CN117076376A

CN117076376A - GPU server architecture based on OCSP mainboard realization

Info

Publication number: CN117076376A
Application number: CN202311336553.XA
Authority: CN
Inventors: 万大炎; 俞跃渊; 杨立; 张少鹏
Original assignee: Enginetech Tianjin computer Co ltd
Current assignee: Enginetech Tianjin computer Co ltd
Priority date: 2023-10-16
Filing date: 2023-10-16
Publication date: 2023-11-17
Anticipated expiration: 2043-10-16
Also published as: CN117076376B

Abstract

The application relates to a GPU server architecture realized based on an OCSP motherboard, which solves the problem of power supply connection between a power supply module and a complete machine by stacking the OCSP motherboard and a GPU expansion board and respectively arranging two horizontally placed and vertically placed power supply backplates; the three power supply circuits are provided for supplying power to the GPU expansion board, so that the flexible power supply requirements of 10GPU large cards and 10 pcie slots are met; by inversely installing the 4 power modules, the partial overlapping of the power connector, the power backboard and the OCSP main board is realized, and the length of the case is saved; through the hollowing out to the power backplate and the setting of table rubberizing copper bar, when guaranteeing power module radiating effect, ensured the overcurrent capacity of power backplate. The application realizes the development of products with low cost and high efficiency by multiplexing the components such as the 2U-height OCSP board type existing structural component, the cooling fan, the power module, the IO board and the like and only customizing the power backboard and the signal cable on the premise of ensuring that the length of the whole machine does not exceed the standard.

Description

GPU server architecture based on OCSP mainboard realization

Technical Field

The application relates to the field of architecture design of server systems, in particular to a GPU server architecture realized based on an OCSP motherboard.

Background

Currently, with the continuous improvement of informatization level, the gradual improvement and breakthrough of computer computing power has prompted the explosion-type prosperity of Artificial Intelligence (AI) related technologies and markets. Thus, the explosive growth of data is presented in various fields, and the requirements for data acquisition, integration, transmission, processing, storage and the like are also increasing.

GPU server, or artificial intelligence server, as a new class in server in recent years, mainly refer to the server bearing a plurality of high-power GPU cards (or modules), which well satisfies the current demands for high data and high power. The structure of the device adopts a 19-inch standard rack type, the height is not less than 4U, and the two typical modes are a 10GPU large card type with the height of 4U and an 8 OAM module type with the height of 6U respectively.

The main board is used as a core component of the whole equipment, and the performance and the stability of the whole server are related. OCSP is an abbreviation of intel-dominant open universal server platform, defines the decoupling standard of open chassis, power supply, main board, hard disk, fan and expansion module, and aims to construct a good server component ecosystem, reduce research and development repetition investment, and further improve market competitiveness of system manufacturers in small-batch customization, quality and cost. Because the system is a general server motherboard specification of Intel main pushing at present, a plurality of motherboard manufacturers are added in the system, and the motherboard selection surface is large.

However, the conventional common method for developing a 4u 10GPU large card model by using an OCSP board type motherboard is limited in CPU performance due to heat dissipation problem, limited in GPU card number due to heat dissipation problem, or greatly improved in development cost due to the need of large-scale re-mold opening.

Therefore, it is particularly necessary to develop a 10GPU large card model with a 4U height more economically and efficiently using the existing OCSP board type motherboard.

Disclosure of Invention

In order to solve various technical problems of developing a 4U 10GPU large card model by using an OCSP board type mainboard, the application provides a GPU server architecture realized based on an OCSP board, which comprises the following steps: OCSP motherboard, GPU expansion board, first power back board, second power back board, power module, IO board, heat dissipation module, memory module;

the power supply module comprises four power supply modules, and the four power supply modules and the IO board are both positioned on a rear window of the chassis; the heat dissipation module and the storage module are sequentially arranged on one side of the OCSP mainboard, which faces the front window of the case; a first expansion card connector, an IO board connector and a second expansion card connector are arranged on one side of the OCSP mainboard facing the rear window of the case;

the OCSP motherboard is positioned on a first plane, and in the first plane, the first power backboard is arranged next to the second expansion card connector in a second direction perpendicular to a first direction formed from the front window to the rear window and far away from the second expansion card connector;

in the first direction, arranging the second power back plate close to the first expansion card connector, the IO board connector, the second expansion card connector and the first power back plate;

one side of the first power backboard is electrically connected with the OCSP main board through a first power connector, the other side of the first power backboard is electrically connected with one side of the second power backboard through a second power connector, and the other side of the second power backboard is electrically connected with the power module through a third power connector.

In some embodiments, the GPU expansion board is located in a second plane parallel to the first plane above the OCSP motherboard;

and, the orthographic projection of the GPU expansion board on the first plane covers a part of the OCSP motherboard, and all of the first power backboard, the second power backboard, the power module and the IO board.

In some embodiments, the first power backplane is disposed within the first plane;

the second power back plate is arranged in a third plane perpendicular to the first plane and extending along the second direction.

In some embodiments, a first power supply path is formed by the power module, the second power supply backboard and the first power supply backboard, and a first power supply is provided for the GPU expansion board;

a second power supply path is formed by the power supply module, the second power supply backboard, the first power supply backboard and the OCSP mainboard, and a second power supply is provided for the GPU expansion board;

and a third power supply path is formed by the power supply module and the second power supply backboard, so that a third power supply is provided for the GPU expansion board.

