CN114079627B - Data transmission device and method - Google Patents

Abstract

The application discloses a data transmission device and a data transmission method, relates to the technical field of data transmission, and is beneficial to improving the data transmission efficiency. The data transmission device comprises a router, a first network interface and a plurality of second network interfaces. After the first network interface receives the data packet including the identifiers of the plurality of destination network interfaces sent by the router, the first network interface may broadcast the data packet in a first subnet, where the first subnet includes the first network interface and a plurality of second network interfaces, and the plurality of second network interfaces includes the plurality of destination network interfaces. After the plurality of second network interfaces in the first subnet receive the data packet sent by the first network interface, the plurality of destination network interfaces can forward the data packet to the corresponding processing node according to the identifiers of the destination network interfaces in the data packet, and other network interfaces except the plurality of destination network interfaces in the plurality of second network interfaces can discard the data packet according to the identifiers of the plurality of destination network interfaces.

Description

Data transmission device and method

Technical Field

The present application relates to the technical field of data transmission, and in particular to a data transmission device and method.

Background technique

In recent years, neural network technology has developed rapidly and has achieved widespread success in many fields such as pattern recognition, prediction and estimation, and image processing. With the increasing complexity of computing tasks, the scale of neural network processing nodes has gradually expanded. How to achieve efficient communication between nodes has become an important task that needs to be solved urgently.

Convolutional neural network is a network structure that is more suitable for image processing and pattern recognition. With the rise of the research boom of convolutional neural network, the model depth of convolutional neural network has been continuously deepened to obtain higher algorithm performance. But at the same time, problems such as large local data volume and intensive data transmission of convolutional neural network have gradually emerged. In this case, how to quickly and efficiently transmit data to improve the computing speed of convolutional neural network has become a technical problem that needs to be solved urgently.

Summary of the invention

The embodiment of the present application provides a data transmission device and method, which helps to improve the data transmission efficiency. When the technical solution is applied to a neural network system, it helps to improve the data transmission efficiency and calculation speed of the neural network system. The neural network system here can be, for example, a convolutional neural network.

In order to achieve the above objectives, the present application provides the following technical solutions:

In a first aspect, a data transmission device is provided, the data transmission device comprising: a router, configured to send a data packet to a first network interface, the data packet including identifiers of multiple destination network interfaces. The first network interface, configured to: receive a data packet sent by the router, and broadcast the data packet in a first subnet, wherein the first subnet comprises a first network interface and multiple second network interfaces, the multiple second network interfaces comprising multiple destination network interfaces. The multiple second network interfaces are configured to receive a data packet sent by the first network interface, wherein the multiple destination network interfaces are further configured to forward the data packet to a corresponding processing node according to the identifier of the destination network interface in the data packet, and the other network interfaces among the multiple second network interfaces except the multiple destination network interfaces are further configured to discard the data packet according to the identifier of the multiple destination network interfaces.

In this technical solution, data packets are sent in a broadcast mode, and received data packets are processed in a multicast mode. Since data packets are sent in a broadcast mode, no path selection is required during the transmission of data packets from the sending end to the destination end. Therefore, it helps to save the complexity of data processing during data packet transmission, thereby saving data transmission time. Processing the data packet in a multicast mode allows the data packet to be sent to a selectable destination (i.e., the above-mentioned multiple destination network interfaces) rather than all destinations in a broadcast mode (i.e., all second network interfaces in the above-mentioned first subnet). Therefore, combining the respective characteristics of the broadcast mode and the multicast mode, in the scenario of transmitting multicast data packets, sending data packets in a broadcast mode and processing received data packets in a multicast mode can help to improve data transmission efficiency while ensuring that the data packets are transmitted to the correct destination. The correct destination here (i.e., the above-mentioned destination network interface) can be pre-configured based on the network topology. When this technical solution is applied to a neural network system, it helps to improve the data transmission efficiency and calculation speed of the neural network system.

In one possible design, the first network interface is used to broadcast a data packet to multiple second network interfaces in the first subnet according to the identifiers of multiple destination network interfaces. For example, if the first network interface 201 parses the header of the data packet and determines that the data packet contains the identifiers of multiple destination network interfaces, the data packet is broadcast to the first subnet.

In one possible design, the first network interface is used to broadcast the data packet to multiple second network interfaces in the first subnet according to the broadcast identifier included in the data packet. For example, if the first network interface 201 parses the header of the data packet and determines that the data packet contains a broadcast identifier, the data packet is broadcast to the first subnet.

In a possible design, the data packet also includes an identifier of the first subnet. The router is used to send the data packet to the first network interface according to the identifier of the first subnet. This creates a condition for "the router to connect multiple subnets and transmit data packets between multiple subnets". That is, the router can also connect to other subnets, thereby sending data packets to other subnets based on other subnet identifiers.

In a possible design, the data transmission device also includes a third network interface, and the router is also used to communicate with the third network interface, wherein the third network interface is a network interface in the second subnet, and the second subnet includes the third network interface and multiple fourth network interfaces. This possible design provides a specific implementation method for connecting multiple subnets to a router. The function of the third network interface can refer to the function of the first network interface in the first subnet. The function of the fourth network interface can refer to the function of the second network interface in the first subnet. No further details are given here.

In a possible design, some network interfaces in the first subnet also belong to the second subnet.

In one possible design, the data transmission device further includes a plurality of configuration switches. A first configuration switch among the plurality of configuration switches is connected to the first network interface and is used to receive data packets sent by the first network interface and to send data packets to other configuration switches connected to a plurality of second network interfaces in the first subnet.

In one possible design, each of the multiple configuration switches includes multiple ports, a first port of the multiple ports is used to connect to a network interface, and other network interfaces of the multiple ports are used to connect to other configuration switches. Each of the multiple configuration switches is also used to control the connection or disconnection status between its own multiple ports based on configuration information.

By setting up a configuration switch, it is helpful to control the connected or disconnected state between network interfaces in the subnet and the data transmission direction by controlling the configuration switch. Similarly, setting up configuration switches in different subnets helps to control the connected or disconnected state between network interfaces between subnets by controlling the configuration switch, thereby realizing that the subnets in the data transmission device are configurable, and the data transmission direction between network interfaces in the subnets is configurable. In this way, a hardware system using a data transmission device can support multiple network topologies.

In one possible design, the data transmission device is a neural network chip, and the processing node is a processing node in the neural network.

In one possible design, the network interface in the first subnet includes: a network interface connecting a first processing node of a first neural network layer in a neural network chip and a network interface connecting a plurality of second processing nodes of a second neural network layer, wherein the first processing node is used to transmit data to the plurality of second processing nodes. For example, the first network interface includes a network interface connecting a first processing node of a first neural network layer in a neural network chip, and the plurality of second network interfaces include a network interface connecting a plurality of second processing nodes of a second neural network layer. This optional implementation provides a method for determining a network interface in a subnet based on a network topology of a neural network. This helps to improve the efficiency of data transmission between neural network layers.

In one possible design, the amount of data packets transmitted between network interfaces in the first subnet is greater than the first threshold. This optional implementation provides a method of using a network interface with a larger amount of data packets transmitted as a subnet. This helps to improve data transmission efficiency.

In a second aspect, a data transmission method is provided, the method comprising: a first network interface receives a data packet sent by a router, the data packet including identifiers of multiple destination network interfaces. Then, the first network interface broadcasts the data packet in a first subnet, wherein the first subnet includes a first network interface and multiple second network interfaces, the multiple second network interfaces including multiple destination network interfaces. The multiple destination network interfaces forward the data packet to a corresponding processing node according to the identifiers of the destination network interfaces in the data packet; and the other network interfaces among the multiple second network interfaces except the multiple destination network interfaces discard the data packet according to the identifiers of the destination network interfaces in the data packet.

In one possible design, the first network interface broadcasts the data packet to the first subnet, including: the first network interface broadcasts the data packet to multiple second network interfaces in the first subnet according to identifiers of multiple destination network interfaces.

In one possible design, the first network interface broadcasts a data packet to a first subnet, including: the first network interface broadcasts the data packet to multiple second network interfaces in the first subnet according to a broadcast identifier in the data packet.

In a third aspect, a data transmission method is provided, the method comprising: a second network interface receives a data packet broadcasted by a first network interface, the data packet including identifiers of multiple destination network interfaces; then, when the second network interface determines that the identifier of the destination network interface in the data packet includes the identifier of the second network interface, the data packet is forwarded to a corresponding processing node; or when the second network interface determines that the identifier of the destination network interface in the data packet does not include the identifier of the second network interface, the second network interface discards the data packet.

