CN111782565A - GPU server and data transfer method - Google Patents

Abstract

The application discloses a GPU server and a data transmission method, and relates to the field of server architecture. The specific implementation scheme is as follows: the method comprises the following steps: the system comprises a GPU, a CPU, a first converter, a second converter and a network interface controller; the GPU is connected with the CPU through the first converter; the GPU is connected with the second converter through the first converter, and the second converter is connected with the network interface controller; the second converter is used for sending the first data received from the network interface controller to the GPU through the first converter and forwarding the second data sent by the GPU to the second converter through the first converter to the network interface controller, so that the performance of the GPU server can be improved, and the method and the system are applied to the fields of high-performance computing, deep learning and the like.

Description

GPU server and data transfer method

technical field

The embodiments of the present application relate to a server architecture in the field of computer technologies, and in particular, to a GPU server and a data transmission method.

Background technique

Graphics Processing Unit (Graphics Processing Unit, GPU for short) has powerful computing capabilities and is widely used in high-performance computing, deep learning and other fields.

Currently, in a GPU server, the GPU is connected to a central processing unit (CPU) through a PICE switch, a network interface controller (NIC) is connected to a CPU, and the GPU communicates with the network interface controller. Requires a PCIE switch and at least one CPU to affect the performance of the GPU server.

SUMMARY OF THE INVENTION

The present application provides a kind of GPU server and data transmission method.

According to an aspect of the present application, a GPU server is provided, including: a GPU, a CPU, a first converter, a second converter, and a network interface controller;

the GPU is connected to the CPU through the first converter;

The GPU is connected to the second converter through the first converter, and the second converter is connected to the network interface controller;

The second converter is configured to send first data received from the network interface controller to the GPU via the first converter, and forward second data received from the GPU to the network interface a controller, wherein the second data is sent by the GPU to the second converter via the first converter.

According to an aspect of the present application, a data transmission method is provided, the method is applied to a GPU server, and the GPU server includes a GPU, a CPU, a first converter, a second converter, a network interface controller, and the GPU The first converter is connected to the CPU, the GPU is connected to the second converter via the first converter, and the second converter is connected to the network interface controller; the method include:

the second converter sends first data received from the network interface controller to the GPU via the first converter;

The second converter forwards second data received from the GPU to the network interface controller, wherein the second data is sent by the GPU to the second converter via the first converter converter.

Techniques in accordance with the present application improve GPU server performance.

It should be understood that the content described in this section is not intended to identify key or critical features of the embodiments of the application, nor is it intended to limit the scope of the application. Other features of the present application will become readily understood from the following description.

Description of drawings

The accompanying drawings are used for better understanding of the present solution, and do not constitute a limitation to the present application. in:

1 is a schematic diagram of a GPU server architecture provided by an embodiment of the present application;

2 is a schematic diagram of another GPU server architecture provided by an embodiment of the present application;

3 is a schematic diagram of another GPU server architecture provided by an embodiment of the present application;

4 is a schematic diagram of another GPU server architecture provided by an embodiment of the present application;

FIG. 5 is a flowchart of a data transmission method provided by an embodiment of the present application.

Detailed ways

Exemplary embodiments of the present application are described below with reference to the accompanying drawings, which include various details of the embodiments of the present application to facilitate understanding, and should be considered as exemplary only. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present application. Also, descriptions of well-known functions and constructions are omitted from the following description for clarity and conciseness.

The present application provides a GPU server, which is applied to the server architecture in the field of computer technology, so as to shorten the communication delay between the GPU and the network interface controller in the GPU server, and provide the technical effect of the performance of the GPU server, which can be applied to High performance computing, deep learning and other fields.

FIG. 1 is a schematic diagram of a GPU server architecture according to an embodiment of the present application. As shown in FIG. 1 , the GPU server includes a GPU 10 , a CPU 11 , a first converter 12 , a second converter 13 , and a network interface controller 14 .

The GPU 10 is connected to the CPU 11 through the first converter 12 . The GPU 10 is connected to the second converter 13 through the first converter 12 , and the second converter 13 is connected to the network interface controller 14 .

The second converter 13 is used to send the first data received from the network interface controller 14 to the GPU 10 via the first converter 12, and to forward the second data received from the GPU 10 to the network interface controller 14, wherein the second Data is sent by the GPU 10 to the second converter 13 via the first converter 12 .

