31 results about "Overhead (computing)" patented technology

A scale-extensible convolutional neural network acceleration system, and the system comprises a processor and at least one convolutional acceleration core, wherein each convolutional acceleration coremainly comprises a computing array, a controller and an on-chip cache; when the scale is extended, only the number of the convolutional acceleration cores needs to be increased, programs running on the processor need to be modified, and other hardware modules do not need to be changed. Namely, a plurality of convolution acceleration kernels can be added to improve the scale and computing performance of the system. The invention further provides a method based on the scale-extensible convolutional neural network acceleration system, extra overhead caused by scale extension can be reduced to agreat extent, and therefore the scale-extensible convolutional neural network acceleration system can be deployed on different hardware platforms. Meanwhile, the software and hardware cooperation modeis good in universality, and different convolutional neural networks can be supported. Compared with other circuits, the method has universality and expandability.

The invention discloses an online computing task unloading scheduling method for an edge computing environment, which is based on a dynamic Cournot game model and comprises the following steps that: aserver periodically publishes the rental price of computing resources; each application of each user equipment generates a calculation task according to a certain frequency; when a calculation task is generated, the user equipment calculates the number of rented calculation resources by using the obtained price information and historical information and a designed iterative algorithm, determinespayment function values obtained by calculation on different servers, selects an optimal scheduling mode in real time, and sends a calculation task request; the server receives a user equipment request, allocates computing resources in real time, and updates the rental price; and after calculation is completed, the user equipment receives result data and updates historical information, so that dynamic circulation is carried out, Nash equilibrium is achieved, and global optimization is realized. The method is low in calculation amount, small in communication overhead, high in real-time performance and suitable for differentiated user experience quality requirements of a dynamic network environment and different applications.

The invention provides a fine-grained low-overhead fault-tolerant system for a GPGPU. The fine-grained low-overhead fault-tolerant system comprises a task division module, a check point backup module,a redundancy execution and error detection module and an error repair module. The fault-tolerant processing of the instantaneous fault of the GPU computing component can be realized, and the problemsof large fault-tolerant granularity, high error repair cost, poor fault-tolerant system performance and the like in the traditional software fault-tolerant method of the GPU can be solved. The beneficial effects of the invention are as follows: thread tasks can be divided; the calculation scale of the kernel is reduced, only the relative active variables need to be backed up during check point backup, the space-time expenditure caused by storage is reduced, only part of objects related to errors need to be recalculated during error repair, the fault-tolerant cost caused by recalculation is reduced, and the asynchronous mechanism of the CPU-GPU heterogeneous system is fully utilized to hide the time delay caused by data transmission and improve the performance of the system.

The embodiment of the invention discloses a main CPU and NPU data interaction method and computing equipment, which can be applied to the field of artificial intelligence, the method is applied to thecomputing equipment, the computing equipment comprises a main CPU and an NPU, a target APP runs on the main CPU, the main CPU comprises an AI Service and an NPU driver, and the NPU comprises an NPU controller, an arithmetic logic unit and N registers. And the NPU allocates a register to the AI model, the main CPU maps a physical address into a virtual address of a virtual memory on the target APPafter receiving the physical address of the register sent by the NPU, and the main CPU can actually read/write the corresponding register on the NPU directly through the target APP, which is equivalent to that a direct connection path exists between the target APP and the NPU. When the main CPU issues the execution command of the AI model to the NPU through the target APP to the whole process ofobtaining the execution result of the AI model, the AI Service and the NPU drive are bypassed by the calculation pass, only the register reading/writing overhead exists, and the real-time performanceof AI model reasoning is improved.

The embodiment of the invention relates to a processing method and device for a language model, a text generation method and device and a medium. The language model is deployed in the electronic equipment, and a plurality of calculation operations between calculation of a target type in calculation of the same feature layer of the language model are combined into one fusion calculation operation, and the processing method for the language model comprises the following steps: when it is determined that the fusion calculation operation is about to be executed, the CPU of the electronic equipment sends an operation instruction containing the plurality of calculation operations to a GPU (Graphics Processing Unit); in response to receiving the operation instruction, the GPU processes the plurality of computing operations. Therefore, the scheduling overhead between the CPU and the GPU and the repeated read-write overhead of the GPU on the video memory in the processing process of the language model can be effectively reduced, so that the calculation efficiency of the GPU can be effectively improved, the calculation efficiency of the language model is further improved, and the delay of text processing based on the language model is effectively reduced.

The invention discloses a data processing method, system and device and a readable storage medium in the technical field of computers. According to the method, the host and the hardware computing platform are connected through the CXL protocol, so that the host and the hardware computing platform can share the memory, IO and cache of each other, training data does not need to be transmitted through storage media such as a host memory, a GPU cache and a GPU memory, and the hardware computing platform directly reads the training data in the host memory; therefore, the data transmission overhead is reduced. Meanwhile, the hardware computing platform can adjust the learning rate based on the moment moving average current value and then compute new parameters of the model, so that the model parameters can be stabilized, the model precision is guaranteed, and the training efficiency is improved. According to the scheme, the data transmission overhead between the host and the hardware module can be reduced, and the model training efficiency is improved. Correspondingly, the invention also provides a data processing system and device and a readable storage medium, which also have the above technical effects.

