side equipment, an edge server and a cloud platform, the end side device is provided with an enhanced processing capability and comprises a lightweight finite-state machine decision module, a local model inference engine and an intelligent data preprocessing module. And the edge server is responsible for multi-modal information fusion, complex scene decision and distributed coordination management. The cloud platform is responsible for super-large scale model training, global knowledge management and system optimization. The system adopts a dynamic capability distribution strategy, and autonomously selects a processing path according to scene complexity. The end-side device can independently process a simple scene, and only requests the edge server for cooperative processing in a complex scene. By means of the design, system delay is remarkably reduced, privacy protection is enhanced, and system reliability is improved. According to the scheme, system delay is remarkably reduced, and millisecond response can be realized in a simple scene; privacy protection is enhanced, sensitive data are processed on the end side, and data transmission is reduced; the system reliability is improved, and basic services can still be provided in a weak network environment; personalized interaction is realized, and the end side continuously learns habits and preferences of the user; resource utilization is optimized, and edge and cloud computing loads are reduced.

The invention relates to the technical field of wireless communication, and discloses a sparse channel estimation method and system for 5G-A5G-A super-large scale MIMO. The system comprises a quantitative sensing preprocessing module, a layered sparse reconstruction module, a high-precision guide correction module and a dynamic threshold decision module, and realizes efficient and accurate recovery of channel sparse characteristics under strong quantitative interference through nonlinear distortion compensation, multi-resolution sparse recovery, high-precision reference signal guide correction and parameter adaptive adjustment. The estimation performance is improved; and the calculation complexity is balanced.

The invention discloses a real-time quality detection method for Internet of Things big data of a new energy power grid. The method comprises the following steps: firstly, collecting equipment time sequence data in a new energy power grid by a lower-layer Internet of Things platform, and uniformly pushing the equipment time sequence data to a Kafka platform; a user needs to deliver quality inspection standard data to the Flink to be loaded in a broadcast stream mode; meanwhile, the system can realize automatic synchronization after change of measuring point grouping, detection rules and parameter thresholds by means of the characteristics of Flink broadcast streams according to metadata reloaded in a period agreed by a user; the method comprises the following steps: firstly, using an asynchronous IO to consume kafka, then indicating Flink to use asynchronous IO to consume kafka, using configuration in a quality inspection standard data broadcast stream, then carrying out data detection, and finally dropping the detected data into a distributed time sequence database cluster by virtue of Flink Sink. The problems that an existing offline scheme is poor in timeliness, insufficient in adaptability and difficult to cope with super-large-scale concurrency are solved.

The invention relates to the technical field of computers, and discloses a multi-FPGA collaborative deep neural network parallel acceleration method. According to the method, unified compiling and automatic calculation graph segmentation mapping based on a dynamic programming algorithm are carried out on a deep neural network model through a host control node, a heterogeneous strategy fusing data, the model and assembly line parallel is generated, and the problems of single FPGA computing power resource limitation and memory wall are effectively solved; direct communication between the FPGAs is achieved through a high-speed point-to-point internet, an asynchronous assembly line execution mechanism and a hardware-level data packaging technology are combined, communication delay is reduced, assembly line bubbles are eliminated, and a multi-FPGA cluster reaches a near-linear speed-up ratio; computing distribution is optimized in real time through a global load balancing unit, and the equipment utilization rate is increased; and meanwhile, a transparent deployment interface is provided through the hardware abstraction layer, so that the super-large-scale neural network can efficiently run on a multi-FPGA platform with low delay, and the overall calculation efficiency is improved.

The invention discloses an industrial chain electricity consumption sensitivity feature extraction method, system and device and a storage medium, and the method comprises the steps: obtaining multi-dimensional electricity consumption data, carrying out the feature extraction and dimension reduction of the multi-dimensional electricity consumption data, and mapping an original high-dimensional feature into a low-dimensional feature; carrying out regression analysis on the low-dimensional features, calculating a feature sensitivity coefficient, reversely associating original high-dimensional features, calculating the total influence degree of the original high-dimensional features on power consumption, and screening to obtain key features; based on the characteristic sensitivity coefficient and the key characteristic, the deviation degree of the low-dimensional characteristic and the normal working condition distribution is calculated, abnormity is detected according to the deviation degree, when abnormity is detected, a dynamic optimization mechanism is triggered, extraction of the electricity utilization sensitivity characteristic is achieved, and online learning and dynamic optimization requirements in a super-large scale time sequence scene are supported.

