CN-122018649-A - AI calculates server convenient to heat dissipation of overwind

Abstract

The invention relates to the technical field of AI (advanced technology attachment) calculation, in particular to an AI calculation server convenient for over-wind heat dissipation. The server comprises a case, wherein an installation area for installing acceleration cards and a heat dissipation area for installing a heat dissipation device are sequentially arranged in the case along the length direction of the case, a main air channel is formed between the heat dissipation area and the installation area, the number of the acceleration cards is multiple, the acceleration cards are arranged at intervals along the width direction of the case, a heat dissipation air channel is formed between two adjacent acceleration cards, a distributor plate is arranged at the top of the case, a plurality of power transmission interfaces for connecting with the acceleration cards are arranged on the distributor plate, and two lateral connectors are respectively arranged at two ends of the distributor plate in the length direction and are used for electrically connecting the distributor plate with a power supply plate arranged at the bottom of the case. The PCIE switching flat cable effectively solves the technical problem that in the prior art, PCIE switching flat cables used for connecting an AI computing board card and a main board occupy a heat dissipation channel space, and increase air flow resistance to cause poor heat dissipation of an AI server.

Inventors

• ZHANG FAN
• LIAO HAIYING
• XIE GAOHUI
• LI YUNLONG
• XIE LUOFU
• HUANG GUANGPU
• WEN WEIBIAO

Assignees

• 广州广电五舟科技股份有限公司

Dates

Publication Date

20260512

Application Date

20251226

Claims (10)

1. 1. The AI computing server is characterized in that a distributor plate is arranged at the top of the machine case, a plurality of power transmission interfaces are arranged on the distributor plate and used for being connected with the accelerator cards, the power transmission interfaces and a power supply cable used for connecting the power transmission interfaces with the accelerator cards are spatially isolated from the main air duct and the heat dissipation air duct, and two lateral connectors are respectively arranged at two ends of the distributor plate in the length direction and used for being electrically connected with the distributor plate and a power supply plate arranged at the bottom of the machine case.
2. 2. The AI computing server of claim 1, wherein a bottom slot for mounting the accelerator card is provided at a bottom of the chassis, a baffle is mounted on a side wall of the chassis, the baffle is located at a side of the accelerator card facing away from the heat sink, and the baffle is used for fixing the accelerator card.
3. 3. The AI computing server for facilitating over-wind heat dissipation according to claim 2, wherein a connecting piece is arranged at one end of the accelerator card facing the power transmission interface, and the connecting piece is detachably connected with the distributor.
4. 4. The AI computing server for facilitating over-wind heat dissipation according to claim 3, wherein the connector comprises a vertical connecting portion and a horizontal connecting portion, the vertical connecting portion is connected with the accelerator card, and the horizontal connecting portion is disposed on top of the vertical connecting portion and is detachably connected with the distributor.
5. 5. The AI computing server for facilitating over-wind heat dissipation of any of claims 2-4, wherein a gold finger is provided at a bottom of the accelerator card, the gold finger being inserted into the bottom slot.
6. 6. The AI computing server for facilitating over-wind heat dissipation according to any of claims 1 to 4, wherein the distribution board comprises a conductive layer and a supporting layer sequentially arranged from top to bottom, the power transmission interface is arranged on the conductive layer, a penetrating hole through which the lateral connector passes is arranged on the supporting layer, and the lateral connector passes through the penetrating hole and is connected with the conductive layer.
7. 7. The AI computing server of claim 6, wherein connecting plates are provided at both ends of the support layer in a length direction for connecting the support layer and the chassis.
8. 8. The AI computing server for facilitating over-wind heat dissipation according to any one of claims 1 to 4, wherein the heat dissipation device includes a plurality of heat dissipation fans arranged at intervals in a width direction of the chassis, the plurality of heat dissipation fans forming a heat dissipation fan wall.
9. 9. The AI computing server for facilitating over-wind heat dissipation according to any of claims 1-4, wherein the power transmission interface is connected with at least one power supply female socket, the accelerator card is provided with a power supply male interface, and the power supply female socket and the power supply male interface are on the same horizontal plane.
10. 10. The AI computing server of claim 9, wherein each power transmission interface connects two power supply female stations via the power supply cable.

