451 results about "Graphics processing unit" patented technology

The invention discloses a data processing method, a product, electronic equipment and a computer readable storage medium, relates to the technical field of computers, and aims to solve the problems of large GPU (Graphic Processing Unit) space occupation and high access delay in data processing in related technologies. Reading corresponding historical key value cache data from a key value cache of the memory extension equipment according to a historical key value cache data acquisition request sent by a host end; returning the historical key value cache data to the host end, and enabling the host end to process the current input data based on the historical key value cache data by adopting a local model to obtain key value data; and obtaining key value cache data corresponding to the key value data sent by the host end, and sending the key value cache data to a key value cache of the memory extension equipment for storage. The key value cache data is stored in the key value cache of the memory extension equipment, and the historical key value cache data is read from the key value cache, so that the occupation of the memory space of the GPU and the consumption of the memory bandwidth can be reduced, and the access delay is reduced.

The embodiments of the present application relate to the technical field of terminals. Provided are a rendering processing method and an electronic device. The method comprises: determining a plurality of rendering instructions corresponding to first graphic data to be rendered of a first application program, wherein the plurality of rendering instructions include a first rendering instruction; and then reading, from a graphics processing unit (GPU) memory, a first rendering resource corresponding to the first rendering instruction, and executing the plurality of rendering instructions on the basis of rendering resources respectively corresponding to the plurality of rendering instructions, so as to render the first graphic data. In this way, a rendering resource is read from a GPU memory, so as to reduce data interaction between a GPU and a main memory, thereby reducing the bandwidth consumption of a terminal device.

The invention discloses a model performance test method and device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence. The theoretical maximum lexical throughput of a target large language model is calculated based on the video memory bandwidth of a graphics processor, the model parameter quantity, the byte number corresponding to the quantization precision and the video memory bandwidth utilization rate; meanwhile, the benchmark performance throughput is obtained, a theoretical corresponding first concurrency number is calculated in combination with the theoretical maximum lexical unit throughput and the concurrency competition loss coefficient, then the model test is executed based on the first concurrency number to obtain the actual maximum lexical unit throughput and a corresponding second concurrency number, and a model performance test result is generated. The problems that in the prior art, due to the fact that manual testing is conducted depending on manual intervention, a continuous approaching attempt mode is adopted, a reasonable test starting point is not deduced in combination with hardware core bottlenecks and key parameters, evaluation is time-consuming and labor-consuming, the result is prone to being affected by artificial factors, and accuracy and consistency are poor can be solved.

The embodiment of the invention provides a color setting method, a graphic processing unit and a system on chip. The method applied to GPU user mode driving comprises the steps that under the condition that a rendering preparation instruction is received, a first bit field of a sampling descriptor is set as a first identifier, the first identifier is used for pointing to a preset table located in a memory, and the table is used for storing boundary colors with the color format exceeding the bit number of the sampling descriptor; writing a target color carried in the rendering preparation instruction into the preset table; acquiring a second identifier of the target color in the preset table; and writing a second identifier of the target color in the preset table into a second bit field of the sampling descriptor, and sending the sampling descriptor to GPU hardware, so that the GPU hardware realizes texture boundary color setting according to the sampling descriptor. According to the embodiment of the invention, the storage format and precision of the target color can be expanded, so that the setting of the texture boundary color can be improved, and the user experience is improved.

The invention discloses a thread group scheduling method and device for a GPU (Graphics Processing Unit), the GPU and equipment. The method comprises the following steps of: 1) receiving a scheduling request of a thread group; 2) task type priority scheduling; according to a thread group weight value set by a user, obtaining execution priorities of the vertex thread group and the fragment thread group in the current scheduling period; 3) instruction type priority scheduling; pre-analyzing to-be-executed instructions of the thread group, and determining a priority sequence of schedulable instruction types; and 4) priority scheduling of the thread groups: selecting the thread group with the highest priority for scheduling according to the specified task type priority and instruction type priority in combination with the thread group generation time. The invention provides a thread group three-level scheduling strategy so as to improve the instruction throughput rate and the key task response speed of the GPU under the complex load.

The invention relates to a GPU (Graphics Processing Unit) check point storage method in large model distributed training, which comprises the following steps of: 1, loading check point configuration information when training is started; 2, monitoring a back propagation completion signal of each layer in the training process, and immediately triggering check point fragment storage operation of a certain layer after parameter updating of the layer is completed; 3, the model state of the layer is asynchronously copied to a CPU memory from a GPU memory, and consistency verification is carried out; 4, asynchronously storing the check point fragments from the CPU memory into a persistent storage; 5, when training needs to be recovered, the check point file is loaded from the persistent storage, integrity verification is carried out, all layers of fragmented data are recombined into a complete model state, and training is recovered. And the synchronous blocking and I / O bottleneck of the check points in the training process are reduced, so that the training efficiency is improved, and the fault recovery time is shortened.

