A blockchain-based AI computing control method, system and device
Through the blockchain AI computing management and control method, GPU heterogeneous pooling design and CXL technology are used to solve the problem of excessive AI computing load on private chain servers, and the reasonable allocation and efficient processing of resources are achieved, which improves response efficiency and user experience. .
Patent Information
- Application Number
- CN202411162335.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-23
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-08-23
AI Technical Summary
The load of blockchain servers in private chains is too high when performing AI calculations, resulting in excessive loss of computing power resources and low response efficiency, affecting user experience.
Through the blockchain-based AI computing control method, the parameter set and demand computing power value of the AI processing request are determined, matched and allocated to the idle GPU computing unit in the appropriate distributed computing pool for task processing, and use GPU heterogeneous pooling to achieve GPU Design and CXL technology for unified management and communication.
It effectively reduces the computing load of blockchain servers, improves AI computing response efficiency, avoids waste of computing power resources, improves user experience, and supports high-performance blockchain data processing.
Smart Images

Figure CN118674610B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of graphics processing unit (GPU) processing technology, and in particular to a blockchain-based AI computing management and control method, system and device. Background Art
[0002] Blockchain is a bookkeeping technology that is jointly maintained by multiple parties, uses cryptography to ensure transmission and access security, can achieve consistent data storage, is difficult to tamper with, and prevents denial. It is also called Distributed Ledger Technology (DLT). Blockchain is divided into the following three categories: public chain, private chain and consortium chain. Among them, private chain refers to a chain that is open to a single individual or entity.
[0003] At present, private chain users also have AI computing needs that need to be met. If the users' many AI computing needs are concentrated on the blockchain server, it will inevitably lead to excessive server load, excessive consumption of computing resources, and affect the execution of other important applications on the server. At the same time, the accumulation of server processes will also lead to low efficiency of blockchain user AI computing response, extremely long user waiting time, and a poor user experience. Summary of the invention
[0004] To solve the above problems, the embodiments of the present application provide a blockchain-based AI computing management method, system and device.
[0005] On the one hand, an embodiment of the present application provides an AI computing management and control method based on blockchain, which is applied to a private chain; the method includes:
[0006] Based on the AI processing request information from the blockchain user node, determine the corresponding AI processing parameter set; wherein the AI processing parameter set at least includes a model scale parameter, a required number of parameters, a precision calculation complexity coefficient, and a reference model coefficient group;
[0007] Determine the required computing power value corresponding to the AI processing request information based on the AI processing parameter set and the preset AI computing power judgment algorithm;
[0008] According to the preset matching rules, the required computing power value is matched with the critical computing power list corresponding to each current blockchain distributed computing pool, so as to determine the allocation target computing pool according to the matching result; wherein the preset matching rule is used to match one or more target pools to be allocated that at least reach the required computing power value; one of the blockchain distributed computing pools includes multiple GPU computing units;
[0009] The AI processing task corresponding to the AI processing request information is allocated to the idle GPU computing unit in the allocation target computing pool, so that the AI processing result after the idle GPU computing unit executes the AI processing task is sent to the blockchain user node.
[0010] In one implementation of the present application, based on the AI processing parameter set and the preset AI computing power judgment algorithm, determining the required computing power value corresponding to the AI processing request information specifically includes:
[0011] Input the model scale parameter, the required number of parameters, the precision operation complexity coefficient and the reference model coefficient group into the preset AI computing power judgment algorithm to calculate the corresponding operation parameter complexity parameter; the reference model coefficient group includes at least a first coefficient representing the scale of the reference model and a second coefficient representing the number of reference model parameters;
[0012] The required computing power value is determined according to the product value of the preset server CPU main frequency parameter, the preset maximum GPU operation parameter and the operation parameter complexity parameter; wherein the preset maximum GPU operation parameter is the maximum value of the GPU operation parameter corresponding to each of the GPU computing units in the current blockchain distributed computing pools.
[0013] In one implementation of the present application, the model scale parameter, the required parameter quantity, the precision operation complexity coefficient and the reference model coefficient group are input into the preset AI computing power judgment algorithm to calculate the corresponding operation parameter complexity parameter, specifically including:
[0014] Determine, by using the preset AI computing power judgment algorithm, a division result of the model scale parameter and a preset maximum reference model scale value as a first intermediate value;
[0015] Determine a second intermediate value as a result of a division operation between the required number of parameters and a preset maximum number of parameters;
[0016] Calculate the product of the first intermediate value and the first coefficient as a third intermediate value, and calculate the product of the second intermediate value and the second coefficient as a fourth intermediate value;
[0017] The sum of the third intermediate value and the fourth intermediate value is calculated, and the product of the sum and the precision operation complexity coefficient is determined as the operation parameter complexity parameter.
[0018] In one implementation of the present application, according to a preset matching rule, the required computing power value is matched with the critical computing power list corresponding to each current blockchain distributed computing pool, so as to determine the allocation target computing pool according to the matching result, specifically including:
[0019] Determine a critical computing power value in the critical computing power list that is greater than or equal to the required computing power value;
[0020] In order from small to large, the identifiers of the blockchain distributed computing pools corresponding to the critical computing power values are added to the sequence of target pools to be allocated;
[0021] According to the sequence of the target pools to be allocated, the idle state of each of the GPU computing units in each of the target pools to be allocated is polled in turn, so as to determine the allocated target computing pool according to the polling result.