In a fourth aspect, a network interface is provided, comprising: a receiving unit, configured to receive a data packet broadcasted by a first network interface, the data packet including identifiers of multiple destination network interfaces; a processing unit, configured to forward the data packet to a corresponding processing node when determining that the identifier of the destination network interface in the data packet includes the identifier of the second network interface; or to discard the data packet when determining that the identifier of the destination network interface in the data packet does not include the identifier of the second network interface.

In a fifth aspect, a network interface is provided.

In one possible design, the network interface includes: a memory and one or more processors, the memory is used to store computer instructions, and the processor is used to call the computer instructions to execute the method provided in the fourth aspect.

In another possible design, the network interface includes: a control circuit for implementing the method provided in the fourth aspect above, and one or more ports.

In a sixth aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is run on a computer, the computer executes the method provided in the fourth aspect.

In a seventh aspect, a computer program product is provided, which, when executed on a computer, enables the method provided in the fourth aspect to be executed.

It can be understood that any of the data transmission methods, network interfaces, computer storage media, computer program products or chip systems provided above can be applied to the corresponding data transmission devices provided above. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding data transmission devices and will not be repeated here.

In this application, the name of the above data transmission device does not limit the device or functional module itself. In actual implementation, these devices or functional modules may appear with other names. As long as the functions of each device or functional module are similar to those of this application, they fall within the scope of the claims of this application and their equivalent technologies.

These and other aspects of the present application will become more apparent from the following description.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of the structure of a neural network system provided in an embodiment of the present application;

FIG2 is a schematic diagram of the structure of a computing node in a neural network chip provided in an embodiment of the present application;

FIG3 is a schematic diagram of the structure of a data transmission device provided in an embodiment of the present application;

FIG4 is a schematic diagram of the structure of another data transmission device provided in an embodiment of the present application;

FIG5 is a schematic diagram of the structure of another data transmission device provided in an embodiment of the present application;

FIG6 is an interactive schematic diagram of a data transmission method provided in an embodiment of the present application;

FIG7 is a schematic diagram of a format of a data packet provided in an embodiment of the present application;

FIG8 is a schematic diagram of nodes involved in a data transmission device and connection relationships therebetween provided in an embodiment of the present application;

FIG9 is a schematic diagram of the structure of a basic unit of a data transmission device provided in an embodiment of the present application;

FIG10 is a schematic structural diagram of a data transmission device provided by an embodiment of the present application based on the basic unit shown in FIG9 ;

FIG11 is a schematic diagram of the structure of a configuration switch provided in an embodiment of the present application;

12 is a schematic diagram of the structure of configuration information stored in a configuration register in a configuration switch provided in an embodiment of the present application;

13 is a schematic diagram of the structure of configuration information stored in a configuration register in another configuration switch provided in an embodiment of the present application;

14 is a flow chart of a data transmission method for configuring a state bus of a switch provided in an embodiment of the present application;

15 is a flow chart of a data transmission method for configuring a data bus of a switch provided in an embodiment of the present application;

FIG16 is a schematic diagram of the structure of a network interface provided in an embodiment of the present application;

17 is a schematic diagram of a format of configuration information stored in a configuration register in a network interface provided by an embodiment of the present application;

FIG18 is a schematic diagram of a method of a workflow of a network interface provided in an embodiment of the present application;

FIG19 is a schematic diagram of the structure of another data transmission device provided in an embodiment of the present application;

FIG20 is a schematic diagram of a configuration result of a configuration switch provided in an embodiment of the present application;

FIG21 is a schematic diagram of the structure of a data packet provided in an embodiment of the present application;

22 is a schematic diagram of configuration information stored in a configuration register in a network interface provided in an embodiment of the present application;

FIG. 23 is a schematic diagram of the structure of a network interface provided in an embodiment of the present application.

Detailed ways

The technical solution provided in the embodiment of the present application can be applied to artificial neural network (ANN). Artificial neural network, referred to as neural network (NN) or quasi-neural network, is a mathematical model or computational model that imitates the structure and function of biological neural network (such as the central nervous system of animals, especially the brain) in the field of machine learning and cognitive science, and is used to estimate or approximate functions. Artificial neural network can include convolutional neural network (CNN), deep neural network (DNN), time delay neural network (TDNN) and multilayer perceptron (MLP) and other neural networks.

FIG1 is a schematic diagram of the structure of a neural network system provided in an embodiment of the present application. As shown in FIG1 , the neural network system 100 may include a host 105 and a neural network circuit 110. The neural network circuit 110 is connected to the host 105 via a host interface. The host interface may include a standard host interface and a network interface. For example, the host interface may include a peripheral component interconnect express (PCIE) interface. As shown in FIG1 , the neural network circuit 110 may be connected to the host 105 via a PCIE bus 106. Therefore, data may be input into the neural network circuit 110 via the PCIE bus 106, and data processed by the neural network circuit 110 may be received via the PCIE bus 106. In addition, the host 105 may also monitor the working status of the neural network circuit 110 via the host interface.

The host 105 may include a processor 1052 and a memory 1054. It should be noted that, in addition to the components shown in FIG1 , the host 105 may also include other components such as a communication interface and a disk as an external memory, which are not limited here.

Processor 1052 is the computing core and control unit of host 105. Processor 1052 may include multiple processor cores. Processor 1052 may be a very large scale integrated circuit. An operating system and other software programs are installed in processor 1052, so that processor 1052 can access memory 1054, cache, disk and peripheral devices (such as the neural network circuit in Figure 1). It can be understood that in an embodiment of the present invention, the Core in processor 1052 may be, for example, a central processing unit (CPU) or other application specific integrated circuits (ASIC).

The memory 1054 is the main memory of the host 105. The memory 1054 is connected to the processor 1052 via a double data rate (DDR) bus. The memory 1054 is usually used to store various running software in the operating system, input and output data, and information exchanged with external memory. In order to improve the access speed of the processor 1052, the memory 1054 needs to have the advantage of fast access speed. In the traditional computer system architecture, dynamic random access memory (DRAM) is usually used as the memory 1054. The processor 1052 can access the memory 1054 at high speed through the memory controller (not shown in FIG. 1), and perform read and write operations on any storage unit in the memory 1054.

The neural network circuit 110 is a chip array composed of multiple neural network chips (chips). For example, as shown in FIG1 , the neural network circuit 110 includes multiple neural network chips 115 for data processing and multiple routers 120. For the convenience of description, the neural network chip 115 in the application is referred to as chip 115 in the embodiment of the present invention. In one example, the multiple chips 115 can be interconnected through a router 120. For example, a chip 115 can be connected to one or more routers 120. Multiple routers 120 can form one or more network topologies. Data can be transmitted between chips 115 through the multiple network topologies. In another example, the multiple chips 115 can be connected to each other through a PCIE bus.

FIG2 is a schematic diagram of the structure of a computing node in a neural network chip provided in an embodiment of the present application. As shown in FIG2, the chip 115 includes multiple routers 120, each of which can be connected to a tile 125. In practical applications, a router 120 can also be connected to multiple tiles 125, wherein each tile 125 can include one or more computing units for performing neural network calculations on data. For ease of description, in the embodiment of the present application, the chip 115 or tile 125 may also be referred to as a computing node.

System-on-a-chip (SoC) refers to a technology that integrates a complete system on a single chip and groups all or part of the necessary electronic circuits. When applied to the embodiment of the present application, SoC is a system on chip 115.

Figures 3, 4 and 5 are schematic diagrams of the structure of a data transmission device 20 provided in an embodiment of the present application. The data transmission device 20 may include a router 200 and at least one subnet. A subnet is a collection of network interfaces.

As shown in FIG3 , at least one subnet includes a first subnet, and the first subnet includes a first network interface 201 and a plurality of second network interfaces 202. The first network interface 201 is a network interface for broadcasting data packets in the first subnet. One first network interface 201 can be connected to at least one processing node 205. The second network interface 202 is a network interface for receiving broadcast data packets in the first subnet. One second network interface 202 can be connected to at least one processing node 205.