In the embodiments of the present application, in the training tasks of many large-scale machine learning models, for the GPU servers participating in the training tasks, a large part of the data sent by other GPU servers through the network does not need to be processed by the CPU. processing, but can be processed directly by the GPU. In this embodiment of the present application, the data that is transmitted to the network interface controller through the Internet without processing by the CPU and can be directly handed over to the GPU for processing is collectively referred to as the first data.

For the transmission process of the first data, the network interface controller 14 sends the first data to the second converter 13, the second converter 13 sends the received first data to the first converter 12, and the first converter 12 Send the received first data to GPU10. In this embodiment of the present application, the first data received by the network interface controller is sent to the GPU through the second converter and the first converter in sequence, without forwarding by the CPU, and the first data can be shortened while reducing the burden on the CPU. The latency of transmission from the network interface controller to the GPU can improve the overall performance of the GPU server.

In the embodiments of the present application, most of the data processed by the GPU does not need to be processed by the CPU, and can be directly transmitted to an Ethernet or other network to be transmitted to other GPU servers or other devices through the network. In the embodiment of the present application, the data processed by the GPU does not require further processing by the CPU, and the data that can be directly transmitted to the Ethernet or other networks is collectively referred to as second data.

For the transmission process of the second data, the GPU sends the second data to the first converter 12, the first converter 12 sends the received second data to the second converter 13, and the second converter 13 sends the received second data to the second converter 13. The second data is sent to the network interface controller 14 to cause the network interface controller 14 to pass the second data into the Ethernet or other network. In this embodiment of the present application, the second data processed by the GPU is sequentially sent to the network interface controller through the first converter and the second converter, so as to be transmitted to the Ethernet or other networks through the network interface controller, without forwarding by the CPU. , while reducing the burden on the CPU, the delay in transmitting the second data from the GPU to the network interface controller can be shortened, and the overall performance of the GPU server can be improved.

According to the technical solution of the embodiment of the present application, a second converter is added to the GPU server, the GPU is connected to the network interface controller through the first converter and the second converter, and the second converter is responsible for the connection between the network interface controller and the first converter. For data forwarding between converters, data transmission between GPU and network interface controller does not need to go through CPU, which can shorten the delay of data transmission between GPU and network interface controller, and can improve the overall performance of GPU server; further, Data transmission between each GPU can be performed through the first converter and the second converter without going through the CPU, which can achieve high bandwidth and low latency of P2P communication between GPUs, and further improve the overall performance of the GPU server.

On the basis of the above embodiment, in the training task of the machine learning model, the CPU needs to perform the distribution of the training task and the parsing of some data, that is to say, in the actual application scenario, a small part of the data needs to be Transfer between the network interface controller and the CPU, and between the CPU and the GPU. In this embodiment of the present application, the data that needs to be transmitted by the network interface controller to the CPU is collectively referred to as third data. The data to be transmitted from the CPU to the network interface controller is collectively referred to as fourth data.

In the embodiment of the present application, based on the GPU server architecture shown in FIG. 1 , in order to realize data transmission between the CPU and the network interface controller, the second converter is further configured to transmit the third data received from the network interface controller via the first converter. A converter is sent to the CPU.

Specifically, the network interface controller sends the third data to the second converter, the second converter forwards the received third data to the first converter, and the first converter forwards the received data to the CPU, which can Data transmission from the network interface controller to the CPU is implemented based on the GPU server architecture provided by the embodiments of the present application.

The second converter is further configured to forward the received fourth data to the network interface controller, where the fourth data is sent by the CPU via the first converter.

Specifically, the CPU sends the fourth data to the first converter, the first converter forwards the received fourth data to the second converter, and the second converter forwards the received fourth data to the network interface controller , data transmission from the CPU to the network interface controller can be implemented based on the GPU server architecture provided by the embodiments of the present application.

In the embodiments of the present application, the data that needs to be transmitted from the GPU to the CPU is collectively referred to as fifth data. The data that needs to be transferred from the CPU to the GPU is collectively referred to as sixth data.

In the embodiment of the present application, based on the GPU server architecture shown in FIG. 1 , data transmission between the GPU and the CPU is implemented by a first converter. The first converter is used to transmit the fifth data received from the GPU to the CPU, and the sixth data received from the CPU to the GPU.

Specifically, the GPU sends the fifth data to the first converter, and the first converter forwards the received fifth data to the CPU; the CPU sends the sixth data to the first converter, and the first converter forwards the received fifth data to the CPU. The sixth data is forwarded to the GPU, and data transmission between the GPU and the CPU can be implemented based on the GPU server architecture provided in the embodiment of the present application.