[0022] In one implementation of the present application, the idle state of each GPU computing unit in each target pool to be allocated is polled in sequence to determine the allocated target computing pool according to the polling result, specifically including:
[0023] Determine whether there is an idle GPU computing unit in each of the GPU computing units in the target pool to be allocated currently being polled;
[0024] If so, determining that the currently polled target pool to be allocated is the allocated target computing pool;
[0025] Otherwise, the next target pool to be allocated after the currently polled target pool to be allocated in the sequence of target pools to be allocated is used as the currently polled target pool to be allocated, until the allocation target computing pool is determined.
[0026] In one implementation of the present application, the blockchain user node exists in a blockchain user cluster composed of multiple blockchain user nodes; one blockchain user cluster corresponds to one user server;
[0027] The blockchain server connects each of the user servers and each of the blockchain distributed computing pools through a computing fast link switch CXL Switch, so as to receive the AI processing request information and distribute the AI processing request information through the CXL Switch.
[0028] In one implementation of the present application, the performance parameters of the GPU computing units in the same blockchain distributed computing pool are the same; the performance parameters of the GPU computing units between the blockchain distributed computing pools are different; and the GPU computing units in the same blockchain distributed computing pool are connected via an advanced expansion interface AXI bus.
[0029] In one implementation of the present application, before determining the corresponding AI processing parameter set based on the AI processing request information from the blockchain user node, the method further includes:
[0030] Obtaining user uploaded data from the blockchain user node;
[0031] Determine whether there is an AI computing demand based on the user uploaded data and the preset AI computing screening model; wherein the preset AI computing screening model is a model trained based on a number of user data samples annotated with AI computing demand labels;
[0032] If it exists, the user uploaded data is used as the AI processing request information.
[0033] On the other hand, the embodiment of the present application also provides an AI computing management and control system based on blockchain, the system comprising:
[0034] A first determination module is used to determine a corresponding AI processing parameter set based on the AI processing request information from the blockchain user node; wherein the AI processing parameter set at least includes a model scale parameter, a required parameter quantity, a precision calculation complexity coefficient, and a reference model coefficient group;
[0035] A second determination module, configured to determine a required computing power value corresponding to the AI processing request information based on the AI processing parameter set and a preset AI computing power judgment algorithm;
[0036] A matching module is used to match the required computing power value with the critical computing power list corresponding to each current blockchain distributed computing pool according to a preset matching rule, so as to determine the target computing pool for allocation according to the matching result; wherein the preset matching rule is used to match one or more target pools to be allocated that at least reach the required computing power value; one of the blockchain distributed computing pools includes multiple GPU computing units;
[0037] An allocation module is used to allocate the AI processing task corresponding to the AI processing request information to the idle GPU computing unit in the allocation target computing pool, so as to send the AI processing result after the idle GPU computing unit executes the AI processing task to the blockchain user node.
[0038] On the other hand, the embodiment of the present application also provides an AI computing management and control device based on blockchain, the device comprising:
[0039] At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a blockchain-based AI computing management and control method as described above.
[0040] Compared with the prior art, the present invention has the following significant effects:
[0041] 1) This application uses the above solution to complete AI calculations using a blockchain distributed computing pool, which can solve the problem of excessive AI computing load on current blockchain servers, resulting in excessive consumption of computing resources. At the same time, it can avoid the accumulation of blockchain servers, improve the corresponding efficiency of AI calculations, and improve the user experience.
[0042] 2) Based on the performance of GPU computing units, group interconnection is used for unified management to achieve GPU heterogeneous pooling design. A blockchain GPU heterogeneous computing resource pool system architecture with two levels, distributed computing pool and GPU computing unit, is constructed, which supports multi-pool interconnection and can be expanded arbitrarily.
[0043] 3) Transferring the computing needs of blockchain users’ AI applications to the constructed blockchain distributed computing pool to achieve unified computing, which greatly reduces the computing load of the blockchain server and facilitates the management of the blockchain server.
[0044] 4) CXL technology is also used to realize the communication between blockchain servers, user servers, and several blockchain distributed computing pools, making data transmission and communication more efficient and more stable.
[0045] 5) Grouping based on GPU computing unit performance, coupled with the application and scheduling strategy of a scientific and reasonable AI computing resource pool architecture, can avoid waste of GPU computing resources, better meet the needs of high-performance exchange and processing of blockchain data in the context of the big data era, and significantly improve system performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:
[0047] Figure 1 A flowchart of a blockchain-based AI computing control method in an embodiment of the present application;
[0048] Figure 2 This is a structural diagram of a blockchain data management system corresponding to a blockchain-based AI computing control method in an embodiment of the present application;
[0049] Figure 3 A schematic diagram of a process for determining and allocating a target computing pool in a blockchain-based AI computing management and control method in an embodiment of the present application;
[0050] Figure 4 A schematic diagram of computing task execution communication in a blockchain-based AI computing management and control method in an embodiment of the present application;
[0051] Figure 5This is another flowchart of a blockchain-based AI computing control method in an embodiment of the present application;
[0052] Figure 6 This is a schematic diagram of the structure of a blockchain-based AI computing management and control system in an embodiment of the present application;
[0053] Figure 7 This is a structural diagram of a blockchain-based AI computing and control device in an embodiment of the present application. DETAILED DESCRIPTION
[0054] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.
[0055] The embodiments of the present application provide a blockchain-based AI computing management and control method, system and device to solve the problem that the AI computing load of the blockchain server in the current private chain is too high, resulting in excessive consumption of computing resources, and easily causing process accumulation, affecting the AI computing response efficiency, and reducing the user experience.
[0056] The blockchain-based AI computing control method of this application is based on a new GPU heterogeneous pooling design concept, which is grouped and interconnected based on the performance of GPU AI computing units and managed in the same way. It includes a blockchain distributed computing pool and a GPU computing unit at two levels of the blockchain GPU heterogeneous computing resource pool system architecture, which supports multi-pool interconnection and can be expanded arbitrarily.