The invention relates to a GPU-based adaptive compressed data direct query method, system and device and a medium, and the method comprises the steps: carrying out the data compression of input column data through employing an adaptive compression strategy, and obtaining a block-level structure suitable for GPU storage and calculation; loading the compressed data into a GPU memory, and performing Tile-level memory management by taking Tile as a basic scheduling unit; and according to the received query statement, executing direct query of the compressed data on the Tile level through the GPU, and outputting a query result. By designing a Tile-level direct query framework, a hardware-aware memory management and control flow coordination mechanism and a self-adaptive compression strategy, high-performance query execution is realized without decompression, and the computing potential of the GPU is fully released. The method can be widely applied to the technical field of big data processing.

The embodiment of the invention provides a video memory control method, device, equipment and system and a computer storage medium. The method comprises the steps that in response to an obtained drive upgrading request of an image processing unit (GPU), a before-upgrading page table corresponding to the GPU is obtained, and the before-upgrading page table comprises a mapping relation between a physical address and a virtual address of a GPU video memory; determining a data cleaning queue corresponding to the GPU, wherein the data cleaning queue is used for performing global cleaning on data in a GPU video memory; under the condition that the data cleaning queue is locked, driving upgrading operation is carried out on the GPU based on the driving upgrading request, an upgraded page table corresponding to the GPU is obtained, and the upgraded page table comprises a mapping relation between a physical address and a virtual address of a GPU video memory; and regulating and controlling data in the GPU video memory based on the pre-upgrading page table and the post-upgrading page table. In the embodiment of the invention, the driver upgrading operation of the GPU can be realized without operations such as video memory backup, shutdown, restart and the like.

Disclosed in the embodiments of the present application are a key-value cache reuse method, and a related apparatus, which can be applied to scenarios such as cloud technology, artificial intelligence, intelligent transportation, assisted driving and the Internet of Things. The method comprises: after acquiring an ith-round prompt, on the basis of historical token groups respectively generated by a plurality of tokens included in the ith-round prompt with the previous (i-1)th-round prompt, and according to token arrangement orders, matching tokens having the same token arrangement order, so as to obtain a prefix token group and remaining token groups; acquiring from a graphics processing unit a key-value cache of the prefix token group, and if the acquisition fails, acquiring from a central processing unit a key-value cache corresponding to the prefix token group; and sending the key-value cache of the prefix token group and the remaining token groups to an inference engine, such that the inference engine performs inference on the basis of the key-value cache of the prefix token group and the remaining token groups to obtain response content for the ith-round prompt. Thus, an inference engine directly reuses a key-value cache of a prefix token group without requiring recalculation, thereby improving the speed of inference computation.

The invention discloses an Ethernet-based distributed cache coherence protocol, an implementation method, an implementation device and an implementation system. A plurality of computing nodes are connected through a packet switching network. Each computing node comprises a CPU / GPU (Central Processing Unit / Graphics Processing Unit) and a local cache thereof, and is provided with a cache agent. The far-end memory is organized in home nodes, and each home node manages a part of physical address space and is equipped with a directory controller. When the CPU of the computing node accesses a far-end memory address and does not hit in the local cache, the CA of the computing node replaces the far-end memory address and communicates with the DC managing the address through the network so as to maintain the cache consistency of the data among all the nodes. Based on a cache consistency protocol of a directory, the CXL.cache consistency of a plurality of independent computing nodes can be maintained in a low-overhead and high-reliability mode on a high-delay and lossy packet switching network, and broadcast storm caused by a monitoring protocol is avoided.

Described are examples for using multiple graphics processing units (GPUs) to accelerate a database query. Data for a database query can be loaded from the database into memories of multiple GPUs for parallel processing by the multiple GPUs. At least a portion of the data loaded into a memory for one of the multiple GPUs can be moved to a memory for a different one of the multiple GPUs. A compute process can be executed, via parallel processing on the multiple GPUs, for the query to perform data processing related to the database query.