In addition, in this embodiment, the number of CPUs in the GPU server may be one or more, which is not specifically limited in this embodiment.

FIG. 2 is a schematic diagram of another GPU server architecture provided by an embodiment of the present application; FIG. 3 is a schematic diagram of another GPU server architecture provided by an embodiment of the present application. On the basis of any of the above embodiments, in this embodiment, the number of CPUs may be multiple, each CPU is connected to at least one first converter, and each first converter is connected to one CPU, so that for each CPU For the CPU, the more first converters between the CPU and the GPU, the more data transmission links between the CPU and the GPU, and the higher the efficiency of data transmission.

In a possible implementation, the GPU server may include a plurality of network interface controllers, and each network interface controller is connected to the second converter, so as to improve the network bandwidth of the GPU. Of course, the GPU server may also include only one network interface controller.

Exemplarily, the GPU server may also include at least one network interface controller connected to the CPU, so that direct data transmission between the network interface controller and the CPU can be implemented, and for some data transmissions that need to be transmitted from the network to the CPU. , which can improve the efficiency and improve the overall performance of the GPU server.

Illustratively, a GPU server may contain a GPU board with multiple GPUs. Each first converter in the GPU server is connected to each GPU on the GPU board. Among them, multiple CPUs can directly perform data communication through QuickPath Interconnect (UPI for short).

In a possible implementation manner, the first converter may be a PCIE switch chip working in a system mode (base mode), and the first converter is connected to the CPU through a PCIEx16 link.

The second converter is a PCIE switch chip operating in a fabric mode (fabric mode), and the second converter is connected to the first converter through a PCIEx16 link.

The PCIEx16 link may be a link of different versions, such as a PCIE4.0x16 link. With the development of the PCIE technology, a higher version of PCIEx16 link may also be used for implementation in this embodiment of the present application, or, according to an actual application scenario, a lower version of PCIEx16 link may also be selected for implementation. Make specific restrictions.

In the embodiment of the present application, both the first converter and the second converter are implemented using a PCIE switch chip, which can improve the stability of data transmission between the first converter and the second converter.

In a possible implementation, the GPU may include a first converter, as shown in FIG. 2 and FIG. 3 .

For example, the GPU server in Figure 2 includes: a GPU board containing multiple GPUs, a second switch (switch2 as shown in Figure 2), two network interface controllers (NIC0 and NIC0 as shown in Figure 2) NIC1), two CPUs (CPU0 and CPU1 as shown in FIG. 2), and each CPU corresponds to a first converter as an example to illustrate the architecture of the GPU server exemplarily. The two CPUs communicate through UPI.

As shown in FIG. 2 , the first converter corresponding to CPU0 is switch0, and the first converter corresponding to CPU1 is switch1. CPU0 and CPU1 are respectively connected to each GPU on the GPU board through switch0 and switch1. Among them, between CPU0 and switch0, and between CPU1 and switch1 are connected through PCIE4.0x16 link. There is a data transmission link between each CPU and GPU board.

As shown in FIG. 2, both network interface controllers NIC0 and NIC1 are connected to the second switch switch2. The second switch switch2 is respectively connected to the first switch switch0 and switch1 through the PCIE4.0x16 link.

In a possible implementation manner, the number of the first converters corresponding to each CPU can be expanded to improve the data transmission efficiency between the CPU and the GPU.

Exemplarily, the number of first converters corresponding to each CPU is equal, so that the first converters can be allocated to each CPU, so that the number of data transmission links between each CPU and the GPU is balanced, which can improve the performance of the GPU server. The balance of the structure is conducive to improving the overall performance of GPU services.

In addition, the number of first converters corresponding to each CPU may also be unequal. For example, the number of first converters corresponding to each CPU can be flexibly adjusted according to the load conditions of each CPU, which is not specifically limited in this implementation example. .

For example, the GPU server in FIG. 3 includes: a GPU board containing multiple GPUs, a second switch (switch2 as shown in FIG. 3 ), n network interface controllers (NIC0 as shown in FIG. 3 , NIC1, . The two CPUs communicate through UPI. where n is a positive integer representing the total number of network interface controllers.