[0057] The following describes in detail various embodiments of the present application in conjunction with the accompanying drawings.
[0058] The present application embodiment provides a blockchain-based AI computing management and control method, which is applied to a private chain, such as Figure 1 As shown, the method may include steps S101-S104:
[0059] S101, the blockchain server determines a corresponding AI processing parameter set based on the AI processing request information from the blockchain user node.
[0060] Among them, the AI processing parameter set includes at least model scale parameters, required parameter quantity, precision calculation complexity coefficient, and reference model coefficient group.
[0061] It should be noted that the blockchain server, as the executor of the blockchain-based AI computing control method, is in a private chain together with the blockchain distributed computing pool and user server.
[0062] The blockchain data management system based on Compute Express Link (CXL) technology in this application consists of a blockchain server, several user servers, several blockchain distributed computing pools, several blockchain user clusters and a connection component computing express link switch CXL Switch. The connection relationship is as follows: Figure 2 As shown. Among them, each blockchain distributed computing pool is composed of several GPU computing units (GPU units in the figure) with the same performance (i.e., the same performance parameters). The performance parameters can be set by the user according to actual needs, such as core architecture, number of stream processors, video memory capacity and type, video memory bandwidth, core frequency and acceleration frequency, etc. The performance of different blockchain distributed computing pools can be differentiated, so as to meet the diversified AI computing needs and avoid waste of computing resources. In other words, the performance parameters of each GPU computing unit in the same blockchain distributed computing pool are the same. The performance parameters of the GPU computing units between different blockchain distributed computing pools are different. Among them, the GPU computing units in the same blockchain distributed computing pool are connected through an Advanced Extensible Interface (AXI) bus.
[0063] Among them, the blockchain server (CXL Type 2 devices) is responsible for the overall management and control of all user information of the blockchain. It is connected to the CXL Switch through CXL3.1, and communicates with the blockchain distributed server across the switch through the three sub-protocols of CXL protocol, CXL.io, CXL.mem, and CXL.cache. CXL.io is a physical layer interface defined in the specification, and the ".io" in its name usually stands for input / output. CXL.mem is a key component of the CXL protocol, which allows the processor to directly access remote memory through the CXL interconnect. CXL.cache is another important extension of the CXL protocol, which focuses on achieving cache consistency between processors and accelerators. The blockchain server, each user server, and each blockchain distributed computing pool communicate across the CXLSwitch through the three sub-protocols of CXL protocol, CXL.io, CXL.mem, and CXL.cache.
[0064] Each private chain can deploy n computing pools, where n is a natural number and can be expanded according to actual user needs. The blockchain distributed computing pool consists of several GPU computing units (CXL Type 1 / 2 devices) with the same performance, which are responsible for processing the AI computing tasks of the blockchain (such as deep learning, convolutional neural network computing, scientific simulation computing, etc.). GPU computing units are such as GPU unit 1, GPU unit 2...GPU unit n. The type of GPU computing unit can be set according to the actual needs of the blockchain; the AI computing unit of the distributed computing pool is connected to the CXL Switch through the PCIe5.0 interface, and is uniformly scheduled and controlled by the blockchain database management server, responsible for the AI processing and computing work of the current blockchain.
[0065] The user server (CXL Type 2 devices) in the blockchain is responsible for the overall collection of AI processing requests of the blockchain user nodes in the current blockchain user cluster, data format conversion processing, and sending the computing requirements to the blockchain server through the CXL Switch; it is connected to the CXL Switch through CXL3.1, and communicates with other blockchain devices across the Switch through the three sub-protocols of CXL.io, CXL.mem, and CXL.cache in the CXL protocol.
[0066] Since there are too many user nodes in the entire blockchain, it is necessary to divide the blockchain user nodes into several blockchain user clusters according to the computing power requirements of the user server. Each blockchain user cluster is composed of several blockchain user nodes and is directly managed by the corresponding blockchain distributed server; each user node terminal device (such as PC, mobile phone, tablet computer) communicates with the user server through the network, and these node devices transmit the current user's AI computing needs to the distributed server in real time, and the server performs subsequent data processing operations.
[0067] In the embodiment of the present application, the blockchain user node exists in a blockchain user cluster composed of multiple blockchain user nodes. One blockchain user cluster corresponds to one user server. The blockchain server connects each user server and each blockchain distributed computing pool through a CXL Switch to receive AI processing request information and distribute AI processing request information.
[0068] For example, user node 1 issues an AI processing request, which is sent through the user server and CXL Switch, so that the CPU (Central Processing Unit) of the blockchain server responds to the AI processing request information, stores the task-related data corresponding to the AI processing request information in the AI computing task queue, and stores the task execution-related data in the relevant storage partition of the server storage, waiting for the blockchain distributed computing pool to read and execute. At the same time, the blockchain server determines the AI processing parameter set in order to make an AI computing power judgment in order to allocate computing tasks. Task execution-related data, for example, is the image pixel data involved when the GPU performs image processing based on a convolutional neural network.
[0069] In addition, before determining the corresponding AI processing parameter set based on the AI processing request information from the blockchain user node, the present application also includes:
[0070] Obtain user uploaded data from blockchain user nodes; determine whether there is an AI computing demand based on the user uploaded data and the preset AI computing screening model; wherein the preset AI computing screening model is a model trained based on several user data samples annotated with AI computing demand labels; when it is determined that there is an AI computing demand, use the user uploaded data as AI processing request information. When it is determined that there is no AI computing demand, it is processed in the way that private chain data is uploaded to the chain, without the need for AI computing control and allocation. The preset AI computing screening model can be a machine learning model or a neural network model, which is not specifically limited in this application.