The invention provides a method for executing a computing task through a GPU and a host machine, the GPU is configured on a target host machine, a first confidential virtual machine containing a first client and a second confidential virtual machine containing a first agent program run on the target host machine, and the second confidential virtual machine loads an encryption and decryption kernel into the GPU in advance. The method comprises the following steps: a first client transmits a calculation task to a first agent program through an encryption channel pre-established with the first agent program; the first agent program encrypts the calculation task through a session key negotiated in advance with the encryption and decryption kernel, and sends an obtained first encryption result to the GPU; the GPU decrypts the first encryption result according to the session key, and restores and executes the calculation task to obtain a calculation result; encrypting the calculation result according to the session key, and sending an obtained second encryption result to the first agent program; and the first agent program decrypts the second encryption result according to the session key, and sends the restored calculation result to the first client through the encryption channel.

The invention relates to the technical field of computers, and belongs to a BIOS (Basic Input / Output System) and BMC (Baseboard Management Controller)-based GPU (Graphics Processing Unit) resource intelligent dynamic optimization method and system. A four-layer collaborative architecture of a firmware layer (BIOS), a management controller layer (BMC), a system layer (operating system kernel) and a hardware layer (GPU equipment) is adopted, and the architecture is a cross-layer closed-loop optimization architecture. An optimization decision function is decoupled from an operating system kernel and is deployed in an independent hardware management unit, namely a baseboard management controller (BMC). The system collects GPU operation data in real time through an operation system kernel space and sends the GPU operation data to a BMC; an intelligent optimization engine deployed in the BMC analyzes and calculates the data to generate an optimal GPU resource scheduling strategy; the strategy is returned to an operating system kernel and finally executed by a hardware layer, and intelligent dynamic optimization of GPU resources is achieved through cross-level interaction. Therefore, a continuous intelligent closed-loop optimization link of'deployment-data acquisition-intelligent decision-making-strategy execution 'is formed.

The invention discloses a GPU shader rendering computer implementation method of a finite element result. The method comprises the steps that a static three-dimensional grid is adopted as a rendering geometry; the 32-bit floating point type scalar data corresponding to the vertexes are subjected to lossless coding at a CPU end to obtain four-channel 8-bit RGBA vertex color attributes; during data updating, lightweight vertex color data are only transmitted to the GPU; hardware interpolation is carried out on the encoded color attributes by using a GPU rasterizer; and finally, in the fragment shader, decoding the interpolation result of each pixel to reconstruct a scalar value, and mapping the scalar value into a final color. According to the method, the bottleneck of topological calculation of a CPU end and transmission of massive geometric data to the GPU is avoided, the calculation load is transferred to the GPU in a large scale for parallel processing, the rendering efficiency is remarkably improved, and high-frame-rate dynamic visualization of a finite element result is realized.

The invention provides a large model reasoning method and device, a medium and electronic equipment. In the method, a central processing unit can divide a batch processing data set of a to-be-executed reasoning task into sub-batch processing data sets and distribute the sub-batch processing data sets to a plurality of graphics processors, so that the graphics processors execute reasoning calculation based on the distributed sub-batch processing data sets respectively, and an execution result of the to-be-executed reasoning task is obtained. In the process, each graphic processor can preload the weight parameter of the next reasoning calculation layer from the storage space corresponding to the central processor and store the weight parameter into a video memory in the process of executing the reasoning calculation task corresponding to any reasoning calculation layer; therefore, when the calculation task of the current reasoning calculation layer is not completed, the graphics processing unit can load the weight parameters required by the subsequent reasoning calculation layer into the video memory in advance, so that storage is replaced by passage, and the bandwidth utilization rate of the video memory and the overall reasoning throughput are improved.

The invention relates to a management method and device for server GPU (Graphic Processing Unit) equipment, and the method comprises the following steps: after a server is started, a basic input / output system BIOS writes GPU static asset information through an H2B PCIE (Peripheral Component Interface Express) shared memory; the baseboard management controller BMC obtains GPU static asset information and extended static asset information of the GPU devices through the H2B PCIE shared memory, and polls all the GPU devices; and if the remote client initiates an information acquisition request of the GPU equipment, the northbound interface of the server acquires and feeds back the GPU equipment information through the callback function. According to the method, the basic input and output system BIOS directly and preferentially transmits the GPU equipment information to the shared memory of the GPU equipment information area predefined by the BMC in the starting stage, and the time from starting of the basic input and output system BIOS to obtaining of the basic information of the GPU equipment by the BMC is shortened.

The density and power consumption of modern integrated circuits, such as Graphic Processing Units (GPUs), Central Processing Units (CPUs), and Network Processing Units (NPUs) is growing rapidly, which necessitates designing advanced cooling systems. Existing solutions for characterizing and validating these cooling system are inadequate. A flexible, scalable Thermal Test Vehicle (TTV) is disclosed which is based on an array of power transistors, measurement / control circuitry, and onboard computer. The TTV is configured for characterizing the performance of electronic cooling solutions under a variety of operating conditions.

