Data processing device, operation accelerator, and data processing method
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ATEN INTERNATIONAL CO LTD
- Filing Date
- 2025-07-24
- Publication Date
- 2026-08-07
AI Technical Summary
然而，此一作法将使得AI的速度受到网络传输速度所限制
Smart Images
Figure CN122528969A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to data processing and computing technologies, and in particular to a data computing device, a computing accelerator, and a data computing method. Background Technology
[0002] With the rapid development of technology, Artificial Intelligence (AI) has been widely applied in various fields, including anomaly detection, machine vision, robot control, and speech recognition. Currently, AI is mostly offered as a network service. In other words, users must connect to the internet and perform AI calculations through cloud servers. However, this approach limits the speed of AI to network transmission speed. Furthermore, users find it difficult to personalize and adjust the AI's computational models, thus leaving much room for improvement. Summary of the Invention
[0003] This disclosure relates to a data processing device, including a processing accelerator. The processing accelerator includes a processing unit and a storage device simulation unit. The processing unit performs calculations based on processing data to generate calculation results. The storage device simulation unit includes a processor and a memory, wherein the processor is coupled to the processing unit and the memory. When the processor confirms that writing a first file to the memory is complete, the processor uses the first file as processing data to obtain the calculation results generated by the processing unit. The processor stores the calculation results as a second file in the memory.
[0004] In one embodiment, when the processor determines that the format of the first file conforms to the operation format, the processor uses the first file as the operation data of the operation unit.
[0005] In one embodiment, the computing unit includes a machine learning model, which includes multiple parameters. When the processor determines that the memory has been written to a model file, the processor updates at least one of the parameters in the machine learning model according to the model file.
[0006] In one embodiment, the processor monitors the memory, and when the memory is written to an end instruction, the processor executes a machine learning model based on a first file.
[0007] In one embodiment, the data processing device further includes a host. The host is coupled to the computing accelerator and identifies the computing accelerator as an external storage device based on the storage device emulation unit.
[0008] In one embodiment, the host is used to provide model files to the computing unit as a machine learning model in the computing unit, or to update the machine learning model in the computing unit.
[0009] This disclosure also relates to a data processing method executed in a computing accelerator. The data processing method includes: determining, by means of a processor in the computing accelerator, whether a memory coupled to the processor has been written to a first file; when the first file has been written to the memory, using the first file as processing data; performing operations on the processing data by a processing unit in the computing accelerator to generate a processing result; and storing the processing result in the memory as a second file.
[0010] In one embodiment, the computing unit includes a machine learning model having multiple parameters, and the data processing method further includes: when it is determined that the memory is written to a model file, updating at least one of the parameters in the machine learning model according to the model file.
[0011] In one embodiment, the method of using the first file as computational data further includes: monitoring the memory; and executing a machine learning model based on the first file when the memory is written with an end instruction.
[0012] This disclosure also relates to a computing accelerator, comprising a computing unit and a storage device simulation unit. The computing unit performs calculations based on computational data to generate computational results. The storage device simulation unit includes a processor and a memory. The processor is coupled to the computing unit and the memory, and is used to configure the memory as an external memory. The processor is also used to determine the type of data written to the memory. When the data written to the memory is a model file, the processor updates the computing unit according to the model file. When the data written to the memory is a first file, the processor uses the first file as computational data to obtain the computational results generated by the computing unit. The types of computational data and model files are different.
[0013] Accordingly, by setting the machine learning model in the storage device simulation unit and configuring the storage device simulation unit and the computing unit as a computing accelerator, the computing accelerator can be used like an external storage device without being limited by the device's operating system. Attached Figure Description
[0014] Figure 1 This is a schematic diagram of a data processing device according to some embodiments of the present disclosure.
[0015] Figure 2 This is a flowchart of a data processing method according to some embodiments of the present disclosure.