The invention provides an online learning type scheduling method based on a container layer dependency relationship in edge computing. The invention relates to edge computing and a deep reinforcement learning method of resource scheduling and machine learning in a distributed system. According to the technical scheme, firstly, modeling is conducted on edge calculation based on the level of a container layer; the task completion time of the user in the edge calculation is considered, and the task completion time comprises the downloading time of a container required by the user task and the running time of the user task. On the basis, an algorithm based on factorization is provided, the dependency relationship of a container layer in edge computing is extracted, and high-dimensional and low-dimensional sparse dependency features in the dependency relationship are extracted. And finally, on the basis of the extracted dependency relationship and task and node resource characteristics, a learning type task scheduling algorithm based on strategy gradient is designed, and thus verifying the whole process through real data. According to the method provided by the invention, resources in the edge computing can be better planned, and the total overhead of tasks of users in an edge computing system and the overhead required for downloading container mirror image files during container running in the edge computing are reduced.

Lazy Persistency (LP), a software persistency method that allows caches to slowly send dirty blocks to the non-volatile main memory (NVMM) through natural evictions. With LP, there are no additional writes to NVMM, no decrease in write endurance, and no performance degradation from cache line flushes and barriers. Persistency failures are discovered using software error detection (checksum), and the system recovers from them by recomputing inconsistent results. LP was evaluated and compared to the state-of-the-art Eager Persistency technique from prior work. Compared to Eager Persistency, LP reduces the execution time and write amplification overheads from 9% and 21% to only 1% and 3%, respectively.

The invention discloses an FPGA-based design method for improving the speed of a BCPNN, and relates to the technical field of artificial intelligence, and the method comprises the steps: updating variables, weights and offsets of synaptic states in the BCPNN on hardware through modular design; and achieving exponential operation on the FPGA through the lookup table. through a parallel algorithm, the speed of the weight and bias updating process of the synaptic state in the BCPNN is increased; through module multiplexing of the adder and the multiplier, the resource overhead is reduced under the condition of keeping the same computing performance. The method provided by the invention not only has higher calculation performance, but also has higher calculation accuracy, and can effectively improve the weight and offset updating speed of the BCPNN.

The invention belongs to the technical field of computers, and discloses a method and a system for improving security multi-party computing efficiency and a storage medium, the method for improving security multi-party computing efficiency is used for a first computing party and a second computing party: the first computing party randomly generates a first fragment of a multiplication triple and generates a random value to cover the fragment; the first computing party sends the generated random value and the masked value to a second computing party; the second computing party receives a random value and a covering value sent by the first computing party, and generates a multiplication triple second fragment; u and v are obtained through calculation with the second fragment, and the u and the v are sent to the first computing party; and the first computing party receives the numerical values u and v sent by the second computing party, calculates c0, and stores the c0 locally. According to the method, the computing overhead and the communication overhead are greatly reduced, and expansibility is achieved; and a strict cryptography theory is used as a support, and it can be guaranteed that privacy is not leaked under the condition that computing parties are not mutually collocated.

The embodiment of the invention discloses a processor device, an instruction execution method thereof and computing equipment. The apparatus comprises one or more single-instruction multi-thread processing units, and the single-instruction multi-thread processing units comprise one or more thread bundles used for executing instructions; a shared register group including a plurality of general purpose registers shared among the thread bundles; a predicate base address register which is arranged corresponding to each thread bundle and is used for indicating a base address of a group of general purpose registers which are used as predicate registers of each thread bundle in the shared register group; wherein each thread bundle performs asserted execution on instructions based on predicate values in the set of general purpose registers used as predicate registers for each thread bundle. According to the embodiment of the invention, the inherent special predicate register of each thread bundle in the original processor architecture can be canceled, the dynamic expansion of the predicate register resource of each thread bundle is realized, the full utilization of processor resources is realized, the overhead of switching instructions is reduced, and the instruction processing performance is improved.

The invention provides a task processing method and device for preventing side channel attacks based on a heterogeneous computing system and a computer readable storage medium. The method comprises the steps of performing task division on a to-be-executed first task to obtain at least two first sub-tasks, determining a random probability number, and dynamically adjusting the at least two first sub-tasks according to the random probability number to obtain at least two to-be-executed first sub-tasks, and allocating the at least two first sub-tasks to be executed to at least one processing unitmatched with the task characteristics, and respectively executing the at least two to-be-executed first sub-tasks by the at least one processing unit corresponding to the at least two to-be-executed first sub-tasks. Through adoption of the method and the device, the side channel attack can be effectively resisted, the resource overhead is reduced, and the information security is improved.

The invention discloses a lightweight high-throughput calculation mode and a fault tolerance method thereof, and particularly relates to the technical field of computer application, and the method comprises the following steps: 1, providing a runtime high-throughput calculation mode; 2, providing an error discrimination method based on a threshold value; and 3, designing a high-flux calculation fault-tolerant technical framework which can achieve efficient fault tolerance of thousands of flux calculation. A father-child task relation model is designed, and the management overhead of super-large-scale concurrent computing tasks is reduced; According to the method, a runtime high-throughput task model is provided for thousands of throughput concurrent calculation modes, a father-child task relation model is designed in a dynamic task expansion mode, and the management overhead of super-large-scale concurrent calculation tasks is reduced; and through the task packaging technology of the job array, the interaction frequency with the supercomputing scheduling system is reduced, and the overhead of job delivery and job state monitoring is further reduced.

The invention discloses a streaming data-oriented cooperative computing method, which comprises the following steps: after local computing of each device is completed, in order to reduce the transmission delay of an intermediate result of each round of iterative transmission and consider that the transmission has random fluctuation, selecting a most suitable edge server; and, in the continuous arrival process of streaming data, carrying out certain local calculation on the equipment, reducing the instantaneous calculation load of the equipment through a local calculation regulation and control mode, and meanwhile, carrying out proper edge server selection during global calculation aggregation, so that the overall cooperative computing overhead is saved under the conditions that the peak load of the equipment is not exceeded and the global computing effect is met.