Disclosed in the embodiments of the present application are a processor, a computing node, and a computing cluster, which are used for simplifying the network architecture of the computing cluster. The processor comprises: a general-purpose protocol port, which is used for supporting a first communication protocol and a second communication protocol, wherein the second communication protocol is different from the first communication protocol; the first communication protocol is a communication protocol of a first network, the second communication protocol is a communication protocol of a second network, the first network is used for supporting communication between a plurality of processors in one computing node including the processor, the second network is used for supporting communication between the processor and other processors comprised in other computing nodes, and the processor is any one of a graphics processing unit (GPU), a tensor processing unit (TPU), a neural processing unit (NPU) and a general-purpose graphics processing unit (GPGPU).

A computing platform may train, using code specific information, temporal information, and compute specific information, a CO2 minimization model to output, for a given processing request, a resource allocation recommendation. The computing platform may input a processing request into the CO2 minimization model to output a resource allocation recommendation indicating processing resources for use in executing the processing request with a minimal amount of CO2 emissions. The computing platform may cause the task execution platform to execute the processing request using the processing resources of the resource allocation recommendation. The computing platform may: dynamically monitor the task execution platform to collect, in real time, metadata associated with execution of the processing request, generate, using the CO2 minimization model, updated resource allocation recommendations based on the real time metadata, and cause the task execution platform to shift the processing for the processing request to the updated resources.

The invention relates to a display method and device, equipment and a chip. The method comprises the following steps: receiving layer information of a to-be-processed layer sent by a graphics processing unit (GPU); determining a target super-resolution multiple according to the layer information; determining a rendering resolution according to the target super-resolution multiple and the display resolution; sending the rendering resolution to a GPU, so that the GPU generates a rendered image according to the rendering resolution and the to-be-processed layer; and performing super-magnification display on the rendered image based on the target super-resolution multiple. By adopting the method, the optimal balance between the performance and the display effect is realized.

Some embodiments provide a method of managing communication between graphics processing units (GPUs) through a network having multiple forwarding elements. The method uses a set of servers to generate instructions for specifying forwarding behaviors for the forwarding elements to forward data messages containing results of computations performed by the GPUs. The GPUs perform the computations in order to collectively execute a distributed application. The method distributes the generated instructions to the forwarding elements to configure the forwarding elements to implement the forwarding behavior.

The invention provides a hardware resource scheduling method and a hardware scheduler used in a graphics processing unit. The hardware resource scheduling method used in the graphics processing unit comprises the following steps: for any hardware computing unit in the graphics processing unit, acquiring task or instruction queue state information indicating whether a task or instruction queue of the hardware computing unit is in an idle state or a full state from the hardware computing unit; receiving, from a driver system of the graphics processing unit, priority ranking information indicating a priority ranking between respective hardware computing units inside the graphics processing unit; and based on the task or instruction queue state information and the priority ranking information of the hardware computing unit, judging whether the task or instruction to be scheduled to the hardware computing unit is scheduled to the hardware computing unit, and if so, scheduling the task or instruction to be scheduled to the hardware computing unit to the hardware computing unit.

The embodiment of the invention provides a GPU computing power scheduling method and device, equipment, a medium and a program product, and relates to the technical field of computing power scheduling, and the method comprises the steps: dividing a computing unit and video memory resources of a GPU into a plurality of dynamically adjustable partitions according to a real-time task demand; calculating the priority weight of each task; preemptive scheduling and cooperative scheduling are combined, resource preemption is carried out on the first priority task, and time division multiplexing is carried out on the non-preemptive task; executing a dynamic voltage frequency adjustment (DVFS) and task merging strategy; through dynamic resource partitioning, dynamic priority evaluation, a hybrid scheduling strategy and an energy efficiency optimization algorithm, the GPU resource utilization rate and the energy efficiency ratio in a multi-task scene are remarkably improved, and meanwhile, energy consumption is reduced by combining a DVFS and a task merging strategy.

This invention discloses a method, device, medium, and system for generating and using 3D character culling data in computer graphics, belonging to the field of computer graphics. The method includes the following steps: pre-generating frame-level cluster bounding box data corresponding to sampled animation frames; submitting the frame-level cluster bounding box data to the graphics processing unit of the graphics card during initialization; updating the current animation frame using an animation state machine during runtime; during the culling operation, searching for the bounding box data of the cluster in the current animation frame based on the current animation frame and the cluster index, and culling it; if the current animation frame does not match the sampled animation frame, interpolating and calculating the bounding box data corresponding to the current animation frame based on the pre-generated frame-level cluster bounding box data of the animation frames before and after the current animation frame. This invention has advantages such as low overhead, good dynamics, and high culling performance.