This invention discloses a phased anomaly detection method for ultra-large-scale industrial software supply chains based on graph neural networks. It divides the software anomaly detection task into two stages. In the rapid contamination stage, a simple model is used to quickly learn global anomaly patterns and filter out potential anomaly nodes and regions. In the precise detection stage, a more complex model is used to detect anomaly nodes. Furthermore, this invention defines the "disorder level" of nodes on the graph based on the Lyapunov coefficients in chaos theory and utilizes the "defect" of rapid oversmoothing in graph neural networks to rapidly propagate the behavioral patterns of anomaly nodes on the graph, simulating the propagation process of the entire chaotic system. In this way, this invention can quickly learn the global node anomaly patterns and quickly and accurately filter out candidate anomaly nodes and regions.

The application provides an automatic disk IO method based on structural mechanics CAE simulation software, comprising the following steps: S1, constructing a memory-disk cooperative computing framework, and dynamically monitoring a data access state in a simulation solving process; S2, automatically exchanging non-active core data structures to a high-performance disk according to a data access frequency, a life cycle and a memory pressure; S3, loading back to the memory on demand when subsequent calculation needs to access the data exchanged to the high-performance disk; wherein the core data structures comprise a stiffness matrix, shape function data, an intermediate result vector and derivative physical quantities, and the exchange process is transparent to the solver. According to the application, the exchange process is highly transparent to the core computing logic of the solver, is automatically managed by the system, and does not need user intervention, so that efficient utilization of memory resources and stable operation of super-large-scale model simulation are realized.

Described herein are techniques for analyzing historic time-series data to determine right-size commissioned resources from a software provider such as a hyperscaler. The historic time-series data may be analyzed to determine whether a commissioned resource should be upsized or downsized based on historical usage of the resource. Advantages to right-sizing a commissioned resource include improved performance and reduced spending.

This invention provides a heterogeneous architecture-based system and method for accelerating large language model inference, aiming to address the limitations of existing acceleration solutions in terms of cost, power consumption, storage capacity, and collaborative efficiency in meeting the deployment requirements of ultra-large-scale MoE models. The system primarily comprises a CPU, an FPGA accelerator card, and a high-speed communication interface. This method, targeting large language models with a hybrid expert architecture, analyzes the computational characteristics and storage requirements of different modules. It deploys computationally intensive attention calculations and shared expert calculations with relatively fixed parameter sizes on the FPGA accelerator card; while storing routing expert weights with large parameter sizes, sparse access, and strong data dependencies in the CPU memory, which then performs routing decisions and routing expert calculations. This invention effectively overcomes the storage capacity limitations of programmable logic accelerator cards, supports the deployment of ultra-large-scale sparse parameter models, and significantly reduces system cost and power consumption, improving the overall energy efficiency and practicality of heterogeneous collaborative inference.

The present disclosure relates to a method, system, storage medium and electronic device for very large scale integrated circuit; the method for very large scale integrated circuit obtains timing information TNS of the entire netlist of the very large scale integrated circuit; all nodes are sorted according to the topology structure of the integrated circuit, and nodes with the same serial number are placed in the same queue; all sorted nodes are traversed in reverse according to the number of queues; the size reduction preprocessing is performed on all nodes in the same queue, and the slack change value of the output pin of the current node and the next level node is obtained; the slack change value of the output pin of the current node and the next level node is checked, the current node with the slack change value of the output pin meeting the requirements is subjected to size reduction processing; the timing information of the current node and the next level node is updated, the timing information new_TNS of the entire netlist is reobtained; the timing information new_TNS and TNS are compared, and the area power consumption in the physical optimization of the very large scale integrated circuit is optimized according to the comparison result. The problem of excessive area power consumption in the physical optimization design process under the condition of unchanged timing can be solved.