[0071] In other words, the present application can perform an identification process on the data uploaded by the user, avoid ineffective AI calculations, and improve the AI computing utilization rate of the blockchain distributed computing pool of the present application.
[0072] S102, the blockchain server determines the required computing power value corresponding to the AI processing request information based on the AI processing parameter set and the preset AI computing power judgment algorithm.
[0073] In the embodiment of the present application, based on the AI processing parameter set and the preset AI computing power judgment algorithm, the required computing power value corresponding to the AI processing request information is determined, specifically including:
[0074] The blockchain server inputs the model scale parameter, the number of required parameters, the precision operation complexity coefficient and the reference model coefficient group into the preset AI computing power judgment algorithm to calculate the corresponding operation parameter complexity parameter. The reference model coefficient group includes at least a first coefficient representing the scale of the reference model and a second coefficient representing the number of reference model parameters. The required computing power value is determined according to the product value of the preset server CPU main frequency parameter, the preset maximum GPU operation parameter and the operation parameter complexity parameter. Among them, the preset maximum GPU operation parameter is the maximum value of the GPU operation parameter corresponding to each GPU computing unit in the current blockchain distributed computing pool. The GPU operation parameter can be, for example, the number of floating-point operations per second.
[0075] Furthermore, the above inputs the model scale parameter, the number of required parameters, the precision calculation complexity coefficient and the reference model coefficient group into the preset AI computing power judgment algorithm to calculate the corresponding calculation parameter complexity parameters, specifically including:
[0076] The division result of the model scale parameter and the preset maximum reference model scale value is determined by the preset AI computing power judgment algorithm. , is the first intermediate value. Determine the result of the division operation between the required number of parameters and the preset maximum number of parameters , is the second intermediate value. The product of the first intermediate value and the first coefficient is calculated as the third intermediate value , calculate the product of the second intermediate value and the second coefficient as the fourth intermediate value . Calculate the sum of the third intermediate value and the fourth intermediate value, and determine the product of the sum and the precision operation complexity coefficient , is the operation parameter complexity parameter.
[0077] Among them, the formula of the preset AI computing power judgment algorithm is as follows:
[0078] ;
[0079] ;
[0080] ;
[0081] in, The required computing power value. To preset the server CPU main frequency parameters. It is the preset maximum GPU computing parameter. is the operation parameter complexity parameter. is the model scale parameter. The preset maximum reference model scale value. is the first coefficient. is the number of required parameters. The maximum number of preset parameters. is the second coefficient. is the precision operation complexity coefficient. It can be confirmed based on the calculation accuracy of the AI operation (such as FP32, FP64, int32, etc.), and is a value set by the user based on actual use. This application does not make any specific limitations on this.
[0082] The above-mentioned preset server CPU main frequency parameters can be set according to the CPU of the actual blockchain server, and this application does not make specific restrictions on this; the preset maximum GPU computing parameters can be understood as the GPU processor computing power of the computing pool with the largest computing power among all blockchain distributed computing pools in the current blockchain, which can be determined according to the actual selection, and no specific restrictions are made here; the preset maximum reference model scale value and the preset maximum number of parameters are set by the user according to actual needs, and this application does not make specific restrictions on this. The first coefficient and the second coefficient are set by the user according to different reference models and different reference model parameter quantities according to needs, and this application does not make specific restrictions on this.
[0083] Through the above scheme, the blockchain distributed computing pool can be quickly and uniformly calculated, so that the on-chain AI application calculations originally performed by the blockchain server can be transferred to the blockchain distributed computing pool in an orderly and efficient manner according to the calculation results. Based on the above scheme, the foundation is laid for realizing reasonable scheduling of computing resource pools.
[0084] S103, the blockchain server matches the required computing power value with the critical computing power list corresponding to each current blockchain distributed computing pool according to the preset matching rules, so as to determine the target computing pool for allocation according to the matching results.
[0085] The preset matching rules are used to match one or more target pools to be allocated that at least reach the required computing power value. A blockchain distributed computing pool includes multiple GPU computing units.
[0086] In the embodiment of the present application, according to the preset matching rules, the required computing power value is matched with the critical computing power list corresponding to each current blockchain distributed computing pool, so as to determine the allocation target computing pool according to the matching result, specifically including:
[0087] The blockchain server determines the critical computing power value in the critical computing power list that is greater than or equal to the required computing power value. In order from small to large, the identifiers of each blockchain distributed computing pool corresponding to each critical computing power value are added to the sequence of target pools to be allocated. According to the sequence of target pools to be allocated, the idle state of each GPU computing unit in each target pool to be allocated is polled in turn to determine the allocated target computing pool according to the polling results.
[0088] That is to say, this application can allocate blockchain distributed computing pools by calculating the required computing power value corresponding to the AI processing request information. The above critical computing power value is set according to the actual GPU computing unit, and this application does not make specific restrictions on this. The sequence of target pools to be allocated includes one or more target pools to be allocated.
[0089] Further, the idle state of each GPU computing unit in each target pool to be allocated is polled in turn to determine the target computing pool to be allocated according to the polling result, which specifically includes:
[0090] Determine whether there are idle GPU computing units in the currently polled target pool. If it is determined that there are idle GPU computing units in the currently polled target pool, determine the currently polled target pool to be allocated as the allocated target computing pool. Otherwise, the next target pool to be allocated of the currently polled target pool to be allocated in the sequence of target pools to be allocated is used as the currently polled target pool to be allocated until the allocated target computing pool is determined.