Optionally, the data transmission device 20 may be a neural network chip (such as the chip 115 in FIG. 1 ), in which case the tile 125 in FIG. 2 may be referred to as a processing node in the neural network. In another case, the data transmission device 20 may also refer to a neural network circuit including a plurality of neural network chips (such as the neural network circuit 110 in FIG. 1 ), in which case the neural network chip 115 may be used as a processing node in the neural network.

The router 200 is used to send a data packet to the first network interface 201. The data packet includes identifiers of multiple destination network interfaces. Optionally, the data packet also includes a first subnet identifier. The router 200 is used to send the data packet to the first network interface 201 according to the first subnet identifier.

The first network interface 201 is used to receive the data packet sent by the router 200 and broadcast the data packet in the first subnet, wherein the first subnet includes the first network interface 201 and multiple second network interfaces 202, and the multiple second network interfaces 202 include the multiple destination network interfaces. The multiple network interfaces may be part or all of the second network interfaces 202, and which network interfaces are specifically determined based on the network topology.

In one implementation, the first network interface 201 is used to broadcast the data packet to the multiple second network interfaces 202 in the first subnet according to the identifiers of the multiple destination network interfaces. For example, if the first network interface 201 parses the header of the data packet and determines that the data packet contains the identifiers of multiple destination network interfaces, the data packet is broadcast to the first subnet.

In another implementation, the first network interface 201 is used to broadcast the data packet to multiple second network interfaces 202 in the first subnet according to the broadcast identifier included in the data packet. For example, if the first network interface 201 parses the header of the data packet and determines that the data packet contains a broadcast identifier, the data packet is broadcast to the first subnet. In addition, if the first network interface 201 determines that the data packet does not contain a broadcast identifier, or contains another identifier (such as a non-broadcast identifier, etc.), the data packet is sent to a processing node connected to the first network interface 201.

The plurality of second network interfaces 202 are used to receive the data packet sent by the first network interface 201. The plurality of destination network interfaces are also used to forward the data packet to the corresponding processing node 205 according to the identifier of the destination network interface in the data packet, and the other network interfaces among the plurality of second network interfaces except the plurality of destination network interfaces are also used to discard the data packet according to the identifier of the plurality of destination network interfaces.

Specifically, for any second network interface 202, after receiving the data packet sent by the first network interface 201, the header portion of the data packet can be parsed to obtain the identifier of the destination network interface included in the data packet. If it is determined that its own identifier is included in the identifier of the destination network interface in the data packet, it means that the second network interface 202 is the destination network interface, and the data packet is forwarded to one or more processing nodes connected to itself. Among them, it can be known to which processing nodes the data packet is forwarded to by further parsing the data packet (such as the header portion of the data packet). For any second network interface, if it is determined that its own identifier is not included in the identifier of the destination network interface in the data packet, it means that the second network interface 202 is not the destination network interface, and the data packet is discarded.

In the data transmission device 20 provided in this embodiment, data packets are sent in a broadcast mode, and received data packets are processed in a multicast mode. It should be noted that in the traditional broadcast mode, the data packet does not carry the identifier of the destination end, while in the technical solution provided in the embodiment of the present application, the data packet transmitted in the broadcast mode carries the identifier of the destination end, so that the second network interface that receives the data packet can process the data packet in a multicast mode.

Since the data packet is sent in a broadcast mode, no path selection is required during the transmission of the data packet from the sender to the destination, which helps to save the complexity of data processing during the data packet transmission, thereby saving data transmission time. Processing the data packet in a multicast mode allows the data packet to be sent to a selectable destination (i.e., the above-mentioned multiple destination network interfaces) rather than all destinations in the broadcast mode (i.e., all second network interfaces in the above-mentioned first subnet). Therefore, combining the respective characteristics of the broadcast mode and the multicast mode, in the scenario of transmitting multicast data packets, sending data packets in a broadcast mode and processing received data packets in a multicast mode can help improve data transmission efficiency while ensuring that the data packets are transmitted to the correct destination. The correct destination here (i.e., the above-mentioned destination network interface) can be pre-configured based on the network topology.

For neural network chips, there is a lot of multicast traffic, so the above technical solution is particularly suitable for neural network chips.

The network interface used to broadcast data packets in a subnet can be considered as the entrance of the subnet. Data packets transmitted to the subnet from the outside (such as a host or other subnet) are sent to the network interface by the router connected to the network interface. For example, the first network interface can be considered as the entrance of the first subnet. Data transmitted to the first subnet from a host or other subnet is first sent to the first network interface by the router connected to the first network interface.

One of the network interfaces in a subnet can be used as the exit of the subnet. Data packets transmitted from the subnet to the outside world (such as a host or other subnets) are sent by the network interface to the router connected to the network interface, and are directly or indirectly sent by the router to the host or other subnets. For example, the first network interface or any second network interface can be used as the exit of the first subnet. Data packets transmitted from the first subnet to the host or other subnets are sent by the first network interface to the router connected to the first network interface, and are directly or indirectly sent by the router to the host or other subnets. Which network interface is used as the exit of the first subnet can be determined based on the network topology.

Optionally, each network interface (including the first network interface 201 and the second network interface 202) can be connected to the same router, or can be connected to different routers, and the different routers are directly or indirectly connected to each other. In this way, it is helpful to support that each network interface can be used as the entrance or exit of the subnet, so that the data transmission device 20 supports multiple network topologies or supports multiple neural network services.

As shown in FIG4 , at least one subnet of the data transmission device 20 includes a first subnet and a second subnet. FIG4 is drawn based on FIG3 . The second subnet may include a third network interface 203 and multiple fourth network interfaces 204. The third network interface 203 is a network interface for broadcasting data packets in the second subnet. A third network interface 203 can be connected to at least one processing node 205. The fourth network interface 204 is a network interface for receiving broadcast data packets in the second subnet. A fourth network interface 204 can be connected to at least one processing node 205. Among them, the function of the third network interface 203 is the same as or similar to that of the first network interface 201, and the function of the fourth network interface 204 is the same as or similar to that of the second network interface 202, which will not be repeated here.

Data is transmitted between different subnets through routers. Network interfaces used to broadcast data packets in different subnets can be connected to the same router or different routers. For example, the first network interface 201 and the third network interface 203 are connected to the same router 200, as shown in FIG4. For another example, the first network interface 201 and the third network interface 203 are respectively connected to different routers, and the routers to which they are connected are directly connected, or indirectly connected through other routers.

The number of processing nodes connected to different network interfaces may be the same or different.

Which network interfaces in the data transmission device 20 constitute a subnet, which network interface in a subnet is used as a network interface for broadcasting data packets, which network interface or network interfaces are used as network interfaces for receiving broadcast data packets, etc. can all be pre-configured, which will be described in detail below.

The following describes several specific implementations of the network interface in the first subnet provided in the embodiments of the present application:

In one implementation, the network interface in the first subnet includes: a network interface connecting a first processing node of a first neural network layer in a neural network chip and a network interface connecting a plurality of second processing nodes of a second neural network layer, wherein the first processing node is used to transmit data to the plurality of second processing nodes. For example, the first network interface includes a network interface connecting a first processing node of a first neural network layer in the neural network chip, and the plurality of second network interfaces include a network interface connecting a plurality of second processing nodes of a second neural network layer. This optional implementation provides a method for determining a network interface in a subnet based on a network topology of a neural network. This helps to improve the efficiency of data transmission between neural network layers.

For example, assuming that the first neural network layer includes processing nodes 11, 12, and 13, the second neural network layer includes processing nodes 21, 22, 23, and 24, and processing node 13 is used to transmit data to processing nodes 21, 22, and 23 respectively, then the network interfaces used to connect processing nodes 13, 21, 22, and 23 respectively are used as network interfaces in a subnet.

In another implementation, the amount of data packets transmitted between network interfaces in the first subnet is greater than the first threshold. This optional implementation provides a method of using the network interface with a larger amount of data packets transmitted as a subnet. This helps to improve data transmission efficiency.

It should be noted that, in a specific implementation, the above two implementation methods can be used in combination to obtain a new technical solution. Specifically, if the amount of data packets transmitted between the first processing node and the plurality of second processing nodes is greater than a first threshold, the above first processing node and the plurality of second processing nodes are used as network interfaces in the first subnet.

As shown in Fig. 5, the data transmission device 20 may further include a plurality of configuration switches 206. A first configuration switch among the plurality of configuration switches 206 is connected to the first network interface 201, and is used to receive a data packet sent by the first network interface 201, and to send the data packet to other configuration switches 206 connected to the plurality of second network interfaces 202 in the first subnet.