The invention provides a multi-shape batch rendering method suitable for a GPU without hardware batch processing capability, and relates to the technical field of computer graphic rendering, the method comprises the following steps: pre-defining path drawing instruction templates of various graphic primitives and text characters; traversing and analyzing all to-be-rendered graphics and text elements of the current frame, and uniformly converting the to-be-rendered graphics and text elements into a path instruction sequence and coordinate parameters; separately storing the instruction and the parameter in a global array; sorting and combining all elements according to the rendering hierarchy to generate a global instruction stream and a parameter stream; and finally, packaging the complete instruction stream and the parameter stream into a single drawing command, and submitting the single drawing command to a GPU (Graphics Processing Unit) through one-time calling to complete the whole-frame rendering. According to the method, unified batch processing of graphs and texts on the path level can be realized, the GPU calling times are remarkably reduced, and the interface rendering efficiency and smoothness on resource-limited equipment such as smart watches and health bracelets are greatly improved.

The embodiment of the invention provides a GPU debugging method and device, electronic equipment, a readable storage medium and a computer program product, and relates to the technical field of computers. The method comprises the following steps: identifying a plurality of address fields in a command stream received by a GPU (Graphics Processing Unit); and executing a read operation on data specified by a selected address field in the plurality of address fields, and inhibiting a read operation on data specified by an unselected address field in the plurality of address fields, the plurality of address fields representing addresses of different data buffers in the GPU memory. According to the method, only the data specified by the selected address field in the command stream is read, and the reading of the data specified by the unselected address field is inhibited, so that the data reading amount is remarkably reduced, and the debugging efficiency is improved.

A rendering method and a rendering apparatus. The rendering method comprises the following steps: receiving a diagram file; interpreting the diagram file to generate a plurality of geometric units and, on the basis of the geometric units, generating a plurality of vertices and a plurality of pieces of entity data; on the basis of the vertices, a graphics processing unit generating a multi-layer picture that contains a plurality of geometric graphics, and executing mouse collision detection between mouse coordinates and the vertices; on the basis of a result of the mouse collision detection, determining that the mouse coordinates contact at least one of the plurality of vertices and, on the basis of the at least one of the plurality of vertices and at least one of the plurality of pieces of entity data corresponding to the at least one of the plurality of vertices, determining a corresponding line segment; and, on the basis of the corresponding line segment, displaying a prompt display effect in the picture. FIG. 2

The embodiment of the invention discloses a GPU chip, a ray tracing method, a graphics card and computer equipment, and relates to the field of GPU graphic rendering. The GPU chip comprises a GPU core, an on-chip cache and a cache controller, the on-chip cache is used for storing ray tracing data required by ray tracing calculation; the GPU core is used for sending a data reading instruction to the cache controller; the cache controller is used for reading the ray tracing data from the on-chip cache based on the data reading instruction and sending the ray tracing data to the GPU core; and the GPU core is used for executing ray tracing calculation based on the ray tracing data. By adopting the GPU chip provided by the invention, a large amount of occupation of a GPU core in ray tracing calculation can be avoided, and a relatively high data access speed is achieved.

The invention relates to the technical field of data communication, and provides a method and device for managing graphics processing unit (GPU) equipment and a computer product, and the method comprises the following steps: determining a target container in a target container group (Pod) of a node (Node); entering a resource isolation environment of the target container based on the first related information of the target container; and in the resource isolation environment of the target container, creating GPU equipment or deleting the GPU equipment, and modifying the access authority of the GPU equipment. Therefore, the target container can dynamically mount or unload the GPU equipment during load running, the use flexibility of the GPU equipment is improved, and the GPU equipment is reasonably distributed.

The present disclosure describes a primitive processing method for a rasterizer stage, a graphic process unit, a computer-readable storage medium, and a computer program product. The method includes: acquiring a primitive and vertex information of the primitive; determining a type of the primitive according to the vertex information; removing the primitive when the type of the primitive is a to-be-removed primitive; and when the type of the primitive is a non-to-be-removed primitive, determining a pixel attribute of the primitive according to the vertex information of the primitive, wherein the pixel attribute is configured to be inputted into a pixel shader. The method can reduce the computational pressure of the graphic process unit during the rasterizer stage, and improve the performance of the GPU.