[0091] In other words, the blockchain server can first allocate a blockchain distributed computing pool that meets the required computing power value and is closest to the required computing power value to execute the computing task corresponding to the AI processing request information; if the GPU computing units in the blockchain distributed computing pool are not idle, then allocate a blockchain distributed computing pool with greater computing power to execute the computing task.
[0092] Through the above solution, excessive waste of computing resources can be avoided, the blockchain distributed computing pool can be applied and scheduled scientifically and reasonably, and the system performance can be improved.
[0093] In an embodiment of the present application, after the blockchain server determines the target computing pool to be allocated, it can send AI query requirements (i.e., AI processing request information) and query source data (i.e., task execution related data) through the CXL bridge.
[0094] Figure 3 A flowchart of using a preset AI computing power judgment algorithm to determine the target computing pool for allocation in an embodiment of the present application is shown in FIG. Figure 3 As shown, the following steps are included:
[0095] S301, the blockchain server CPU extracts the AI processing parameter set including the model scale parameter, required parameter quantity, precision calculation complexity coefficient, parameter model coefficient group, etc. of the current AI processing;
[0096] S302, inputting the parameters of the extracted AI processing parameter set into the AI computing power judgment algorithm to calculate the required computing power value;
[0097] S303, comparing the required computing power value with the critical computing power of the GPU computing units in each blockchain distributed computing pool, and selecting a target pool to be allocated that is suitable for the AI processing task;
[0098] S304, query the status of the GPU computing units in the target pool to be allocated, and poll the idle status of all GPU computing units contained in each target pool to be allocated according to the number. After confirming that there is no error, send the AI query request and query source data through the CXL bridge.
[0099] Through the above solution, AI processing tasks are reasonably allocated to the GPU computing units in the blockchain distributed computing pool for processing, alleviating the computing pressure of the blockchain server.
[0100] S104, the blockchain server allocates the AI processing task corresponding to the AI processing request information to the idle GPU computing unit in the allocation target computing pool, so as to send the AI processing result after the idle GPU computing unit executes the AI processing task to the blockchain user node.
[0101] In an embodiment of the present application, the blockchain server can assign AI processing tasks to idle GPU computing units in the assigned target computing pool. The idle GPU computing units are connected to the blockchain server through the high-speed serial computer expansion bus standard (Peripheral Component Interconnect Express, PCIe) 5.0 interface via the CXL3.1 transmission link to realize information exchange. The GPU computing unit receives and stores AI processing tasks from the blockchain server through the DDR5 peripheral. The GPU on-chip logic organically combines a series of complex data reception (i.e., Data Rx), value acquisition, decoding, execution of AI operations (i.e., deep learning, convolutional neural networks, scientific simulation calculations, etc.), result collection, re-encoding, write-back, and result sending (i.e., Data Tx) and other hierarchical applications to finally complete the AI processing task. The communication process of computing task execution is as follows: Figure 4 The figure only shows 9 GPU computing units, namely GPU computing unit 0 to GPU computing unit 8, and the number of GPU computing units in actual use is not specifically limited.
[0102] When the blockchain server receives the user's AI processing task, it assigns the corresponding task to the qualified blockchain distributed computing pool and the specific internal GPU computing unit, and notifies it through CXL; then, the specific GPU computing unit reads the instruction signal stored in the specific address of the server main memory and the data set to be processed to the DDR5 peripheral, and the corresponding AI calculation can be performed after the reading is completed; after the calculation is completed, the calculation result is written back to the DDR5 memory in time, and the corresponding blockchain user server is notified to read the AI processing result. In order to realize the AI processing result after the idle GPU computing unit executes the AI processing task and sends it to the blockchain user node.
[0103] Figure 5 Another flowchart of the AI computing control method based on blockchain in the embodiment of the present application is shown in FIG. Figure 5 As shown, the following steps are included:
[0104] S501: The user terminal sends AI processing request information to the user cluster server.
[0105] User terminals are blockchain user nodes, including but not limited to user mobile phones, computers and other devices, and user cluster servers are user servers.
[0106] S502, the blockchain server receives AI processing request information from the user server through CXL.
[0107] S503, the blockchain server CPU stores the corresponding task-related data to the AI computing task sequence, and stores the task execution-related data to the relevant storage partition of the blockchain server storage, waiting for the blockchain distributed computing pool to read and execute.
[0108] S504, the blockchain server CPU executes the AI computing power judgment algorithm to determine the required computing power value for the current task execution, and allocates a blockchain distributed computing pool that matches the required computing power to the current task.
[0109] S505, query the status of the target pool to be allocated, and poll the idle status of all GPU computing units included in the computing pool according to the numbers.
[0110] S506: Determine whether there is an idle GPU computing unit.
[0111] S507: If not, assign the task to a higher-level distributed computing pool that can meet the required computing power.
[0112] A higher level can be understood as the critical computing power value being greater than the target pool to be allocated with no idle GPU computing units currently.
[0113] S508: If yes, send the current task to the GPU computing unit m included in the blockchain distributed computing pool n through the CXL bridge.
[0114] The blockchain distributed computing pool n can be understood as the allocation target computing pool, and the GPU computing unit m is the idle GPU computing unit in it. Where n and m are natural numbers.
[0115] S509, the target computing pool is allocated to read the task execution related data stored in the main memory of the blockchain server through the PCIe interface, and execute the current task.
[0116] S5010, after the current task is completed, the AI processing result is sent to the user server according to the address through the CXL bridge, and the user server sends the AI processing result to the user terminal through the network.