The application discloses a kind of high-efficiency parallel simulation method and system of super large scale low-orbit satellite network, to carry out high-efficiency parallel simulation while being able to simulate low-orbit satellite network dynamic characteristic.The application can be in the topological partition of simulation process, topological construction and parallel simulation of multiple CPU computing cluster automation;Meanwhile, a kind of satellite network topology division mechanism proposed in the application can quickly divide network topology into multiple simulation processes, reduce inter-process synchronization overhead, improve parallel simulation efficiency;And a kind of conservative synchronization mechanism for satellite dynamic network is proposed, which can ensure the topological consistency between multiple simulation processes, ensure causality, and promote the correct operation of parallel simulation.

The application provides a high-dimensional multi-objective evolutionary method based on an improved dominance criterion, which is used for optimizing the wiring design of a very large scale integrated circuit physical design and comprises the following steps: step one, convergence of a non-dominated solution set is ensured according to a defined convergence index, and an adaptive parameter based on a genetic algorithm niche is combined to control the diversity of the solution set, the radius target and the line length target of the wiring design are optimized by minimizing MOP, and the dominance criterion is improved; step two, a convergence index and a diversity index are designed, the two indexes jointly constitute a dynamic fitness function, and individuals with better convergence and diversity are adaptively reserved to perform MaOEA-IDR environment selection; and step three, an adaptive t distribution crossover operator is provided, which can balance the global search capability of a Cauchy operator and the local exploration capability of a Gaussian operator in a high-dimensional space, and the ASDX adaptive distribution crossover operator is adaptively adjusted. t The application can effectively solve the wiring problem of the very large scale integrated circuit physical design.

The invention discloses an intelligent check point scheduling method and system for super-large-scale large model training, and belongs to the technical field of large model training fault tolerance. Aiming at the pain points of large data loss, low single-index prediction precision and no adaptive optimization caused by passive response of a traditional check point scheme, the method comprises the following core schemes: firstly, collecting hardware, software and historical fault features, and constructing a multi-dimensional feature matrix through standardization processing; an LSTM + attention mechanism lightweight model is adopted to output fault probabilities in the future 1-10 minutes, and risk levels are divided; finally, differential scheduling is executed according to grades, and meanwhile closed-loop optimization is formed through false report and adjustment threshold values, missing report and compensation features and incremental training; the system comprises a multi-source acquisition module, a model prediction module, a scheduling execution module and a feedback optimization module, and a collaborative support scheme is landed. According to the method, the fault can be early warned 1-10 minutes in advance, the method is adaptive to a 1000 + GPU super-large-scale cluster, and the long-period training requirement of a trillion parameter model is met.

A distributed inference communication method of a super-large-scale graph neural network comprises the following steps: dividing a global topology to be processed into a plurality of local sub-domains matched with the number of GPUs, and determining a local master node and a boundary master node allocated to each GPU; calculating coloring priorities of nodes, and performing parallel graph coloring based on the coloring priorities to construct a global independent set sequence; mapping the global independent set sequence to the local sub-domains of the GPUs to generate a local task slice sequence, which comprises boundary node slices and internal node slices, evaluating communication load and calculation load of the slices, and taking the internal node slices as calculation fillers to be executed in the communication waiting period of the boundary node slices; and each GPU executes an inference task by using a multi-stream asynchronous pipeline constructed locally according to the local task slice sequence. The method can improve the throughput of distributed inference.

Provided in the present application is a data routing system. The system comprises a first topological network, a second topological network and a third topological network, wherein the network structure of the first topological network is a toroidal network topology structure consisting of a first node group, the first node group comprises a first number of network nodes, and the network nodes in the toroidal network topology structure have connection relationships with upstream nodes and downstream nodes in each dimension; the second topological network is a fat-tree network topology structure consisting of the network nodes in the first topological network and a switch; there is a cascade relationship between the network nodes in the first topological network and the switch; the first topological network is nested in the third topological network, and the network structure of the third topological network is a fully connected network topology structure consisting of a second node group; and the second node group comprises a second number of first node groups, and there is a connection relationship between the first node groups. Thus, large-scale or ultra-large-scale networking can be realized at a relatively low cost.