As shown in Figure 3, the first converters corresponding to CPU0 are switch01 and switch02, and the first converters corresponding to CPU1 are switch11 and switch12. CPU0 is connected to each GPU on the GPU board through switch01 and switch02, and CPU1 is connected to the GPU through switch11 and switch12. Each GPU connection on the GPU board. Among them, between CPU0 and switch01 and switch02, and between CPU1 and switch11 and switch12 are connected through PCIE4.0x16 link. There are two data transmission links between each CPU and GPU board. Compared with the GPU server architecture shown in Figure 2, there are more data communication links between each CPU and GPU, and the transmission efficiency is higher.

As shown in FIG. 3, the n network interface controllers NIC0 to NICn are all connected to the second switch switch2. The second switch switch2 is respectively connected with the first switch switch01, switch02, switch11, and switch12 through a PCIE4.0x16 link.

In another implementation manner of the embodiment of the present application, the number of second converters in the GPU server may be multiple, so as to improve the bandwidth of data transmission between the network interface controller and the GPU, and to improve the network bandwidth of the GPU server.

Wherein, each first converter is respectively connected with each second converter; each network interface controller is connected with one second converter, and the multiple network interface controllers are evenly distributed to the multiple second converters, so that the GPU The network balance of the server is better, which can improve the overall performance of the GPU server.

In the technical solution of the embodiment of the present application, the GPU server may include a plurality of network interface controllers connected to the second converters, so as to improve the network bandwidth of the GPU; further, the number of the first converters corresponding to each CPU can be expanded, And the number of first converters corresponding to each CPU is equal, which can improve the data transmission efficiency between CPU and GPU, and make the number of data transmission links between each CPU and GPU balanced, which can improve the overall performance of GPU services. ; Further, by expanding the number of second converters, the bandwidth of data transmission between the network interface controller and the GPU can be increased, the network bandwidth of the GPU server can be increased, and the overall performance of the GPU service can be further improved.

FIG. 4 is a schematic diagram of another GPU server architecture provided by an embodiment of the present application. Based on any of the foregoing embodiments, in this embodiment of the present application, the GPU server further includes at least one third converter. The GPU is connected with at least one third converter, each third converter is connected with each second converter, and at least one third converter is not connected with the CPU.

The second converter is further configured to send data received from the network interface controller to the GPU via the third converter, and to forward data sent by the GPU to the second converter via the third converter to the network interface controller.

In the embodiment of the present application, by adding a third converter between the second converter and the GPU, and sharing the data transmission task of the first converter with the third converter, the data transmission between the second converter and the GPU can be increased. The number of links can increase the network bandwidth of the GPU and improve the overall performance of the GPU.

Specifically, for the first data that needs to be received by the first converter from the second converter, the second converter may send part of the data in the first data to the third converter, and then send the data to the GPU via the third converter.

For the second data that the first converter needs to receive from the GPU, the GPU may send part of the second data to the third converter, and the third converter sends the second data to the second converter.

For example, in FIG. 4 , the architecture of the GPU server is exemplarily illustrated by adding two third converters on the basis of the GPU server architecture provided in FIG. 3 as an example. As shown in FIG. 4 , two third converters switch31 and switch32 are added on the basis of the GPU server architecture provided in FIG. 3 . Wherein, switch31 and switch32 are respectively connected with each second converter switch2, and switch31 and switch32 are respectively connected with each GPU on the GPU board. As shown in FIG. 4 , the third converters switch31 and switch32 are not connected to each CPU, and are exclusively used for data transmission between the second converter switch2 and each GPU on the GPU board.

In this embodiment of the present application, the number of third converters may be an integer multiple of the number of CPUs, and the number of third converters corresponding to each CPU is equal, so that the added third converters can be evenly distributed to each CPU for use in Processing part of the data of the first data and part of the data of the second data forwarded by each first converter connected to the corresponding CPU can improve the balance of the network bandwidth of the GPU server and further improve the overall performance of the GPU server.

Exemplarily, in another implementation manner, the total number of third converters may not be an integer multiple of the number of CPUs, and the number of third converters corresponding to each CPU may be unequal, for example, the number of third converters corresponding to each CPU The number of CPUs can be flexibly adjusted according to the load of each CPU, which is not specifically limited in this embodiment.

In a possible implementation manner, the third converter may be a PCIE switch chip operating in a system mode (base mode), and the third converter and the second converter are connected through a PCIEx16 link. In the embodiment of the present application, both the third converter and the second converter are implemented using a PCIE switch chip, which can improve the stability of data transmission between the third converter and the second converter.