[0117] This application uses the above solution to complete AI calculations using a blockchain distributed computing pool, which can solve the problem of excessive AI computing load on current blockchain servers, resulting in excessive consumption of computing resources. At the same time, it can avoid the accumulation of blockchain servers, improve the corresponding efficiency of AI calculations, and improve the user experience.
[0118] At the same time, the GPU computing unit performance is used as the basis for grouping and interconnection for unified management to realize GPU heterogeneous pooling design. A blockchain GPU heterogeneous computing resource pool system architecture including two levels of distributed computing pool and GPU computing unit is constructed, which supports multi-pool interconnection and can be expanded arbitrarily. And the blockchain user AI application computing needs are transferred to the constructed blockchain distributed computing pool to realize unified computing, which greatly reduces the computing load of the blockchain server and facilitates the management of the blockchain server. In addition, CXL technology is also used to realize the communication between blockchain servers, user servers, and several blockchain distributed computing pools, with higher data transmission and communication efficiency and stronger stability. This application also groups based on the performance of GPU computing units, and uses scientific and reasonable AI computing resource pool architecture application and scheduling strategies to avoid waste of GPU computing power resources, which can better meet the needs of high-performance exchange and processing of blockchain data in the context of the big data era, and significantly improve system performance.
[0119] Figure 6 A schematic diagram of the structure of a blockchain-based AI computing control system provided in an embodiment of the present application is shown in FIG. Figure 6 As shown, the blockchain-based AI computing management and control system 600 includes:
[0120] The first determination module 601 is used to determine the corresponding AI processing parameter set based on the AI processing request information from the blockchain user node. Among them, the AI processing parameter set includes at least a model scale parameter, a required number of parameters, a precision calculation complexity coefficient, and a reference model coefficient group. The second determination module 602 is used to determine the required computing power value corresponding to the AI processing request information based on the AI processing parameter set and the preset AI computing power judgment algorithm. The matching module 603 is used to match the required computing power value with the critical computing power list corresponding to each current blockchain distributed computing pool according to the preset matching rule, so as to determine the allocation target computing pool according to the matching result. Among them, the preset matching rule is used to match one or more target pools to be allocated that at least reach the required computing power value. A blockchain distributed computing pool includes multiple GPU computing units. The allocation module 604 is used to allocate the AI processing task corresponding to the AI processing request information to the idle GPU computing unit in the allocation target computing pool, so as to send the AI processing result after the idle GPU computing unit executes the AI processing task to the blockchain user node.
[0121] In the embodiment of the present application, the second determining module 602 is specifically used for:
[0122] The model scale parameter, the number of required parameters, the precision operation complexity coefficient and the reference model coefficient group are input into the preset AI computing power judgment algorithm to calculate the corresponding operation parameter complexity parameter. The reference model coefficient group includes at least a first coefficient representing the scale of the reference model and a second coefficient representing the number of reference model parameters. The required computing power value is determined according to the product value of the preset server CPU main frequency parameter, the preset maximum GPU operation parameter and the operation parameter complexity parameter. Among them, the preset maximum GPU operation parameter is the maximum value of the GPU operation parameter corresponding to each GPU computing unit in the current blockchain distributed computing pool. The GPU operation parameter is the number of floating-point operations per second.
[0123] In the embodiment of the present application, the second determining module 602 is further configured to:
[0124] By using the preset AI computing power judgment algorithm, the division result of the model scale parameter and the preset maximum reference model scale value is determined as the first intermediate value. The division result of the required number of parameters and the preset maximum number of parameters is determined as the second intermediate value. The product value of the first intermediate value and the first coefficient is calculated as the third intermediate value, and the product value of the second intermediate value and the second coefficient is calculated as the fourth intermediate value. The sum of the third intermediate value and the fourth intermediate value is calculated, and the product value of the sum and the precision operation complexity coefficient is determined as the operation parameter complexity parameter.
[0125] In the embodiment of the present application, the matching module 603 is specifically used for:
[0126] Determine the critical computing power value in the critical computing power list that is greater than or equal to the required computing power value. In order from small to large, add the identifiers of each blockchain distributed computing pool corresponding to each critical computing power value to the sequence of target pools to be allocated. According to the sequence of target pools to be allocated, poll the idle status of each GPU computing unit in each target pool to be allocated in turn, and determine the allocation target computing pool according to the polling results.
[0127] The matching module 603 is further specifically used for:
[0128] Determine whether there are idle GPU computing units in the currently polled target pool. If so, determine that the currently polled target pool is the allocated target computing pool. Otherwise, use the next target pool to be allocated of the currently polled target pool to be allocated in the target pool sequence as the currently polled target pool to be allocated until the allocated target computing pool is determined.
[0129] The system also includes: a blockchain user node exists in a blockchain user cluster composed of multiple blockchain user nodes. One blockchain user cluster corresponds to one user server. The blockchain server is connected to each user server and each blockchain distributed computing pool through a CXL Switch, and is used to receive AI processing request information and distribute AI processing request information.
[0130] In an embodiment of the present application, the system further includes: the performance parameters of each GPU computing unit in the same blockchain distributed computing pool are the same. The performance parameters of the GPU computing units between the blockchain distributed computing pools are different. The GPU computing units in the same blockchain distributed computing pool are connected via an AXI bus.