Description

AI calculates server convenient to heat dissipation of overwind Technical Field The invention relates to the technical field of AI (advanced technology attachment) calculation, in particular to an AI calculation server convenient for over-wind heat dissipation. Background The AI calculation server is a special server for providing high-efficiency calculation support for training, reasoning and data processing of an Artificial Intelligence (AI) model, and has the core of realizing high-efficiency support for training and reasoning of the AI model by matching a high-bandwidth interconnection, high-capacity high-speed storage, high-efficiency heat dissipation system and AI framework adaptation software through a heterogeneous calculation architecture of a CPU and a special acceleration card, and is a core hardware carrier for carrying AI applications such as a large language model, computer vision, voice recognition and the like. At present, patent document with an authorized bulletin number of CN110134205B discloses an AI computing server, which comprises a case, wherein the case is divided into a first area, a second area and a third area by a bracket, a server main board and an image processing board are arranged in the first area, the image processing board is connected with a display card interface of the server main board, a plurality of AI computing boards and a hard disk array which are arranged in a stacked manner are arranged in the third area, the AI computing boards are electrically connected with the server main board by PCIE switching flat cables, the side part of each AI computing board is fixed with the side wall of the server, the AI computing boards further comprise an adapter board and a computing board, the adapter board is provided with an M.2 socket, a bridge chip and a PCIE interface, the computing board is provided with an M.2 plug and an AI chip, the bridge chip are detachably connected with the M.2 socket by the M.2 plug, the bridge chip is electrically connected with the server main board by a data flat cable, the hard disk array is electrically connected with the server main board by the first area, the AI computing board comprises a radiator which is arranged between the first area and the third area, the bracket and the two radiator are arranged on the bracket, and the radiator channels are communicated with the radiator are arranged in the radiator, and the radiator channels are used for realizing the inner radiator channel of the radiator. The prior art can realize the basic heat dissipation function by arranging the special heat dissipation channel and the heat dissipation fan, but has obvious defects in the practical application process, because the PCIE switching flat cable for connecting the AI computing board card and the main board is directly positioned in a heat dissipation airflow path (namely the heat dissipation channel), the physical structure of the PCIE switching flat cable occupies part of the space of the flow channel, and the effective flow cross section area of the heat dissipation channel is reduced. The air flow resistance is directly increased, the uniform flow of cooling air flow is disturbed, and the air supply efficiency of the cooling fan and the heat exchange effect of the surface of the radiator are seriously influenced, so that the heat radiation performance of the whole AI computing server is poor. When the AI computing board card with high power consumption in the server works for a long time under full load, if accumulated heat cannot be taken away effectively in time, the problems of chip overheat and frequency reduction, system stability reduction, even hardware damage and the like are easily caused. Disclosure of Invention The invention provides an AI (advanced technology interface) computing server convenient for over-wind heat dissipation, which aims to solve the technical problem that PCIE switching flat cables used for connecting an AI computing board card and a main board occupy a heat dissipation channel space in the prior art, and increase air flow resistance to cause poor heat dissipation of the AI server. In order to solve the problems, the AI computing server convenient for over-wind heat dissipation provided by the invention adopts the following technical scheme: The AI computing server comprises a case, wherein an installation area for installing accelerator cards and a heat dissipation area for installing a heat dissipation device are sequentially arranged in the case along the length direction of the case, a main air channel is formed between the heat dissipation area and the installation area, the accelerator cards are multiple in number and are arranged at intervals along the width direction of the case, a heat dissipation air channel is formed between every two adjacent accelerator cards, a distributor plate is arranged at the top of the case, a plurality of power transmission interfaces for being connected with the accelerator