By setting a configuration switch in the data transmission device 20, it is helpful to control the connectivity or disconnection status between the network interfaces in the subnets and the direction of data transmission by controlling the configuration switch. Similarly, setting configuration switches in different subnets helps to control the connectivity or disconnection status between the network interfaces between the subnets by controlling the configuration switches, thereby achieving that the subnets in the data transmission device 20 are configurable, and the data transmission direction between the network interfaces in the subnets is configurable. In this way, a set of hardware systems using the data transmission device 20 can support multiple network topologies. When applied to a neural network system, the network topology refers to a neural network topology. That is to say, the technical data transmission device 20 provided in the embodiment of the present application supports processing different neural network services.

For neural network chips, by configuring the configuration information in the configuration switch, the network interfaces connected to the neural network nodes with dense local traffic can be formed into a subnet, so that multicast data packets can be transmitted in the subnet using broadcast mode to improve data transmission efficiency. Since the configuration information in the configuration switch can be pre-configured, the pre-configured information has set the transmission path of the data packets between the network interfaces. Therefore, in the actual application process, there is no need to rebuild the links between the network interfaces, nor is there any need to arbitrate the data packets and parse multiple messages. Among them, the parsing of multiple messages and the arbitration of data packets are used by the network interface to determine whether the data packet is a data packet sent to itself. This is conducive to reducing the complexity of data processing, thereby improving data transmission efficiency.

Optionally, the first network interface 201 and each second network interface 202 are respectively connected to a configuration switch. Different configuration switches 206 are directly or indirectly connected to each other, as shown in FIG5 .

Optionally, each configuration switch 206 includes multiple ports, a first port among the multiple ports is used to connect a network interface, and other network interfaces among the multiple ports are used to connect other configuration switches. As shown in Figure 4. Each configuration switch 206 is also used to control the connected or disconnected state between its multiple ports based on the configuration information. And, each configuration switch 206 is also used to control the data transmission direction between its multiple ports based on the configuration information. The specific implementation method can be referred to below.

Optionally, some network interfaces in one subnet may belong to other subnets, for example, some network interfaces in the first subnet belong to the second subnet.

It should be noted that if a network interface belongs to multiple subnets, the network interface acts as a second network node in one of the subnets and as an intermediate network interface in the other subnets. Among them, if a network interface is an intermediate network interface in a subnet, then in the subnet, the intermediate network interface will not receive data packets sent by the configuration switch to which it is connected, and therefore will not send data packets to the processing node connected to itself. The first network interface and the second network interface can send data packets to the processing node connected to themselves. In specific implementation, the configuration switch is controlled not to send data packets to the intermediate network interface by configuring the connectivity and disconnection of the port of the configuration switch connected to the intermediate network interface. For specific examples of intermediate network interfaces, please refer to the following, which will not be repeated here.

Optionally, the configuration switch includes one or more configuration registers, and one configuration register may correspond to one subnet, and is used to configure a data transmission path in the subnet. Specific examples may be referred to below, and will not be described in detail here.

As shown in Figure 6, it is an interactive schematic diagram of a data transmission method provided in an embodiment of the present application. The method shown in Figure 6 can be based on the data transmission device 20 provided in any of Figures 3 to 5. The method shown in Figure 6 includes the following steps:

S101: The router sends a data packet to the first network interface.

After executing S101, execute S102 and/or S106.

S102: If the first network interface determines that the data packet contains the identifier of the first network interface, the first network interface sends the data packet to a processing node connected to the first network interface.

S103: After the processing node connected to the first network interface processes the data packet, it generates a new data packet and sends the new data packet to the first network interface. The new data packet may include the identifier of the target subnet or the identifier of the host. The target subnet includes the first subnet and/or other subnets.

After executing S103, execute S104 or S105.

S104: If the first network interface determines that the new data packet contains the identifier of the first subnet, the new data packet is broadcast in the first subnet. After S104 is executed, the data transmission process ends.

S105: If the first network interface determines that the new data packet contains the identifier of another subnet, it sends the data packet to the network interface for broadcasting data packets in the other subnet through the router connected to the first network interface. If it is determined that the new data packet contains the identifier of the host, it sends the data packet to the host through the router connected to the first network interface. After executing S105, the data transmission process ends.

S106: If the first network interface determines that the data packet contains identifiers of other network interfaces, the first network interface broadcasts the data packet in the first subnet.

S107: After receiving the data packet, the second network interface in the first subnet forwards the data packet to one or more processing nodes connected to the second network interface if it determines that its own identifier is included in the identifier of the destination network interface, or discards the data packet if it determines that its own identifier is not included in the identifier of the destination network interface.

S108: After the processing node connected to the second network interface processes the data packet, it generates a new data packet and sends the new data packet to the second network interface. The data packet may include an identifier of another subnet (such as an identifier of the second subnet), or may include an identifier of a new destination network interface in the first subnet.

After executing S108, execute S109 and/or S110.

S109: The second network interface sends the new data packet to the router connected to the second network interface.

Subsequently, if the route determines that the data packet contains the identifier of other subnets, the router directly sends the new data packet to the network interface for broadcasting data packets in other subnets, or indirectly sends the new data packet to the network interface for broadcasting data packets in other subnets through the connection relationship between routers. If the router determines that the data packet contains the identifier of the new destination network interface in the first subnet, the router directly sends the new data packet to the first network interface, or indirectly sends the new data packet to the first network interface through the connection relationship between routers by the router connected to the second network interface.

The description of the beneficial effects that can be achieved by this embodiment can be found above and will not be repeated here.

The following is a description of the data packets provided in the embodiments of the present application:

As shown in Figure 7, it is a schematic diagram of the format of a data packet provided in an embodiment of the present application. The data packet shown in Figure 7 includes: a header part and a data part. The header part is used to carry control information of the data packet, and the data part is used to carry valid data.

The packet header may include the following flags/fields:

Field 1: Destination network interface flag, used to carry the identifier of the destination network interface.

In the embodiment described above, the destination network interface is part or all of the second network interface. Here, for the convenience of description, the destination network interface is defined as the destination to which the data packet is sent. Based on this, in specific implementation, taking the current subnet as the first subnet as an example, the destination network interface may include any multiple network interfaces in the current subnet, wherein the network interfaces may all be the second network interface, or include the first network interface and multiple second network interfaces.

In one example, since each network interface in the current subnet can be used as a destination network interface, the destination network interface flag can be set to include N bits, where N represents the number of network interfaces included in the current subnet, and each bit represents whether a network interface is a destination network interface. For example, if a network interface is a destination network interface, the bit corresponding to the network interface in the destination network interface flag is marked as 1; if the network interface is not a destination network interface, the bit corresponding to the network interface in the destination network interface flag is marked as 0.

In another example, if the data transmission device includes multiple subnets, and the multiple subnets contain different numbers of network interfaces, then in order to enable data packets to be transmitted between different subnets and recognized by nodes in each subnet, optionally, the flag bit of the target network interface contains N bits, where N represents the number of network interfaces contained in the target subnet, and the target subnet is the subnet in the data transmission device that contains the largest number of network interfaces.

Field 2: Data packet type flag, used to carry the type of data packet, which can be broadcast type or non-broadcast type.

Optionally, the header may also include:

Field 3: Subnet flag, used to carry the identifier of the subnet to which the data packet belongs. For example, if the data packet needs to be sent to the first subnet, the subnet flag of the data packet carries the identifier of the first subnet. After any router receives the data packet, it can directly or indirectly send the data packet to the first network interface in the first subnet.

In addition, the packet header may also include other flags/other fields.

The data part may include several fields. The specific field names and the meanings represented by the fields, etc., can all be referred to in the prior art and will not be described in detail here.

The above description is based on the perspective of one or more subnets to describe the data transmission device 20. The following description is based on the structure of the data transmission device 20 from another perspective:

In one example, as shown in FIG8 , it is a schematic diagram of the nodes involved in the data transmission device 20 and the connection relationship between them. Among them, these nodes include a router (R), a configuration switch (S), a network interface (N) and a processing node (T). The network interface (N) is respectively connected to the router (R), the configuration switch (S) and the processing node (T).

Wherein, the router (R) and the configuration switch (S) can be multi-port. A router (R) can be connected to one or more network interfaces (N), and different configuration switches (S) can be connected in a certain manner, as shown in FIG9. The structure shown in FIG9 can be considered as a basic unit structure of a data transmission device 20.