The PCIEx16 link may be a link of different versions, such as a PCIE4.0x16 link. With the development of the PCIE technology, a higher version of PCIEx16 link may also be used for implementation in this embodiment of the present application, or, according to an actual application scenario, a lower version of PCIEx16 link may also be selected for implementation. Make specific restrictions.

For example, as shown in Figure 4, two third switches (switch30 and switch31 as shown in Figure 4) can be added between the second switch switch2 and the GPU, and each CPU corresponds to one third switch , wherein the third converter corresponding to CPU0 is switch30, and the third converter corresponding to CPU1 is switch31.

As shown in Figure 4, the third switch corresponding to CPU0 is switch30, the first switch connected to CPU0 includes switch01 and switch02, and the second switch switch2 needs to pass through switch01 and switch02 from the first data received from the network interface controller. Part of the data forwarded by switch02 to the GPU is sent to the third converter switch30, and the third converter switch30 forwards the received first data to the GPU; the GPU needs to forward the second data to the second converter switch2 via switch01 and switch02 in the second data Part of the data received is sent to the third switch switch30, and the third switch switch30 forwards the received second data to the second switch switch2.

As shown in Figure 4, the third switch corresponding to CPU1 is switch31, the first switch connected to CPU1 includes switch11 and switch12, and the second switch switch2 needs to pass through switch11 and switch12 from the first data received from the network interface controller. Part of the data forwarded by switch12 to the GPU is sent to the third converter switch31, and the third converter switch31 forwards the received first data to the GPU; the GPU needs to forward the second data to the second converter switch2 via switch11 and switch12 in the second data Part of the data received is sent to the third switch switch31, and the third switch switch31 forwards the received second data to the second switch switch2.

As shown in FIG. 4 , a PCIE4.0x16 link is used between the third switches switch30 and switch31 and the second switch switch2.

In addition, in this embodiment, the first converter, the second converter and the third converter may also be implemented by a chip similar to the PCIEswitch chip, which is not specifically limited in this embodiment. The first converter, the second converter, and the third converter may also be implemented by different types of chips, which are not specifically limited in this embodiment.

In the technical solutions of the embodiments of the present application, by adding at least one third converter to the GPU server, the GPU is connected to at least one third converter, each third converter is connected to each second converter, and at least one third converter The converter is not connected to the CPU; using the third converter to share the data transmission task of the first converter can increase the number of data transmission links between the second converter and the GPU, increase the network bandwidth of the GPU, and improve the GPU performance. the overall performance; further, the number of third converters can be an integer multiple of the number of CPUs, and the number of third converters corresponding to each CPU is equal, so that the added third converters can be evenly distributed to each CPU. For processing part of the data of the first data and part of the data of the second data forwarded by each first converter connected to the corresponding CPU, the balance of the network bandwidth of the GPU server can be improved, and the overall performance of the GPU server can be further improved.

The present application provides a data transmission method. The method is applied to a GPU server. The GPU server includes a GPU, a CPU, a first converter, a second converter, and a network interface controller. The GPU is connected to the CPU through the first converter, and the GPU is connected to the CPU through the first converter. A switch is connected to the second switch, and the second switch is connected to the network interface controller. FIG. 5 is a flowchart of a data transmission method provided by an embodiment of the present application. As shown in Figure 5, the method includes the following steps:

S101. The second converter sends the first data received from the network interface controller to the GPU via the first converter.

S102. The second converter forwards the second data received from the GPU to the network interface controller, where the second data is sent by the GPU to the second converter via the first converter.

In this embodiment, the steps of the data transmission method performed by the second converter may be specifically referred to in the above-mentioned embodiments, and details are not described herein again in this embodiment.

According to the technical solution of the embodiment of the present application, a second converter is added to the GPU server, the GPU is connected to the network interface controller through the first converter and the second converter, and the second converter is responsible for the connection between the network interface controller and the first converter. For data forwarding between converters, data transmission between GPU and network interface controller does not need to go through CPU, which can shorten the delay of data transmission between GPU and network interface controller, and can improve the overall performance of GPU server; further, Data transmission between each GPU can be performed through the first converter and the second converter without going through the CPU, which can achieve high bandwidth and low latency of P2P communication between GPUs, and further improve the overall performance of the GPU server.