[0131] In an embodiment of the present application, the system further includes: obtaining user uploaded data from a blockchain user node. Determine whether there is an AI computing requirement based on the user uploaded data and a preset AI computing screening model. The preset AI computing screening model is a model trained based on a number of user data samples labeled with AI computing requirement tags. If there is, the user uploaded data is used as AI processing request information. Figure 7 A schematic diagram of the structure of a blockchain-based AI computing and control device provided in an embodiment of the present application, such as Figure 7 As shown, the device includes:
[0132] At least one processor. And a memory in communication with the at least one processor. The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute:
[0133] Based on the AI processing request information from the blockchain user node, the corresponding AI processing parameter set is determined. Among them, the AI processing parameter set includes at least a model scale parameter, a required number of parameters, a precision calculation complexity coefficient, and a reference model coefficient group. Based on the AI processing parameter set and the preset AI computing power judgment algorithm, the required computing power value corresponding to the AI processing request information is determined. According to the preset matching rules, the required computing power value is matched with the critical computing power list corresponding to each current blockchain distributed computing pool to determine the allocation target computing pool according to the matching result. Among them, the preset matching rules are used to match one or more target pools to be allocated that at least reach the required computing power value. A blockchain distributed computing pool includes multiple GPU computing units. The AI processing task corresponding to the AI processing request information is allocated to the idle GPU computing unit in the allocation target computing pool, so that the AI processing result after the idle GPU computing unit executes the AI processing task is sent to the blockchain user node.
[0134] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system and device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.
[0135] The system and equipment provided in the embodiments of the present application correspond one-to-one to the method. Therefore, the system and equipment also have similar beneficial technical effects as the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the system and equipment will not be repeated here.
[0136] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.
[0137] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.
Claims
1. A blockchain-based AI computing management and control method, characterized in that: The method is applied to a private chain; the method comprises: Based on the AI processing request information from the blockchain user node, determine the corresponding AI processing parameter set; wherein the AI processing parameter set at least includes a model scale parameter, a required number of parameters, a precision calculation complexity coefficient, and a reference model coefficient group; Determine the required computing power value corresponding to the AI processing request information based on the AI processing parameter set and the preset AI computing power judgment algorithm; According to the preset matching rules, the required computing power value is matched with the critical computing power list corresponding to each current blockchain distributed computing pool, so as to determine the allocation target computing pool according to the matching result; wherein the preset matching rule is used to match one or more target pools to be allocated that at least reach the required computing power value; one of the blockchain distributed computing pools includes multiple GPU computing units; Allocate the AI processing task corresponding to the AI processing request information to the idle GPU computing unit in the allocation target computing pool, so as to send the AI processing result after the idle GPU computing unit executes the AI processing task to the blockchain user node; Wherein, based on the AI processing parameter set and the preset AI computing power judgment algorithm, determining the required computing power value corresponding to the AI processing request information specifically includes: Input the model scale parameter, the required number of parameters, the precision operation complexity coefficient and the reference model coefficient group into the preset AI computing power judgment algorithm to calculate the corresponding operation parameter complexity parameter; the reference model coefficient group includes at least a first coefficient representing the scale of the reference model and a second coefficient representing the number of reference model parameters; Determine the required computing power value according to the product value of the preset server CPU main frequency parameter, the preset maximum GPU computing parameter and the computing parameter complexity parameter; wherein the preset maximum GPU computing parameter is the maximum value of the GPU computing parameter corresponding to each of the GPU computing units in each of the current blockchain distributed computing pools; The model scale parameter, the required number of parameters, the precision calculation complexity coefficient and the reference model coefficient group are input into the preset AI computing power judgment algorithm to calculate the corresponding calculation parameter complexity parameter, specifically including: The division result of the model scale parameter and the preset maximum reference model scale value is determined by the preset AI computing power judgment algorithm. , is the first intermediate value; Determine the result of dividing the required number of parameters by the preset maximum number of parameters , is the second intermediate value; Calculate the product of the first intermediate value and the first coefficient as a third intermediate value , calculate the product of the second intermediate value and the second coefficient as the fourth intermediate value ; Calculate the sum of the third intermediate value and the fourth intermediate value, and determine the product of the sum and the precision operation complexity coefficient , is the operation parameter complexity parameter; The formula of the preset AI computing power judgment algorithm is as follows: ; ; ; in, is the required computing power value; The CPU main frequency parameter of the preset server; The preset maximum GPU computing parameter; is the operation parameter complexity parameter; is the model scale parameter; is the preset maximum reference model scale value; is the first coefficient; is the number of required parameters; The maximum number of preset parameters; is the second coefficient; is the precision operation complexity coefficient; wherein, The calculation accuracy of the AI operation is confirmed; the first coefficient and the second coefficient are set according to different reference models and different reference model parameter quantities; Among them, according to the preset matching rules, the required computing power value is matched with the critical computing power list corresponding to each current blockchain distributed computing pool, so as to determine the allocation target computing pool according to the matching result, specifically including: Determine a critical computing power value in the critical computing power list that is greater than or equal to the required computing power value; In order from small to large, the identifiers of the blockchain distributed computing pools corresponding to the critical computing power values are added to the sequence of target pools to be allocated; According to the sequence of the target pools to be allocated, the idle state of each of the GPU computing units in each of the target pools to be allocated is polled in sequence, so as to determine the allocated target computing pool according to the polling result; The idle state of each GPU computing unit in each target pool to be allocated is polled in sequence to determine the allocated target computing pool according to the polling result, specifically including: Determine whether there is an idle GPU computing unit in each of the GPU computing units in the target pool to be allocated currently being polled; If so, determining that the currently polled target pool to be allocated is the allocated target computing pool; Otherwise, the next target pool to be allocated of the currently polled target pool to be allocated in the sequence of target pools to be allocated is used as the currently polled target pool to be allocated, until the allocated target computing pool is determined; Among them, the performance parameters of the GPU computing units in the same blockchain distributed computing pool are the same; the performance parameters of the GPU computing units between the blockchain distributed computing pools are different; the GPU computing units in the same blockchain distributed computing pool are connected through the advanced expansion interface AXI bus; Wherein, before determining the corresponding AI processing parameter set based on the AI processing request information from the blockchain user node, the method further includes: Obtaining user uploaded data from the blockchain user node; Determine whether there is an AI computing demand based on the user uploaded data and the preset AI computing screening model; wherein the preset AI computing screening model is a model trained based on a number of user data samples annotated with AI computing demand labels; If so, the user uploaded data is used as the AI processing request information; Wherein, the blockchain user node exists in a blockchain user cluster composed of multiple blockchain user nodes; one blockchain user cluster corresponds to one user server; The blockchain server connects each of the user servers and each of the blockchain distributed computing pools through a computing fast link switch CXL Switch, so as to receive the AI processing request information and distribute the AI processing request information through the CXL Switch.