[0016] The reference numerals in the attached figures are explained as follows:
[0017] 100: Data processing device
[0018] 110: Storage device emulation unit
[0019] 111: Processor
[0020] 112: Memory
[0021] 120: Arithmetic Unit
[0022] AID: Computational Accelerator
[0023] D1: First File
[0024] D2: Second document
[0025] M1: Machine Learning Model
[0026] HD: Host
[0027] S201-S208: Steps Detailed Implementation
[0028] The following describes several embodiments of the present invention with reference to the accompanying drawings. For clarity, many practical details will be described in conjunction with the following description. However, it should be understood that these practical details are not intended to limit the invention. That is, in some embodiments of the present invention, these practical details are not essential. Furthermore, for the sake of simplicity, some conventional structures and elements will be shown in the drawings in a simple schematic manner.
[0029] In this document, when an element is referred to as a “connection” or “coupled,” it may mean an “electrical connection” or “electrical coupling.” “Connection” or “coupled” can also be used to indicate the operation or interaction between two or more elements. Furthermore, although terms such as “first,” “second,” etc., are used herein to describe different elements, these terms are merely used to distinguish elements or operations described using the same technical terminology. Unless the context clearly indicates otherwise, these terms do not specifically refer to or imply any order or sequence, nor are they intended to limit the invention.
[0030] An AI computing accelerator is a computing processing circuit designed for a specific computing model. Terminal devices equipped with a computing accelerator can perform calculations directly through the accelerator without needing to access AI services from a cloud server.
[0031] For example, when the processor of a terminal device (such as a computer or smartphone) receives computational data, it transmits the data to a computation accelerator to obtain the computational result using the hardware and computational model within the accelerator. However, since computation accelerators must work in conjunction with the processor, they need to be designed separately for different operating systems. In other words, different operating systems require different computation accelerators. For users, this not only limits their choices but also may lead to problems installing / configuring the computation accelerator (such as needing to correctly install the corresponding driver), making the application inconvenient.
[0032] This disclosure describes the design of a computing accelerator as an external device, enabling its data processing methods to be independent of the terminal device's operating system, thereby improving ease of use. For ease of explanation, the computing accelerator and its associated host device will be collectively referred to as a "data processing device" in the following paragraphs.
[0033] Figure 1 The diagram illustrates a data processing device 100 according to a partial embodiment of this disclosure. The data processing device 100 includes a computing accelerator AID and a host HD. The host HD can be a terminal device or a server, or an electronic device with a built-in embedded system (e.g., an automotive control system). The computing accelerator AID can be connected to the host HD via a connector (e.g., a Universal Serial Bus). In other words, the computing accelerator AID can be connected to the host HD in a plug-and-play manner to provide AI computing services.
[0034] The computing accelerator AID includes a storage device simulation unit 110 and a computing unit 120. The storage device simulation unit 110 and the computing unit 120 may each be a processing circuit or a microprocessor, and may include hardware and / or firmware. The computing unit 120 includes a machine learning model M1, and the computing unit 120 is used to perform calculations based on received computing data and according to the machine learning model M1 to generate computing results.
[0035] A machine learning model M1 is used to provide specific computational functions and can be implemented by a neural network. In one embodiment, the neural network of the machine learning model M1 includes one or more layers, including an input layer, a hidden layer, and an output layer. Each layer has multiple neurons, and these neurons have corresponding parameters (i.e., weight values generated through training) to determine the strength and direction of signal transmission in the neural network. The machine learning model M1 can process received computational data, such as: analyzing the semantics of text to provide a response, organizing data, identifying content in images, generating data according to input instructions (e.g., generating images or music, identifying data errors, generating more similar content based on input data), etc. Since those skilled in the art will understand the meaning and composition of machine learning models, further details are omitted here.
[0036] The storage device emulation unit 110 is coupled to the arithmetic unit 120 and includes a processor 111 and a memory 112. The processor 111 is coupled to the arithmetic unit 120 and the memory 112 and is used for communication with the host HD. The memory 112 may be a flash memory used to store data and files. In one embodiment, the storage device emulation unit 110 configures the memory 112 as an external memory, and the processor 111 is used to determine the type of data written to the memory 112, for example, determining the filename extension of the written data, or analyzing the content of the write command through voice recognition or text recognition. The memory 112 can be any type of storage medium, and this disclosure does not limit its type.