FIG9 is only an example, and does not limit the structure of the basic unit of the data transmission device 20 provided in the embodiment of the present application. For example, in actual implementation, the number of network interfaces (N) connected to a router (R), the number of processing nodes (T) connected to a network interface (N), and the number of other configuration switches (S) connected to a configuration switch (S) can be set as needed.

Different routers (R) can be connected in a certain manner, as shown in FIG10. FIG10 is a schematic diagram of the structure of a data transmission device 20 provided based on the basic unit shown in FIG9. FIG10 is only an example, and does not limit the structure of the data transmission device 20 provided in the embodiment of the present application. For example, the number of other routers (R) that a router (R) can connect to can be configured as needed.

As an example, in FIG10 , routers (R) are connected in a certain way to form a first layer network for transmitting data packets between subnets, and a network interface "for receiving broadcast data packets" in a subnet (such as the second network interface) transmits data packets to a network interface "for broadcasting data packets" in the subnet (such as the first network interface). Configuration switches (S) are connected in a certain way to form a second layer network for distinguishing subnets and controlling the data transmission direction between network interfaces in subnets.

The router (R) in FIG10 is an 8-port router, 4 of which are used to connect to other routers to form a network, and the remaining 4 ports are used to connect to different network interfaces. For example, an 8-port router can be obtained by expanding the ports of a classic router structure, for example, a classic 5-port router is expanded by 3 ports to become an 8-port router.

The configuration switch (S) in FIG10 has five ports, one of which is used to connect to a network interface, which can be referred to as a network interface port, and the remaining four ports: east, west, south, and north ports are used to connect to other configuration switches to form a mesh shape.

The following describes the configuration switch provided in the embodiment of the present application:

As shown in FIG11, a schematic diagram of the structure of a configuration switch provided in an embodiment of the present application is shown. The configuration switch shown in FIG11 includes east, west, south, north ports and a network interface port. It should be noted that the east, west, south, and north ports here are only examples and can be replaced with other names in actual implementation. In addition, the number of ports of the configuration switch used to connect to other configuration switches is not limited to 4, but can also be any other number.

The signal lines connecting each port of the configuration switch (S) and other nodes are bidirectional interfaces. The signal lines can be used as input signal lines or output signal lines, but can only be used as input signal lines or output signal lines at the same time.

Optionally, the signal lines connected to each port of the configuration switch (S) include a data bus, a data control bus and a status bus.

The data bus is the signal line that transmits data.

The data control bus is a signal line that indicates whether the data on the data bus is valid.

The status bus is a signal line indicating the status of the network interface connected to the configuration switch (such as a busy status or an idle status), and the status of the network interface sent by other configuration switches and received by the configuration switch.

For example, in the first subnet, configuration switch 1 is connected to the first network interface, configuration switches 2 and 3 are respectively connected to the second network interface, and configuration switch 3 sends its own status information to configuration switch 1 via configuration switch 2. In this case, the status information sent by configuration switch 3 to configuration switch 2 indicates that the network interface connected to configuration switch 3 is busy or idle. The status information sent by configuration switch 2 to configuration switch 1 indicates that the network interface connected to configuration switch 3 and the network interface connected to configuration switch 2 are busy or idle.

If a network interface connected to a configuration switch has free cache space to receive a data packet, the configuration switch is in an idle state, otherwise, the configuration register is in a busy state. For example, when the remaining cache space (or the remaining cache space accounts for the total cache space) in the network interface connected to the configuration switch for caching data from the configuration switch is greater than or equal to a certain threshold, the network interface has free cache space.

It is understandable that the data transmission directions on the data bus and the data control bus are the same, for example, both are in the "sending" direction, or both are in the "receiving direction".

It is understandable that the data transmission direction on the data bus is opposite to that on the status bus. Specifically:

If the network interface connected to a configuration switch is a network interface for broadcasting data packets (such as the first network interface mentioned above), the configuration switch is used to send data to other configuration switches connected to other network interfaces in the subnet. Therefore, the configuration switch needs to receive the status information of other configuration switches (i.e., busy state or idle state), so as to send data to other configuration switches when it is determined that other configuration switches are all in idle state. In other words, if the data transmission direction on the data bus is "send", the data transmission direction on the status bus is "receive".

Correspondingly, if the network interface connected to a configuration switch is a network interface for receiving broadcast data packets (such as the second network interface mentioned above), the configuration switch is used to directly or indirectly receive data sent by the configuration switch connected to the "network interface for broadcast data packets" in this subnet. Therefore, the configuration switch needs to report its own status information to the configuration switch connected to the "network interface for broadcast data packets". In other words, if the data transmission direction on the data bus is "receive", the data transmission direction on the status bus is "send".

Each configuration switch may include a configuration register therein, and the configuration register is used to store configuration information. The configuration information is used by the configuration switch to control the connection or disconnection of each port of the configuration switch, as well as the data transmission direction.

Optionally, the number of configuration registers configured in a configuration switch is related to the number of subnets for which the configuration switch needs to transmit data at the same time. The number of subnets for which a configuration switch needs to transmit data at the same time is determined based on the configuration information and the maximum number of subnets for which the configuration switch can simultaneously support data transmission. The maximum number of subnets for which a configuration switch can simultaneously support data transmission is determined based on the number of ports of the configuration switch.

Taking the configuration switch as shown in FIG11 as an example, since the configuration switch includes 5 ports, and data transmission between different subnets cannot share the ports of the configuration switch, the configuration switch can support data transmission for 2 subnets at most, for example, the east port and the south port serve as the input port and output port of the data transmission path of one subnet, and the west port and the north port serve as the input port and output port of the data transmission path of another subnet. In this case, the configuration switch can include 2 configuration registers, one of which is used to store the configuration information of one subnet, and the other is used to store the configuration information of another subnet.

Optionally, the format of the configuration information stored in the configuration register in the configuration switch mainly includes the following flag bits/fields, as shown in FIG12:

Field 1: data bus input port flag bit and/or status bus output port flag bit, used to carry the identifier of the data bus input port and/or the identifier of the status bus output port.

Field 2: Data bus output port flag, used to carry the identifier of the data bus output port.

Field 3: Status bus input port flag, used to carry the identification of the status bus input port.

It should be noted that, since the data transmission directions on the data bus and the status bus are opposite, the data bus input port flag bit is the same as the status bus output port flag bit, so the two can share a field.

It should also be noted that since the data transmission direction on the data bus and the data control bus is the same, the format of the configuration register in the configuration switch does not include the input port flag and output port flag of the data control bus. The specific flags can refer to the data bus input port flag and the data bus output port flag respectively.

As shown in FIG13, a schematic diagram of the structure of configuration information stored in a configuration register in a configuration switch provided in an embodiment of the present application is shown. The format of the configuration information shown in FIG13 includes:

Field 1: Bits 0-2 are used to indicate both the data input port flag and the output port flag of the status bus.

Field 2: Bits 3-7 indicate the data bus output port flag. In one example, bit 3 indicates the north port, bit 4 indicates the west port, bit 5 indicates the south port, bit 6 indicates the east port, and bit 7 indicates the network interface port (NI). When a bit is at a high level, it indicates that the data bus output port corresponding to the bit is valid, that is, the port corresponding to the bit is the data bus output port. When a bit is at a low level, it indicates that the data bus output port corresponding to the bit is invalid, that is, the port corresponding to the bit is not the data bus output port.

Field 3: Bits 8-12 indicate the status bus input port flag. In one example, bit 8 indicates the north port, bit 9 indicates the west port, bit 10 indicates the south port, bit 11 indicates the east port, and bit 12 indicates the network interface port. When a bit is at a high level, it indicates that the status bus input port corresponding to the bit is valid, that is, the port corresponding to the bit is the status bus input port. When a bit is at a low level, it indicates that the status bus input port corresponding to the bit is invalid, that is, the port corresponding to the bit is not the status bus input port.

As shown in Figure 14, it is a flow chart of a method for data transmission of a state bus of a configuration switch provided in an embodiment of the present application. The method shown in Figure 14 comprises the following steps:

S201: The configuration switch initializes the configuration register contained in itself. The format of the configuration information stored in the configuration register may be as shown in FIG. 13 .

S202: The configuration switch determines whether the output port flag bits (such as bits 0-2 in FIG. 13 ) of the status bus are valid based on the configuration information.

If invalid, execute S203. If valid, execute S204.

