Graphics processor topology switching device and graphics processor board card

Abstract

The invention discloses a graphics processor topology switching device and a graphics processor board card. The device comprises: a first longitudinal PCIE cable connecting a first central processor with a first port of a first PCIE switcher of a graphics processor; the second longitudinal PCIE cable connected with the second central processing unit and a first port of a second PCIE switcher of the graphics processor; the transverse PCIE optical fiber cable connected with the second port of the first PCIE switcher and the second port of the second PCIE switcher; and a substrate management controller. The substrate management controller is connected to the first longitudinal PCIE cable, the second longitudinal PCIE cable and the transverse PCIE optical fiber cable through an internal integrated circuit bus and is configured to obtain and configure firmware of the first PCIE switcher and the second PCIE switcher based on configuration identifications of the first longitudinal PCIE cable, the second longitudinal PCIE cable and the transverse PCIE optical fiber cable so as to form a graphics processor topology. According to the invention, the node performance under each topology can be improved and the application scene can be widened.

Description

technical field [0001] The present invention relates to the field of artificial intelligence, and more specifically, refers to a graphics processor topology switching device and a graphics processor board. Background technique [0002] With the rapid development of artificial intelligence-related industries, the demand for AI (artificial intelligence) servers represented by GPU (graphics processing unit) servers has increased significantly. Among them, PCIE (Computer and External Equipment Connection Standard Extension) standard card form GPU or AI accelerator cards are widely used because they follow the PCIE specification, have strong applicability, and are relatively cheap. However, for a GPU card in the form of a PCIE standard card, since there is no separate communication link between the GPU card and the GPU card, the communication between the CPU and the GPU and the communication between the GPU and the GPU need to be completed through the PCIE link. [0003] Differe...

AI Technical Summary

Problems solved by technology

However, the use of PCIE MUX will increase the SI (signal-to-noise ratio) loss of the link, resulting in the need to increase signal enhancement components, resulting in cost increases and debugging difficulties

In addition, MUX supports few types of topologies, which do not meet actual needs. The performance of a single topology can only be applied to different application scenarios, and cannot cover all application scenarios.

[0004] There is currently no effective solution to the problems of poor performance and limited use environment of topology switching in the prior art

Images