CN-121996406-A - Power calculation scheduling system for supporting Inlet Weida and flint original cards simultaneously

CN121996406ACN 121996406 ACN121996406 ACN 121996406ACN-121996406-A

Abstract

The invention discloses a power calculation scheduling system for supporting an Inlet and flint original card simultaneously, which comprises a hardware layer, a driving adaptation layer, a power calculation scheduling layer and an application interface layer, wherein bidirectional data interaction is carried out between the layers through a data bus, the hardware layer is used for carrying out hardware state monitoring and data transmission, the driving adaptation layer is used for converting power calculation feedback data into a standard data format and transmitting the data to the power calculation scheduling layer, the power calculation scheduling layer is used for distributing power calculation tasks, and the application interface layer is used for providing an API interface for adapting to a third-party application program. The invention has the advantages that the calculation force request instruction can be respectively converted into the instruction format conforming to the CUDA architecture and the instruction format conforming to the Bolt architecture through the driving adaptation layer, and the calculation force request instruction is sent to the Injeida display card module or the flint original display card module through the corresponding driving sub-module, so that the simultaneous support of two different types of the Injeida and the flint original display cards is realized, an independent dispatching system is not required to be built, and the hardware and software cost is reduced.

Inventors

YANG JIE

Assignees

甘肃燧弘绿色算力有限公司

Dates

Publication Date: 20260508
Application Date: 20251223

Claims (5)

1. The power calculation scheduling system supporting both the Inlet and the flint original cards is characterized by comprising a hardware layer, a drive adaptation layer, a power calculation scheduling layer and an application interface layer, wherein the layers are subjected to bidirectional data interaction through a data bus; the hardware layer is used for carrying out hardware state monitoring and data transmission; the driving adaptation layer is used for converting the calculation force feedback data into a standard data format and transmitting the calculation force feedback data to the calculation force scheduling layer; the power calculation scheduling layer is used for distributing power calculation tasks; The application interface layer provides an API interface for adapting to third party applications.
2. The system for simultaneously supporting both an Injean and a flint original card according to claim 1, wherein the hardware layer comprises an Injean display card module, a flint original display card module, a central control module and a power management module; The Injean display card module is used for providing computational power based on CUDA ecology; the flint original display card module is used for providing calculation power based on the belt ecology; the central control module adopts a processor of ARM architecture, is electrically connected with the Injean display card module and the flint original display card module through a PCIe switch and is used for monitoring the hardware states of the two types of display card modules and transmitting data; The power management module is used for supplying power to the Injean display card module, the flint original display card module and the central control module.
3. The system for simultaneously supporting both the Injew's and the flint's raw cards according to claim 2, wherein the driving adaptation layer is disposed on the central control module and comprises an Injew's driving sub-module, a flint's raw driving sub-module and a driving conversion sub-module; The Injean driving sub-module is used for enabling the central control module to be matched with the communication of the Injean display card module; The flint original driving sub-module is used for adapting the communication between the central control module and the flint original display card module; The driving conversion submodule is used for respectively converting the calculation force request instruction of the application task into an instruction format conforming to the CUDA architecture and an instruction format conforming to the Bolt architecture, transmitting the instruction format to the Inlet Weida video card module or the flint original video card module through the corresponding driving submodule, uniformly converting calculation force feedback data into a standard data format and transmitting the standard data format to the calculation force scheduling layer.
4. The system for simultaneously supporting the power scheduling of the Inlet and the flint cards according to claim 3, wherein the power scheduling layer comprises a task analysis module, a resource monitoring module, a scheduling decision module and a task allocation module; the task analysis module is used for receiving the calculation force task request of the application interface layer and analyzing calculation force demand parameters of the task to generate a task feature tag; the resource monitoring module acquires hardware state parameters of the Inlet video card module and the flint original video card module in real time through the driving adaptation layer, and establishes a computing power resource state library; the scheduling decision module adopts a dynamic weight scheduling algorithm to make decisions based on task feature labels and an algorithm resource state library; and the task allocation module allocates the calculation task to the corresponding display card module through the driving adaptation layer according to the decision result of the scheduling decision module, and tracks the task execution progress in real time.
5. The system for simultaneously supporting both Injeida and flint original cards according to claim 4, wherein the application interface layer comprises a calculation force request interface, a task state query interface and a calculation force statistics interface; the computing force request interface supports a third party application program to send a computing force task request through HTTP or RPC protocol and receive task submission feedback; the task state query interface is used for querying the execution progress of the submitted task, the type of the currently allocated display card module and the task completion time estimation in real time; The computing power statistics interface is used for generating computing power use reports of the Injeida display card module and the flint original display card module.