[0037] Specifically, when the computing accelerator AID is connected to the host HD, the host HD identifies the computing accelerator AID as an external storage device, such as a USB flash drive / portable hard drive, based on the storage device emulation unit 110. For example, the host HD's connection circuitry sends an identification request signal to the computing accelerator AID, and the computing accelerator AID sends an identification response signal back to the host HD accordingly. The identification response signal may include a device descriptor, such as the manufacturer's name code, product identification code, type code, etc., so that the host HD can identify the computing accelerator AID as an external storage device based on the device descriptor.
[0038] Continuing from the above, in one embodiment, after the host HD recognizes the computing accelerator AID as an external storage device, the host HD allocates a driver corresponding to "external storage device" in the operating system to read the memory 112. At this time, the host HD uses the memory 112 as an external data storage space. For example, the host HD will display a data management program on the display screen and show a list of files stored in the memory 112. In one embodiment, the host HD can also use a general-purpose driver to communicate with the memory 112 and perform reading. In other words, after the computing accelerator AID is connected to the host HD, no additional driver installation is required; bidirectional communication and transmission can be performed simply through the connection interface. The connection interface can be the USB transmission protocol, but this disclosure is not limited to this.
[0039] The processor 111 receives operation instructions from the host HD and executes the corresponding actions. Since the computing accelerator AID is recognized as an external storage device, the content of the "operation instructions" is the same as that of the flash drive, including writing data, reading data, and clearing data. The processor 111 is also used to continuously monitor the memory 112, for example, to confirm whether the data writing is complete.
[0040] If the operation instruction is a write instruction, the processor 111 will write data into the memory 112 according to the write instruction to form a first file D1. When the first file D1 is written, the processor 111 will use the first file D1 as the calculation data and input the calculation data into the machine learning model M1 to generate the calculation result corresponding to the calculation data through the machine learning model M1. Then, the processor 111 will store the calculation result as a second file in the memory 112.
[0041] Accordingly, users of the host HD will be able to use the computing accelerator in a manner similar to a "USB flash drive". Furthermore, since the computing accelerator AID is recognized as an external storage device, as long as the host HD can read the external storage device, it can correctly establish communication with the processor 111 without being limited by the operating system.
[0042] In some embodiments, the machine learning model M1 may be provided by the host HD. In other words, the host HD may provide a model file to the computing unit 120 as the machine learning model M1 in the computing unit 120, or update the existing machine learning model M1 in the computing unit 120 according to the model file.
[0043] One thing that deserves special mention here is... Figure 1The "Host HD" shown is for illustrative purposes only; the compute accelerator AID can connect to various devices. For example, the compute accelerator AID can connect to electronic devices with embedded systems to enhance or expand the functionality of those systems through machine learning models. Alternatively, the compute accelerator AID can connect to a computer (e.g., a maintenance personnel's computer) to set or update internal machine learning models. The aforementioned "electronic devices with embedded systems" and "maintenance personnel's computer" can both be... Figure 1 The host HD shown, and the computing accelerator AID application is not based on Figure 1 The diagram shown is for reference only.
[0044] For ease of understanding, several application scenarios of the computing accelerator AID are described below. In one embodiment, the machine learning model M1 can be a neural network for conversational AI. Users can record questions as text (e.g., file extension .txt) or voice (e.g., file extension .mp3) and store the file in memory 112, i.e., the first file D1. The processor 111 inputs the first file D1 into the machine learning model M1 and outputs the results generated by the machine learning model M1 in the same type / format as a second file D2 (e.g., file extension .txt or .mp3). Users can open the second file D2 to see the response content of the machine learning model M1.
[0045] In other embodiments, the machine learning model M1 can be an analytical AI. The user can store the data to be analyzed in memory 112, i.e., a first file D1. The first file D1 can be a video (e.g., with the extension .mp4) or an image (e.g., with the extension .jpg). The processor 111 inputs the first file D1 into the machine learning model M1, causing the machine learning model M1 to analyze the first file D1 (e.g., removing noise, adding effects, etc.). Then, the processor 111 outputs the calculation results in the same type / format as a second file D2.