S203: The configuration switch determines that the status signal is not currently output. After executing this step, the current process ends.

S204: The configuration switch records the identifier of the valid port carried by the status bus input port flag bit (such as bits 8-12 in FIG. 13 ).

S205: The configuration switch performs an “OR” operation on the recorded status information of the valid ports.

This step is described by taking the busy state information as "1" and the idle state information as "0" as an example. If the state information of any one or more valid ports recorded by the configuration switch is "1", the result of the "or" operation in S205 is 1, indicating busy. If the state information of all valid ports recorded by the configuration switch is "0", the result of the "or" operation in S205 is 0, indicating idle.

S206: The configuration switch assigns the result value of the “OR” operation to the output port of the status bus.

Each configuration switch performs an OR operation on the input status information and assigns it to the output status information line. For the first subnet, each configuration switch can only perform an OR operation on its own input status information and assign it to its own output port, thereby transmitting the status information of the network interfaces connected to all the second network interfaces in the first subnet to the first network interface through the transmission of the configuration switches one by one.

As shown in FIG15 , a flow chart of a data transmission method for configuring a data bus of a switch provided in an embodiment of the present application is shown. The method shown in FIG15 includes the following steps:

S301: The configuration switch initializes the configuration register contained in itself. The format of the configuration information stored in the configuration register may be as shown in FIG. 13 .

S302: The configuration switch determines whether the input port flag bits (such as bits 0-2 in FIG. 13 ) of the data bus are valid based on the configuration information.

If invalid, execute S303; if valid, execute S304.

S303: The configuration switch determines that there is no data input at present. After executing this step, this process ends.

S304: The switch is configured to record the valid port number in the data bus output port flag (such as bits 3-7 in FIG. 13 ). The port indicated by the port number is the valid data bus output port.

S305: The configuration switch assigns the data of the input port of the data bus to the valid data bus output port, so that the data of the input port of the data bus can be output through the valid data bus output port.

The following describes the network interface provided in the embodiment of the present application:

As shown in FIG16, a schematic diagram of the structure of a network interface 30 provided in an embodiment of the present application is shown. For the convenience of explanation, FIG13 also schematically shows a router, a configuration switch, and a processing node connected to the network interface 30. The network interface 30 shown in FIG13 includes:

The first cache block 31 is used to cache data packets sent by the router to the network interface 30 .

The second cache block 32 is used to cache data packets sent to the network interface 30 by the configuration switch connected to the network interface 30 .

The configuration register (REG) 33 is used to store configuration information, which may be predefined and can be modified after being predefined.

The controller 34 is used to select the data packets in the first cache block 31 and the second cache block 32 to be output to the appropriate port, where the appropriate port may be a port connected to the processing node and/or a port connected to the configuration switch. And, according to the configuration information stored in the REG 341, the controller 34 performs information exchange with the configuration switch and the second cache block 32.

The third cache block 35 is used to cache data packets sent from the processing node connected to the network interface 30 to the router.

As shown in Figure 17, it is a schematic diagram of the format of configuration information stored in a configuration register in a network interface provided by an embodiment of the present application. Specifically, the configuration information includes the following fields:

Field 1: Master-slave working mode flag, used to carry the identification of the working mode of the control bus of the network interface. The working mode is the master working mode or the slave working mode. The master working mode indicates that the network interface is a network interface for broadcasting data packets, such as the first network interface mentioned above. The slave working mode indicates that the network interface is a network interface for receiving broadcast data packets, such as the second network interface mentioned above.

Field 2: The flag of the network interface in this subnet, used to carry the identification of the network interface in this subnet.

It should be noted that a network interface may belong to different subnets, and its identifiers in different subnets may be different. Therefore, the flag bit here specifically refers to the identifier of the network interface in this subnet.

Optionally, the format of the configuration information may also include other flag bits/fields.

As shown in Figure 18, a schematic diagram of a method of a workflow of a network interface provided in an embodiment of the present application is provided. Figure 18 can be applied to the network interface shown in Figure 16, and to the data packet shown in Figure 7. The method shown in Figure 18 may include the following steps:

S401: Initialize the configuration register inside the current network interface.

S402: The controller in the current network interface determines whether the working mode of the current network interface is the master node working mode or the slave node working mode based on the configuration information stored in the configuration register.

If it is configured as the master node working mode, execute S403.

If it is configured as the slave node working mode, execute S409.

S403: The controller in the current network interface detects whether there is data in the first cache block.

If yes, then S404 is executed. If no, then the data transmission process ends.

S404: The controller in the current network interface reads the data packet in the first cache block, and determines whether the “data packet type flag” of the data packet is a broadcast mark or a non-broadcast mark.

If the data packet type flag is a non-broadcast type, execute S405.

If the data packet type flag is a broadcast type, execute S406.

S405: The controller in the current network interface transmits the read data packet to the processing node.

After executing S405, the data transmission process ends.

S406: The controller in the current network interface determines whether the current network interface is valid based on the “destination network interface flag bit” of the data packet.

If valid, it means that the current network interface is a destination network interface of the data packet (or in other words, the current network interface is a valid destination network interface), and then S407 is executed.

If invalid, it means that the current network interface is not the destination network interface of the data packet (or in other words, the current network interface is not a valid destination network interface), and then S408 is executed.

It should be noted that the destination network port here may include a network interface for broadcasting data packets, and may also include a network interface for receiving broadcast data packets.

For example, if the current network interface is network interface n, where a subnet includes N network interfaces, and the network interfaces in the subnet are numbered 1-N, N is an integer greater than or equal to 3, then network interface n refers to the network interface numbered n in the subnet, 1≤n≤N, and n is an integer. If the controller of network interface n determines that the nth bit of the "destination network interface flag bit" of the data packet is valid, then the network interface is determined to be valid.

S407: The controller in the current network interface transmits the read data packet to the processing node connected to the current network interface, and when the status bus transmitted to the current network interface by the configuration switch connected to the current network interface is valid, the controller in the current network interface transmits the read data packet to the configuration switch connected to the current network interface.

The current network interface executing S407 is a network interface for broadcasting data packets, so the state bus transmitted to (i.e., output to) the current network interface by the configuration switch connected to the current network interface is valid, indicating that all network interfaces "for receiving broadcast data packets" in this subnet are in an idle state. At this time, the current network interface can broadcast data packets in the subnet, so the read data packets are transmitted to the configuration switch connected to the current network interface. Subsequently, the configuration switch can send the data packet to the configuration switch connected to it, and the configuration switch that receives the data packet can execute the method shown in Figures 14 and 15, thereby achieving the purpose of the current network interface broadcasting data packets in the subnet.

After executing S407, the data transmission process ends.

S408: When the status bus transmitted from the configuration switch connected to the current network interface to the current network interface is valid, the controller in the current network interface transmits the read data packet to the configuration switch connected to the current network interface.

After executing S408, the data transmission process ends.

S409: The controller in the current network interface detects whether the configuration switch connected to the current network interface transmits data, that is, determines whether a data packet is stored in the second cache block.

If yes, then S410 is executed. If no, then the data transmission process ends.

S410: The controller in the current network interface determines whether the current network interface is valid based on the “destination network interface flag bit” of the data packet.

If valid, it means that the current network interface is a valid destination network port, and then S411 is executed.

If invalid, the data transmission process ends.

S411: The controller in the current network interface writes the data in the second cache block to the processing node connected to the current network interface.

The method provided in the embodiment of the present application is described below by using a specific example:

As shown in FIG19, it is a schematic diagram of the structure of a data transmission device provided in an embodiment of the present application. The data transmission device includes a first subnet and a second subnet. Among them:

The first subnet includes the following network interfaces: network interface 1, 2, 3, and 4. Among them, network interface 1 is the first network interface, and network interfaces 2, 3, and 4 are the second network interfaces. The second subnet includes the following network interfaces: network interface 5, 3, 6, and 7. Among them, network interface 5 is the third network interface, and network interfaces 6 and 7 are the fourth network interfaces.

Among them, the network interface 3 belongs to the first subnet and also belongs to the second subnet. The network interface 3 serves as the second network interface in the first subnet and as the intermediate network interface in the second network node. This information can be stored in the configuration register of the network interface 3 as part of the configuration information.