The application provides a metasurface array, which is composed of several metasurface subarrays of different shapes; the metasurface subarrays cannot coincide with the original metasurface subarrays after being rotated by 90°, or 180°, or 270°, or being mirror flipped, the application adopts irregular subarray division for a super large scale reflector array, compared with a traditional structure without subarrays, the number of bias lines is greatly reduced, that is, the number of control interfaces of the smart reflector is reduced, under the condition that the performance indicators such as scattering pattern gain and sidelobe are slightly reduced; compared with regular subarray division, the grating lobe level can be obviously suppressed due to breaking the periodicity of the phase center of the subarray, and the application can effectively reduce the system cost, power consumption and engineering implementation difficulty.

The invention discloses a large language model softmax function hardware acceleration circuit and method, and belongs to the field of neural network hardware acceleration of super-large scale integrated circuits. An input sequence is divided into a plurality of data blocks which are processed in parallel, and three-stage pipeline division is adopted, so that average single calculation delay is shortened to G clock cycles, the calculation parallelism is improved, the calculation speed of a softmax function is improved, and the reasoning delay is reduced; and a sparse mask strategy of sparse threshold comparison is introduced, the sparsity of data is fully utilized, the problems of high calculation complexity and high calculation delay of the softmax function and the data access bottleneck of the softmax function are solved, the calculation cost is remarkably reduced, and the calculation efficiency is improved. In addition, a softmax function hardware circuit adaptive to software optimization is constructed, calculation delay and memory access pressure are reduced through the hardware circuit, and data processing efficiency is improved.

The invention provides a system architecture for cross-data center large model training and a resource allocation method and system, and belongs to the technical field of cross-wide area distributed large model training. According to the method, large-scale model cooperative training across multiple data centers can be realized, and the bottleneck that the computing power of a single data center is limited is broken through. Through unified modeling and scheduling of calculation, memory and network resources, task loads can be intelligently allocated according to hardware performance and network bandwidth of different data centers, and efficient collaborative utilization of computing power resources is realized. The training task of the super-large-scale model can be rapidly completed in the heterogeneous computing power environment, and the training time is remarkably shortened. The provided flexible parallelism degree allocation method can be adaptive to different task and resource conditions, the proportion of data parallelism, model parallelism and pipeline parallelism is automatically adjusted, the parallelism efficiency is improved, and the communication overhead is reduced. A training time estimation function is integrated, the overall time delay and resource requirements can be predicted before task execution, and a basis is provided for scheduling decision making.

This invention discloses a method, electronic device, and storage medium for ultra-large-scale MIMO synchronization, relating to the fields of wireless communication and channel measurement technology. The method includes: first, estimating the delay and direction parameters of at least one strong path as an anchor path based on reference subarray channel data; then, for each pair of adjacent subarrays, extracting the phase of the same path on the boundary elements based on the anchor path parameters, calculating the observed phase difference, and thus obtaining the observed value of the common phase drift between subarrays; next, fusing multiple observations to obtain a robust estimate of the common phase difference between adjacent subarrays; finally, accumulating the estimated values ​​of each segment with the first subarray as a reference to obtain the common phase drift of each subarray, performing phase compensation, and then stitching the data of each subarray according to coordinates into an equivalent ultra-large-scale array to achieve ultra-large-scale MIMO synchronization. Therefore, this invention can estimate and compensate for the common phase drift of subarrays without additional hardware time synchronization.

This application provides a method, apparatus, medium, and product for ultra-large-scale MIMO channel modeling, relating to the field of wireless communication technology. The method includes: obtaining the cluster power of each cluster according to a preset channel model; determining the spatially non-stationary SnS state of each cluster according to the cluster power or a preset statistical distribution model; determining a first power attenuation factor for different clusters at the base station based on the SnS state of each cluster; determining a second power attenuation factor for each element of the user equipment; and generating channel coefficients based on the first and second power attenuation factors. The solution of this application provides a simple, physically meaningful, and efficient channel modeling method, effectively overcoming the shortcomings of existing technologies and providing reliable theoretical support for the design and optimization of ultra-large-scale MIMO systems.