[0046] In some embodiments, the processor 111 determines the type of data written to the memory 112 to perform different processing. Specifically, the processor 111 can determine the file extension of the first file D1 to confirm whether the first file is computational data. As in the aforementioned embodiment, if the data type received by the machine learning model M1 is "text", the processor 111 will determine whether the file extension of the first file D1 is ".txt" or other text format. Only when the type / file extension matches will the processor 111 input the first file D1 as computational data into the machine learning model M1.
[0047] In other embodiments, the processor 111 may also determine whether the type of the data being written belongs to the type of a model file. A "model file" is all or part of the configuration data (e.g., firmware, parameters, etc.) of the machine learning model M1, such as an update file for the machine learning model M1, with a file extension of ".onnx" or ".pth". Therefore, the type / file extension of the "model file" will be different from the type / file extension of the "operational data" (the input data received by the machine learning model M1). If the processor 111 determines that the type of the first file D1 being written is a model file, then the processor 111 will update the machine learning model M1 in the operation unit 120 according to the first file D1 / model file.
[0048] Figure 2 The diagram shown is a flowchart of a data processing method according to some embodiments of this disclosure. Figure 1 and Figure 2 For example, the operation of the data processing device 100 will be explained. First, the computing accelerator AID will be connected to the host HD so that the host HD recognizes the computing accelerator AID as an external storage device. In addition, the processor 111 can monitor the type of received signal or monitor the state of the memory 112.
[0049] In step S201, the processor 111 receives an operation instruction from the host HD and identifies the content of the operation instruction. Steps S202 to S208 involve judgment and subsequent processing based on different contents of the operation instruction.
[0050] In step S202, the processor 111 determines whether the operation instruction is a write instruction. In other words, it determines whether the operation instruction requires writing a file to memory 112. If the operation instruction is a write instruction, then in step S203, the processor 111 writes the first file D1 to memory 112 according to the operation instruction.
[0051] In step S204, the processor 111 determines whether the operation instruction contains an end instruction. The "end instruction" can be a specific character, such as "End of File (EOF)". When the processor 111 determines that the operation instruction contains an end instruction, it can confirm that the data (first file D1) that the host HD is currently writing has been completed.
[0052] In step S205, after the first file D1 is written to the memory 112, the processor 111 uses the first file D1 as computation data and inputs the computation data into the machine learning model M1 to execute the computation function of the machine learning model M1 based on the first file D1 (computation data). In one embodiment, the processor 111 monitors the memory 112 to confirm whether an end instruction has been written into the memory 112.
[0053] In step S206, the machine learning model M1 performs calculations based on the computational data to generate a result. After receiving the result, the processor 111 stores it as a second file D2 in the memory 112. Therefore, users of the host HD will see that the memory 112 now contains a second file D2, which represents the result of the machine learning model M1's computation.
[0054] As mentioned above, in some embodiments, before executing step S205, the processor 111 may first determine whether the format of the first file D1 conforms to the operation format of the operation data (e.g., .txt, .jpg, etc.). If the format of the first file D1 conforms to the operation format, the processor 111 will execute step S205 and use the first file D1 as the operation data.
[0055] Continuing from the above, in other embodiments, the processor 111 may further determine whether the first file D1 conforms to the model format of the model file, wherein the model format and the operation format are different. If the format of the first file D1 conforms to the model format, it means that the first file D1 belongs to the model file. At this time, the processor 111 does not execute step S205, but instead installs / updates the machine learning model M1 according to the model file, or updates at least one of the multiple parameters in the machine learning model M1.
[0056] The aforementioned action of "determining the type / format of the first file D1" can be performed after step S204, that is, when the end instruction is detected, the type of the first file D1 is determined. However, this disclosure is not limited thereto. In other embodiments, the processor 111 may also determine the type of the first file D1 in step S201 to confirm whether the first file D1 belongs to a type of computational data or a model file.
[0057] In step S207, processor 111 determines whether the operation instruction is a read instruction. If the operation instruction is a read instruction, it means that host HD wants to read the calculation result of machine learning model M1. In step S208, processor 111 uses the second file D2 as output data and transmits the output data (second file D2) to host HD to present the calculation result.