The configuration registers connected to the network interfaces 1-7 (labeled as N1-7 in FIG. 19 ) are respectively configured switches 1-7 (labeled as S1-7 in FIG. 19 ). The direction of data flow between the configuration switches 1, 2, 3, 4 is shown by the solid arrows in FIG. 19 , and the direction of data flow between the configuration switches 3, 5, 6, 7 is shown by the dotted arrows in FIG. 19 .

In one example, the configuration results of the configuration switches respectively connected to the network interfaces 1 - 7 are shown in FIG. 20 .

In FIG20 , two configuration registers are set in each configuration switch, which are marked as REG1 and REG2. Among them, REG1 is a configuration register for the first subnet, and REG2 is a configuration register for the second subnet. The structure of the configuration information stored in each configuration register and the meaning of each field in the structure are shown in FIG13 .

In one example, the meaning of the information carried in the data bus input port flag/status bus output port flag is as follows: "000" represents the north port (N), "001" represents the west port (W), "010" represents the south port (S), "011" represents the east port (E), "100" represents the network interface port (NI), and "111" represents invalid (INVALID).

In the following, in conjunction with FIG. 20 , the meaning of the configuration information stored in the configuration register is explained by taking the configuration switch 3 as an example.

Field 1 "011" of REG1 indicates that the input port of the data bus is the east port and the output port of the status bus is the east port.

"00001" in field 2 of REG1 indicates that the NI port is the data bus output port, and the other ports are invalid. Therefore, the data flow of the data bus is: input from the east port and output from the NI port. That is to say, for the first subnet, after configuration switch 3 receives the data packet from configuration switch 2, it sends it to network interface 3.

Field 3 "00001" of REG1 indicates that the NI port is the input port of the status bus, and the other ports are invalid. Therefore, the data flow of the status bus is: input from the NI port and output from the East port.

Field 1 "001" of REG2 indicates that the input port of the data bus is the west port, and the output port of the status bus is the west port.

Field 2 "10000" of REG2 indicates that the north port is the data bus output port, and the other ports are invalid. Therefore, it can be seen that the data flow direction of the data bus is: input from the west port and output from the north port. In other words, for the second subnet, after configuration switch 3 receives the data packet from configuration switch 5, it sends it to configuration switch 6 instead of network interface 3. This is the configuration information designed considering that network interface 3 is an intermediate network interface in the second subnet.

Field 3 "10000" of REG2 indicates that the north port is the input port of the status bus, and the other ports are invalid. Therefore, the data flow direction of the status bus is: input from the north port and output from the west port.

It is understandable that, for a configuration switch, the configuration information in the corresponding configuration register of the configuration switch can be determined based on the preset data packet transmission path, so that during the data transmission process, the configuration register reads and analyzes the configuration information to achieve data packet transmission along the preset path. Other examples are not described one by one.

In one example, examples of data packets in the first subnet and the second subnet are shown in Figure 21. Among them, "11" in the data packet type flag indicates that the data packet type is a broadcast type. Figure 21 is illustrated by taking the destination network interface flag as 16 bits as an example, which is not shown in the specific implementation. Among them, the nth bit is a high level "1" indicating that network interface n is the destination network interface, and a low level "0" indicating that network interface n is not the destination network interface. 1≤n≤16, n is an integer. "Other fields" indicate other information contained in the data packet, and the relevant description can refer to the above text or the prior art.

In one example, the configuration information stored in the configuration register in the network interface 1 - 7 is as shown in FIG. 22 .

In conjunction with FIG. 20 to FIG. 22 , for the first subnet:

Based on the master-slave node flag bit "1" shown in FIG22, the network interface 1 determines that its own working mode is the master working mode, thereby broadcasting the data packet in the first subnet. Specifically, the data packet is sent to the configuration switch 2-4 by the configuration switch connected to the network interface 1. The configuration switch 2-4 sends the received data packets to the network interface 2-4 respectively based on the corresponding configuration information shown in FIG20.

Network interface 2-4 determines that its own working mode is slave working mode based on the master-slave node flag "0" shown in Figure 22, thereby receiving the data packet broadcasted in the first subnet. Then, network interface 2-4 determines whether to send the data packet to the processing node connected to itself or discard the data packet based on the destination network interface flag of the data packet shown in Figure 21. Specifically, from the destination network interface flag "11100000000000000" of the data packet in the first subnet in Figure 21, and the predefined 1-4 bits starting from the high position correspond to network interface 1-4 in sequence, it can be known that: network interface 1-3 is the target network interface, and network interface 4 is not the target network interface. Therefore, network interface 1-3 sends the data packet to the processing node connected to itself, and network interface 4 discards the received data packet.

In conjunction with FIG. 20 to FIG. 22 , for the second subnet:

Based on the master-slave node flag "1" shown in FIG22, the network interface 5 determines that its own working mode is the master working mode, thereby broadcasting the data packet in the second subnet. Specifically, the data packet is sent to the configuration switches 3, 6, and 7 by the configuration switch connected to the network interface 5. The configuration switches 3, 6, and 7 send the received data packets to the network interfaces 6 and 7 respectively based on the corresponding configuration information shown in FIG20.

Network interfaces 6 and 7 respectively determine that their own working mode is slave working mode based on the master-slave node flag "0" shown in Figure 22, thereby receiving the data packet broadcasted in the second subnet. Then, network interfaces 6 and 7 respectively determine whether to send the data packet to the processing node connected to themselves or discard the data packet based on the destination network interface flag of the data packet shown in Figure 21. Specifically, from the destination network interface flag "011000000000000" of the data packet in the second subnet in Figure 21, and the predefined 1-3 bits starting from the high position correspond to network interfaces 5, 6, and 7 in sequence, it can be known that: network interfaces 6 and 7 are the target network interfaces, and network interface 5 is not the target network interface. Therefore, network interfaces 6 and 7 respectively send the received data packets to the processing nodes connected to themselves.

The network interface 3 is an intermediate network interface in the second subnet, and therefore, will not receive the data packet sent by the configuration switch 3. For the detailed description of the configuration information in the configuration switch 3 in this case, reference can be made to the above text.

The above mainly introduces the solution provided by the embodiment of the present application from the perspective of the method. In order to achieve the above functions, it includes hardware structures and/or software modules corresponding to the execution of each function. Those skilled in the art should easily appreciate that, in combination with the method steps of each example described in the embodiment disclosed herein, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present application.

The embodiment of the present application can divide the network interface into functional modules according to the above method example. For example, each functional module can be divided according to each function, or two or more functions can be integrated into one processing module. The above integrated module can be implemented in the form of hardware or in the form of software functional modules. It should be noted that the division of modules in the embodiment of the present application is schematic and is only a logical function division. There may be other division methods in actual implementation.

As shown in Figure 23, it is a schematic diagram of the structure of a network interface 23 provided in an embodiment of the present application. The network interface 23 is used to execute the steps performed by the second network interface in the method provided in any one of the embodiments above. For example, the network interface 23 may include: a receiving unit 2301, used to receive a data packet broadcast by the first network interface, and the data packet includes identifiers of multiple destination network interfaces. A processing unit 2302 is used to forward the data packet to a corresponding processing node when it is determined that the identifier of the destination network interface in the data packet includes the identifier of the second network interface; or to discard the data packet when it is determined that the identifier of the destination network interface in the data packet does not include the identifier of the second network interface. For example, in combination with Figure 18, the processing unit 2302 can be used to execute S411.

The embodiment of the present application also provides a network interface. The network interface 24 includes: a memory and one or more processors, the memory is used to store computer instructions, and the processor is used to call the computer instructions to execute the steps performed by the second network interface in the method provided in any of the embodiments above.

The embodiment of the present application also provides a network interface for executing the steps performed by the first network interface in the method provided in any one of the embodiments above. In one implementation, the network interface can be divided into logical functions, such as including a transceiver unit and a processing unit, the transceiver unit is used to perform the transceiver step, and the processing unit is used to perform other steps other than the transceiver step. In another implementation, the network interface includes a memory and one or more processors, the memory is used to store computer instructions, and the processor is used to call the computer instructions to execute the steps performed by the first network interface in the method provided in any one of the embodiments above.

The embodiment of the present application also provides a configuration switch for executing the steps performed by the configuration switch in the method provided in any one of the embodiments above. In one implementation, the configuration switch can be divided into logical functions, such as including a transceiver unit and a processing unit, the transceiver unit is used to execute the transceiver step, and the processing unit is used to execute other steps other than the transceiver step. In another implementation, the configuration switch includes a memory and one or more processors, the memory is used to store computer instructions, and the processor is used to call the computer instructions to execute the steps performed by the configuration switch in the method provided in any one of the embodiments above.

