GPU box server cascade communication method, device, and system

A communication apparatus for cascaded graphics processing unit (GPU) Box servers, comprising:

• a processor and a memory storing computer instructions,
• wherein the processor, when executing the computer instructions, is configured to:
• detect an inter-integrated circuit (i2c) communication bus through a baseboard management controller (BMC) in a GPU Box server being powered on, to determine whether there is a next-stage GPU Box server corresponding to the GPU Box server;
• read an IP address and location information of the next-stage GPU Box server from the i2c communication bus through the BMC in the GPU Box server, in response to a positive determination;
• store the IP address and the location information of the next-stage GPU Box server into a data structure of the GPU Box server, such that a master control server reads the data structure of the GPU Box server and establishes network communication between the master control server and the next-stage GPU Box server based on the IP address and the location information stored in the data structure of the GPU Box server;
• read an IP address of the GPU Box server from the i2c communication bus through the master control server, and establish network communication between the master control server and the GPU Box server based on the IP address of the GPU Box server;
• read a data structure of the next-stage GPU Box server from the i2c communication bus through the BMC in the GPU Box server, wherein the data structure of the next-stage GPU Box server is stored with an IP address and location information of a later-stage GPU Box server corresponding to the next-stage GPU Box server;
• and
• add the IP address and the location information of the later-stage GPU Box server corresponding to the next-stage GPU Box server stored in the data structure of the next-stage GPU Box server to the data structure of the GPU Box server, such that the master control server establishes network communication between the master control server and each GPU Box server based on each IP address and corresponding location information stored in the data structure of the GPU Box server.