2. A blockchain-based AI computing management and control system, characterized in that: The system comprises: A first determination module is used to determine a corresponding AI processing parameter set based on the AI processing request information from the blockchain user node; wherein the AI processing parameter set at least includes a model scale parameter, a required parameter quantity, a precision calculation complexity coefficient, and a reference model coefficient group; A second determination module, configured to determine a required computing power value corresponding to the AI processing request information based on the AI processing parameter set and a preset AI computing power judgment algorithm; A matching module is used to match the required computing power value with the critical computing power list corresponding to each current blockchain distributed computing pool according to a preset matching rule, so as to determine the target computing pool for allocation according to the matching result; wherein the preset matching rule is used to match one or more target pools to be allocated that at least reach the required computing power value; one of the blockchain distributed computing pools includes multiple GPU computing units; An allocation module, configured to allocate the AI processing task corresponding to the AI processing request information to an idle GPU computing unit in the allocation target computing pool, so as to send the AI processing result after the idle GPU computing unit executes the AI processing task to the blockchain user node; Wherein, the second determining module is specifically used for: Input the model scale parameter, the required number of parameters, the precision operation complexity coefficient and the reference model coefficient group into the preset AI computing power judgment algorithm to calculate the corresponding operation parameter complexity parameter; the reference model coefficient group includes at least a first coefficient representing the scale of the reference model and a second coefficient representing the number of reference model parameters; Determine the required computing power value according to the product value of the preset server CPU main frequency parameter, the preset maximum GPU computing parameter and the computing parameter complexity parameter; wherein the preset maximum GPU computing parameter is the maximum value of the GPU computing parameter corresponding to each of the GPU computing units in each of the current blockchain distributed computing pools; Wherein, the second determining module is specifically used for: The division result of the model scale parameter and the preset maximum reference model scale value is determined by the preset AI computing power judgment algorithm. , is the first intermediate value; Determine the result of dividing the required number of parameters by the preset maximum number of parameters , is the second intermediate value; Calculate the product of the first intermediate value and the first coefficient as a third intermediate value , calculate the product of the second intermediate value and the second coefficient as the fourth intermediate value ; Calculate the sum of the third intermediate value and the fourth intermediate value, and determine the product of the sum and the precision operation complexity coefficient , is the operation parameter complexity parameter; The formula of the preset AI computing power judgment algorithm is as follows: ; ; ; in, is the required computing power value; The CPU main frequency parameter of the preset server; The preset maximum GPU computing parameter; is the operation parameter complexity parameter; is the model scale parameter; is the preset maximum reference model scale value; is the first coefficient; is the number of required parameters; The maximum number of preset parameters; is the second coefficient; is the precision operation complexity coefficient; wherein, The calculation accuracy of the AI operation is confirmed; the first coefficient and the second coefficient are set according to different reference models and different reference model parameter quantities; Wherein, the matching module is specifically used for: Determine a critical computing power value in the critical computing power list that is greater than or equal to the required computing power value; In order from small to large, the identifiers of the blockchain distributed computing pools corresponding to the critical computing power values are added to the sequence of target pools to be allocated; According to the sequence of the target pools to be allocated, the idle state of each of the GPU computing units in each of the target pools to be allocated is polled in sequence, so as to determine the allocated target computing pool according to the polling result; Wherein, the matching module is further specifically used for: Determine whether there is an idle GPU computing unit in each of the GPU computing units in the target pool to be allocated currently being polled; If so, determining that the currently polled target pool to be allocated is the allocated target computing pool; Otherwise, the next target pool to be allocated of the currently polled target pool to be allocated in the sequence of target pools to be allocated is used as the currently polled target pool to be allocated, until the allocated target computing pool is determined; The system further includes: the performance parameters of the GPU computing units in the same blockchain distributed computing pool are the same; the performance parameters of the GPU computing units between the blockchain distributed computing pools are different; the GPU computing units in the same blockchain distributed computing pool are connected via an advanced expansion interface AXI bus; Among other things, the system can also: Obtaining user uploaded data from the blockchain user node; Determine whether there is an AI computing demand based on the user uploaded data and the preset AI computing screening model; wherein the preset AI computing screening model is a model trained based on a number of user data samples annotated with AI computing demand labels; If so, the user uploaded data is used as the AI processing request information; The system further includes: the blockchain user node exists in a blockchain user cluster composed of multiple blockchain user nodes; one blockchain user cluster corresponds to one user server; The blockchain server connects each of the user servers and each of the blockchain distributed computing pools through a computing fast link switch CXL Switch, so as to receive the AI processing request information and distribute the AI processing request information through the CXL Switch.
3. A blockchain-based AI computing and control device, characterized in that: The device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the blockchain-based AI computing management method as described in claim 1 above.