Description

Power calculation scheduling system for supporting Inlet Weida and flint original cards simultaneously Technical Field The invention relates to the technical field of computer power dispatching, in particular to a power dispatching system for supporting an Injeida and an original flint card simultaneously. Background With the rapid development of the fields of artificial intelligence, big data processing, high-performance computing and the like, the demand for computing power on computer hardware is increasing. The main high-performance display cards in the market at present mainly comprise an Inlet-Weida series display card and a flint original technology series display card, and the two types of display cards are widely applied in different application scenes by means of respective hardware architecture and software ecology. However, in the prior art, the power dispatch system is usually designed only for a single brand or architecture of graphics card. For example, the scheduling system for the Iwei reach display card is mostly developed based on the CUDA architecture, while the scheduling system for the flint original display card depends on the Bolt system architecture which is independently developed, and the two systems lack compatibility and cannot realize unified scheduling of the two display cards, so that in a computing cluster where the Iwei reach and the flint original cards are deployed at the same time, independent scheduling systems are required to be built respectively, thus not only increasing hardware purchase cost and software maintenance cost, but also having the problems of labor resource waste, low scheduling efficiency and the like. In addition, because the hardware parameters (such as the memory capacity, the number of computing cores and the power consumption limit) and the software interfaces of the Injeida and the flint original cards are different, the existing dispatching system cannot dynamically allocate computing power according to the characteristics of the two types of display cards, so that the differentiated requirements of different tasks on computing power are difficult to meet, and the overall operation efficiency of the computing cluster is further reduced. Disclosure of Invention The invention aims to overcome the defects of the prior art and provide a computing power dispatching system for supporting both an Injeida and an original flint card. The invention aims at realizing the technical scheme that the power calculation scheduling system for supporting the Injeam and the flint original card simultaneously comprises a hardware layer, a driving adaptation layer, a power calculation scheduling layer and an application interface layer, wherein the layers are subjected to bidirectional data interaction through a data bus; the hardware layer is used for carrying out hardware state monitoring and data transmission; the driving adaptation layer is used for converting the calculation force feedback data into a standard data format and transmitting the calculation force feedback data to the calculation force scheduling layer; the power calculation scheduling layer is used for distributing power calculation tasks; the application interface layer provides an API interface to adapt to third party applications. Preferably, the hardware layer comprises an Injean display card module, a flint original display card module, a central control module and a power management module; The Inlet-Wei-Da display card module is used for providing computational power based on CUDA ecology; The flint original display card module is used for providing calculation power based on the belt ecology; The central control module adopts a processor of ARM architecture, is electrically connected with the Injean display card module and the flint original display card module through a PCIe switch, and is used for monitoring the hardware states of the two types of display card modules and transmitting data; the power management module is used for supplying power to the Injean display card module, the flint original display card module and the central control module. Preferably, the driving adaptation layer is deployed on the central control module and specifically comprises an Injean driving sub-module, a flint original driving sub-module and a driving conversion sub-module; the Injean driving sub-module is used for adapting the communication between the central control module and the Injean display card module; the flint original driving sub-module is used for adapting the communication between the central control module and the flint original display card module; The driving conversion submodule is used for respectively converting the calculation force request instruction of the application task into an instruction format conforming to the CUDA architecture and an instruction format conforming to the Bolt architecture, transmitting the instruction format to the Injeida display card module or the flint or