According to the technical solutions of the embodiments of the present application, the GPU server architecture adopts the system mode + mesh connection mode of PCIE switch, and uses a plurality of first converters working in the system mode to communicate with the CPU through the PCIEx16 link, respectively. A converter is connected to each GPU on the GPU board, and these first converters are equally distributed to each CPU to realize data transmission between GPU and CPU; through multiple second converters working in mesh connection mode Connected with a plurality of network interface controllers, the plurality of network interface controllers are equally distributed to each second converter, the second converter is also connected with each first converter, and these second converters are used to realize the network interface controller and the Data transmission between the first converters, so that data can be dynamically and non-blockingly transmitted between the network interface controller and the GPU, and data transmission between each GPU can be performed through the first converter and the second converter without going through the CPU. , which can achieve high bandwidth and low latency of P2P communication between GPUs, which can further improve the overall performance of the GPU server.

It should be understood that steps may be reordered, added or deleted using the various forms of flow shown above. For example, the steps described in the present application can be performed in parallel, sequentially or in different orders, and as long as the desired results of the technical solutions disclosed in the present application can be achieved, no limitation is imposed herein.

The above-mentioned specific embodiments do not constitute a limitation on the protection scope of the present application. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modifications, equivalent replacements and improvements made within the spirit and principles of this application shall be included within the protection scope of this application.

Claims (14)

1. A graphics processor GPU server, comprising: a GPU, a CPU, a first converter, a second converter, and a network interface controller; the GPU is connected to the CPU through the first converter; The GPU is connected to the second converter through the first converter, and the second converter is connected to the network interface controller; The second converter is configured to send first data received from the network interface controller to the GPU via the first converter, and forward second data received from the GPU to the network interface a controller, wherein the second data is sent by the GPU to the second converter via the first converter. 2. The GPU server of claim 1, wherein the second converter is further configured to send third data received from the network interface controller to the CPU via the first converter; The second converter is further configured to forward the received fourth data to the network interface controller, where the fourth data is sent by the CPU via the first converter. 3. The GPU server of claim 2, wherein the first converter is configured to transmit fifth data received from the GPU to the CPU and sixth data received from the CPU to the CPU GPU. 4. The GPU server according to claim 1, wherein the number of the CPUs is multiple, each of the CPUs is connected to at least one of the first converters, and each of the first converters is connected to one of the first converters. CPU connection. 5. The GPU server of claim 4, wherein the number of the first converters connected to each of the CPUs is equal. 6 . The GPU server according to claim 1 , wherein the number of the network interface controllers is plural, and each of the network interface controllers is connected to the second converter. 7 . 7. The GPU server according to claim 6, wherein the number of the second converters is multiple, each of the first converters is connected to each of the second converters, respectively; Each of the network interface controllers is connected to one of the second switches, and a plurality of the network interface controllers are equally distributed to a plurality of the second switches. 8 . The GPU server according to claim 1 , wherein the first converter is a PCIEswitch chip operating in a system mode, and the first converter is connected to the CPU through a PCIE x16 link. 9 . 9 . The GPU server according to claim 8 , wherein the second converter is a PCIE switch chip operating in a mesh connection mode, and the second converter is connected to the first converter through a PCIE x16 link. 10 . 10. The GPU server according to any one of claims 1-9, further comprising: at least one third converter, the GPU is connected to the at least one third converter, each of the third converters is connected to each of the second converters, and the at least one third converter is not connected to the CPU; The second converter is further configured to send data received from the network interface controller to the GPU via the third converter, and to send the GPU to the third converter via the third converter. The data of the two converters are forwarded to the network interface controller. 11. The GPU server according to claim 10, wherein the number of the third converters is an integer multiple of the number of the CPUs, and the number of the third converters corresponding to each of the CPUs is equal. 12 . The GPU server according to claim 10 , wherein the third converter is a PCIE switch chip operating in a system mode, and the second converter and the third converter are connected through a PCIE x16 link. 13 . 13. The GPU server according to any one of claims 1-9, further comprising: a network interface controller connected to the CPU. 14. A data transmission method, the method is applied to a GPU server, the GPU server comprises a GPU, a CPU, a first converter, a second converter, and a network interface controller, and the GPU passes through the first converter connected with the CPU, the GPU is connected with the second converter through the first converter, and the second converter is connected with the network interface controller; the method includes: the second converter sends first data received from the network interface controller to the GPU via the first converter; The second converter forwards second data received from the GPU to the network interface controller, wherein the second data is sent by the GPU to the second converter via the first converter converter.

Images