An embodiment of the present application further provides a network interface, including a control circuit for implementing the above-mentioned first network interface function and/or a control circuit for implementing the second network interface function, and one or more ports.

An embodiment of the present application also provides a configuration switch, including a control circuit for implementing the above-mentioned configuration switch function, and one or more ports.

An embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program runs on a computer, the computer executes any method performed by the network interface provided above, or the method performed by the configuration switch.

For explanations of the relevant contents and descriptions of the beneficial effects of any of the network interfaces and configuration switches provided above, reference may be made to the corresponding embodiments described above, and no further details will be given here.

The embodiment of the present application also provides a chip. The chip integrates a control circuit and one or more ports for realizing the functions of the above-mentioned data transmission device. Optionally, the functions supported by the chip can be referred to above and will not be repeated here. A person of ordinary skill in the art can understand that all or part of the steps of implementing the above-mentioned embodiment can be completed by instructing the relevant hardware through a program. The program can be stored in a computer-readable storage medium. The storage medium mentioned above can be a read-only memory, a random access memory, etc. The above-mentioned processing unit or processor can be a central processing unit, a general-purpose processor, an application specific integrated circuit (ASIC), a microprocessor (digital signal processor, DSP), a field programmable gate array (field programmable gatearray, FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof.

The embodiment of the present application also provides a computer program product including instructions, when the instructions are run on a computer, the computer executes any one of the methods in the above embodiments. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the process or function according to the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from a website site, a computer, a server, or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means to another website site, computer, server, or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server, a data center, etc. that includes one or more servers that can be integrated with the medium. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., an SSD), etc.

It should be noted that the above-mentioned devices for storing computer instructions or computer programs provided in the embodiments of the present application, such as but not limited to the above-mentioned memories, computer-readable storage media, and communication chips, etc., are all non-transitory.

In the process of implementing the claimed application, those skilled in the art can understand and implement other changes to the disclosed embodiments by viewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "one" or "an" does not exclude multiple situations. A single processor or other unit can implement several functions listed in the claims. Certain measures are recorded in different dependent claims, but this does not mean that these measures cannot be combined to produce good results. Although the present application is described in conjunction with specific features and embodiments thereof, various modifications and combinations may be made to it without departing from the spirit and scope of the present application. Accordingly, this specification and the drawings are merely exemplary illustrations of the present application as defined by the appended claims, and are deemed to have covered any and all modifications, changes, combinations or equivalents within the scope of the present application.

Claims (16)

1. A data transmission apparatus, the data transmission apparatus being applied to a neural network system, the data transmission apparatus comprising:

The router is used for sending a data packet to the first network interface, wherein the data packet comprises identifiers of a plurality of destination network interfaces;

the first network interface is configured to: receiving the data packet sent by the router and broadcasting the data packet in a first subnet, wherein the first subnet comprises the first network interface and a plurality of second network interfaces, the plurality of second network interfaces comprise the plurality of destination network interfaces, the first network interface is a network interface for broadcasting the data packet in the first subnet, and the second network interface is a network interface for receiving the broadcasted data packet in the first subnet;

The plurality of second network interfaces are used for receiving the data packet sent by the first network interface, wherein the plurality of destination network interfaces are also used for forwarding the data packet to a corresponding processing node according to the identification of the destination network interface in the data packet, and other network interfaces except the plurality of destination network interfaces in the plurality of second network interfaces are also used for discarding the data packet according to the identification of the plurality of destination network interfaces.

2. The data transmission apparatus according to claim 1, wherein: the first network interface is configured to broadcast the data packet to the plurality of second network interfaces in the first subnet according to the identifiers of the plurality of destination network interfaces.

3. The data transmission apparatus according to claim 1, wherein: the first network interface is configured to broadcast the data packet to the plurality of second network interfaces in the first subnet according to a broadcast identifier included in the data packet.

4. A data transmission device according to any one of claims 1 to 3, wherein: the data packet further includes an identification of the first subnet;

the router is configured to send the data packet to the first network interface according to the first subnet identification.

5. The data transmission apparatus according to claim 1, wherein: the data transmission device further comprises a third network interface, and the router is further used for communicating with the third network interface, wherein the third network interface is a network interface in a second subnet, and the second subnet comprises the third network interface and a plurality of fourth network interfaces.

6. The data transmission device of claim 5, wherein: part of the network interfaces in the first subnetwork also belong to the second subnetwork.

7. A data transmission device according to claim 5 or 6, characterized in that: the data transmission device further comprises a plurality of configuration switches;

A first configuration switch of the plurality of configuration switches is connected to the first network interface and is configured to receive the data packet sent by the first network interface, and send the data packet to other configuration switches in the first subnetwork connected to the plurality of second network interfaces.

8. The data transmission apparatus according to claim 7, wherein: each configuration switch of the plurality of configuration switches comprises a plurality of ports, a first port of the plurality of ports being for connecting to one network interface, and other network interfaces of the plurality of ports being for connecting to other configuration switches;

Each of the plurality of configuration switches is further configured to control a connection or disconnection state between the plurality of ports of itself based on configuration information.

9. The data transmission apparatus according to claim 1, wherein: the data transmission device is a neural network chip, and the processing node is a processing node in the neural network.

10. The data transmission apparatus according to claim 9, wherein:

the first network interface comprises a network interface connected with a first processing node of a first neural network layer in the neural network chip, and the plurality of second network interfaces comprises a network interface connected with a plurality of second processing nodes of a second neural network layer, wherein the first processing node is used for transmitting data to the plurality of second processing nodes.

11. The data transmission device according to claim 9 or 10, characterized in that: the data volume of the data packets transmitted between the network interfaces in the first subnetwork is greater than a first threshold.

12. A data transmission method, wherein the data transmission method is applied to a neural network system, the method comprising:

The method comprises the steps that a first network interface receives a data packet sent by a router, wherein the data packet comprises identifiers of a plurality of destination network interfaces;

the first network interface broadcasts the data packet in a first sub-network, wherein the first sub-network comprises the first network interface and a plurality of second network interfaces, the plurality of second network interfaces comprise the plurality of destination network interfaces, the first network interface is a network interface for broadcasting the data packet in the first sub-network, and the second network interface is a network interface for receiving the broadcasted data packet in the first sub-network;

The plurality of destination network interfaces forward the data packet to corresponding processing nodes according to the identifiers of the destination network interfaces in the data packet;

and discarding the data packet according to the identification of the destination network interface in the data packet by other network interfaces except the destination network interfaces in the second network interfaces.

13. The method according to claim 12, wherein: the first network interface broadcasts the data packet into a first subnet, comprising:

The first network interface broadcasts the data packet to the plurality of second network interfaces in the first subnet according to the identifiers of the plurality of destination network interfaces.

14. The method according to claim 12, wherein: the first network interface broadcasts the data packet into a first subnet, comprising:

The first network interface broadcasts the data packet to the plurality of second network interfaces in the first subnet according to the broadcast identification in the data packet.

15. A data transmission method, wherein the data transmission method is applied to a neural network system, a first subnet includes a first network interface and a plurality of second network interfaces, the first network interface is a network interface for broadcasting data packets in the first subnet, and the second network interface is a network interface for receiving the broadcasted data packets in the first subnet, the method comprising:

The second network interface receives a data packet broadcast by the first network interface, wherein the data packet comprises identifiers of a plurality of destination network interfaces;

The second network interface forwards the data packet to a corresponding processing node when determining that the identification of the destination network interface in the data packet comprises the identification of the second network interface; or (b)

And discarding the data packet by the second network interface when the identification of the destination network interface in the data packet is determined not to comprise the identification of the second network interface.

16. A network interface for use in a neural network system, the network interface comprising:

the receiving unit is used for receiving a data packet broadcast by the first network interface, wherein the data packet comprises identifiers of a plurality of destination network interfaces;

A processing unit, configured to forward the data packet to a corresponding processing node when it is determined that the identifier of the destination network interface in the data packet includes the identifier of the second network interface; or discarding the data packet when the identification of the destination network interface in the data packet is determined not to include the identification of the second network interface;

The first subnet comprises the first network interface and a plurality of the second network interfaces, the first network interface is a network interface for broadcasting data packets in the first subnet, and the second network interface is a network interface for receiving the broadcasted data packets in the first subnet.