A server structure supporting a 16GPU

Abstract

The invention discloses a server structure supporting a 16GPU. The device comprises an 8U case, a 4U server, a bridging module, bolts and screws, The 8U case is provided with a front window structureand a rear window structure. The device comprises a base, Upper cover and middle partition plate, Rear window perforated plate, Rear window support, The invention designs a 4U server structure capableof accommodating eight GPUs. Meanwhile, an 8U case capable of storing two 4U servers is designed. The invention discloses a server structure supporting a 16GPU. GPU high-density ports of the two 4U servers are connected through the bridging module; the two 4U servers are arranged; Sixteen GPU cooperative work, the space requirement of the novel server is met; meanwhile, due to the special structural design of the wind scooper, the heat dissipation performance of the server is improved; the rear window opening plate is connected with the upper cover and the base through bolts at the rear window, so that connection between the upper cover and the base is reinforced, heat dissipation of the server can be guaranteed, and technicians can conveniently observe whether the server has abnormal conditions or not at the rear window.

Description

technical field [0001] The invention relates to the technical field of servers, in particular to a server structure supporting 16 GPUs. Background technique [0002] With the rapid development of AI (artificial intelligence), the application of artificial intelligence technologies such as image recognition and speech recognition has been widely popularized. As the core computing power providing equipment, GPU server is facing a major test in its structure and performance. [0003] The GPU servers in the prior art are mainly 4GPU and 8GPU, and the chassis adopts 2U and 4U structural designs. The HGX-2 super computing platform launched by NVIDIA uses sixteen Tesla V100 GPUs and twelve NVSwitches to Make each GPU communicate with other GPUs at a speed of 2.4TB per second. While the performance of this type of server is powerful, the requirements for heat dissipation and space are also increased. The traditional 2U or 4U chassis can no longer meet the space requirements of new s...

AI Technical Summary

Problems solved by technology

[0003] The GPU servers in the prior art are mainly 4GPU and 8GPU, and the chassis adopts 2U and 4U structural designs. The HGX-2 super computing platform launched by NVIDIA uses sixteen Tesla V100 GPUs and twelve NVSwitches to Make each GPU communicate with other GPUs at a speed of 2.4TB per second. While the performance of this type of server is powerful, the requirements for heat dissipation and space are also increased. The traditional 2U or 4U chassis can no longer meet the space requirements of new servers , and cannot meet the heat dissipation requirements of high-performance servers, there are great limitations in the structure, which will lead to many problems such as poor server integration, poor heat dissipation, and shortened service life. Therefore, it is urgent to develop a new Server structure to meet the above requirements

Images