[0058] In some embodiments, the processor 111 can identify the second file D2 in response to a read instruction (e.g., the user selects the second file D2 and performs a copy or open operation). However, this disclosure is not limited to this. In other embodiments, after the processor 111 uses the first file D1 as processing data (i.e., step S205), it can retain or delete the first file D1, and the read instruction will output the second file D2 in the memory 112 to the host HD. However, this disclosure is not limited to this. Accordingly, even if the read instruction does not actively specify / select the second file D2, the processor 111 can actively output all data (the second file D2) in the memory 112 to the host HD.
[0059] This disclosure utilizes the "write to file, read from file" method as the input and output method for machine learning models. In terms of data operation, it is only a simple data copying and moving. Therefore, the computing accelerator AID does not require setting up specific drivers or specific program frameworks, and can be used without being restricted by the operating system, thus improving the convenience and versatility of use.
[0060] The components, method steps, or technical features in the foregoing embodiments can be combined with each other, and are not limited to the order of textual description or graphical presentation in this disclosure.
[0061] Although the present disclosure has been described above with reference to embodiments, it is not intended to limit the present disclosure. Any person skilled in the art may make various modifications and refinements without departing from the spirit and scope of the present disclosure. Therefore, the scope of protection of the present disclosure shall be determined by the appended claims.
Claims
1. A data processing device, characterized in that, Includes a computing accelerator, the computing accelerator comprising: The arithmetic unit is used to perform calculations based on the data to produce the results; and A storage device emulation unit includes a processor and a memory, wherein the processor is coupled to the arithmetic unit and the memory; When the processor confirms that the memory has been written to the first file, the processor uses the first file as the operation data to obtain the operation result generated by the operation unit; The processor is used to store the calculation result as a second file in the memory.
2. The data processing device as described in claim 1, characterized in that, When the processor determines that the format of the first file conforms to the operation format, the processor uses the first file as the operation data of the operation unit.
3. The data processing device as described in claim 1, characterized in that, The computing unit includes a machine learning model, which includes multiple parameters. When the processor determines that the memory has been written to a model file, the processor updates at least one of the multiple parameters in the machine learning model according to the model file.
4. The data processing device as described in claim 3, characterized in that, The processor is used to monitor the memory, and when the memory is written with an end instruction, the processor is used to execute the machine learning model according to the first file.
5. The data processing device as described in claim 1, characterized in that, Also includes: A host computer is coupled to the computing accelerator, and the host computer identifies the computing accelerator as an external storage device based on the storage device emulation unit.
6. The data processing device as described in claim 5, characterized in that, The host is used to provide model files to the computing unit as machine learning models in the computing unit, or to update the machine learning models in the computing unit.
7. A data processing method, characterized in that, Executed on a computing accelerator, the data processing method includes: The processor in the computing accelerator determines whether the memory coupled to the processor has been written to the first file. When the first file is written to the memory, the first file is used as computation data; The computational unit in the computational accelerator performs computation on the computational data to generate computational results; as well as The calculation result is stored in the memory as a second file.
8. The data processing method as described in claim 7, characterized in that, The computing unit includes a machine learning model, which has multiple parameters, and the data processing method further includes: When it is determined that the memory has been written to the model file, at least one of the plurality of parameters in the machine learning model is updated according to the model file.
9. The data processing method as described in claim 8, characterized in that, The method of using the first file as the computation data further includes: Monitor the memory; and When the memory is written with an end instruction, the machine learning model is executed according to the first file.
10. A computing accelerator, characterized in that, Include: The arithmetic unit is used to perform calculations based on the data to produce the results; and A storage device emulation unit includes a processor and a memory, wherein the processor is coupled to the arithmetic unit and the memory and is configured to set the memory as an external memory, and the processor is also configured to determine the type of data written to the memory. When the data written to the memory is a model file, the processor updates the computing unit according to the model file; When the data written to the memory is a first file, the processor uses the first file as the calculation data to obtain the calculation result generated by the calculation unit; The types of the computational data and the model files are different.