In some embodiments, in the first power supply path, two groups of copper pillars are welded vertically on the first power supply backboard, and the GPU expansion board is inserted in a pluggable manner to supply power.

In some embodiments, the four power modules of the power module are all installed upside down and electrically connected to an end of the second power back plate remote from the OCSP motherboard through the third power connector.

In some embodiments, the orthographic projection of the third power connector and the second power backplane on the first plane partially overlaps the OCSP motherboard.

In some embodiments, the end of the second power backboard adjacent to the OCSP motherboard includes a first area and a second area, and the second area is hollowed out.

In some embodiments, in the third plane, the copper bar is disposed adjacent to the second area in a third direction perpendicular to the first direction and the second direction and away from the OCSP motherboard;

and the surface copper bar is positioned below the third power connector, and the length of the surface copper bar is the same as that of the second area in the second direction.

In some embodiments, a first space is provided between the first expansion card connector, the IO board connector, and the power module, and the cable terminals of the first expansion card connector, the IO board connector are disposed within the first space.

According to the application, a 2U-height OCSP board type main board is used for developing a 4U-height 10GPU large card type server, and two power supply backboard which are horizontally placed and vertically placed are respectively arranged by stacking an OCSP main board and a GPU expansion board, so that the problem of power supply connection between a power supply module and a complete machine is solved; the three power supply circuits are provided for supplying power to the GPU expansion board, so that the flexible power supply requirements of 10GPU large cards and 10 pcie slots are met; by inversely installing the 4 power modules, the partial overlapping of the power connector, the power backboard and the OCSP main board is realized, so that the length of the case is saved; through the hollowing out to the power backplate and the setting of table rubberizing copper bar, when guaranteeing power module radiating effect, ensured the overcurrent capacity of power backplate. In general, the application realizes the development of products with low cost and high efficiency by multiplexing the components such as the 2U-height OCSP board type existing structural component, the cooling fan, the power module, the IO board and the like and only customizing the power backboard and the signal cable on the premise of ensuring that the length of the whole machine does not exceed the standard.

Drawings

FIG. 1 is a 4U height 10GPU large card type GPU server architecture based on a T-shaped motherboard in the prior art;

FIG. 2 is a schematic diagram of a GPU server architecture of an OCSP board motherboard with 2U height in the prior art;

FIG. 3 is a 4U-height 10GPU large card type GPU server architecture based on an OCSP board type motherboard according to the present application;

FIG. 4 is a schematic diagram of three power paths for powering a GPU expansion board in accordance with the present application;

FIG. 5 is a side view of a power module, a second power back plate, and an OCSP motherboard contact location according to the present application;

fig. 6 is a schematic diagram of a hollowed-out area of a second power backboard and related arrangement of surface copper bars according to the present application.

Reference numerals illustrate: a first power connector 1, a second power connector 2, a third power connector 3, a first expansion card connector 4, an IO board connector 5, and a second expansion card connector 6.

Detailed Description

The techniques described below are susceptible to various modifications and alternative embodiments, and are described in detail herein with reference to the accompanying drawings. However, this is not meant to limit the techniques described below to particular embodiments. It should be understood that the application includes all similar modifications, equivalents and alternatives falling within the spirit and scope of the techniques described below.

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used in this specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used herein refers to and encompasses any or all possible combinations of one or more of the associated listed items.

It should be understood that although the terms first, second, third, etc. may be used herein to describe various information, these information should not be limited by these terms. These terms are only used to distinguish one type of information from another. For example, first information may also be referred to as second information, and similarly, second information may also be referred to as first information, without departing from the scope of the application. The word "if" as used herein may be interpreted as "at … …" or "when … …", depending on the context.

A typical 4U-height 10GPU large card model GPU server architecture includes: motherboard, memory module, OAM module, backplate, power module, pcie expansion board etc..

As shown in fig. 1, the ultra-micro 10GPU large card type GPU server with 4U height adopts a T-shaped motherboard, the motherboard originally supports connection of 4 power modules, the 4 power modules can be conveniently placed at the lower 1U height of the rear window, and GPU expansion boards are stacked above the power modules, so that 10GPU large cards are arranged in parallel at the upper 3U height of the rear window, and the transverse width of the whole 19 inch chassis is occupied.

Since OCSP is a general server motherboard specification currently being pushed by Intel, OCSP board type motherboards with 2U height have been widely used by a number of motherboard manufacturers, and the typical architecture is shown in FIG. 2.

At present, three common schemes for developing a 10GPU large card type with a 4U height by using an OCSP board type mainboard with a 2U height are as follows:

(1) The 10GPU large cards are arranged in the middle, and 4 power modules are stacked on the right side of the rear window

According to the scheme, the 10GPU card is placed in the middle of the case and above the CPU, so that the 10GPU card and 4 power modules are prevented from overlapping in a longitudinal area. The advantage of doing so is can be with two liang stacks of power module, sacrifice the height that power module occupy, and every power module height is 1U promptly, and this kind stacks 4 power module of mode and needs to occupy 2U height altogether to change the little increase of complete machine length.

A customized power panel assembly is added between every two power modules stacked and the original 2 power interfaces of the main board, and the length is about 60mm, so that the length of the whole 4U 10GPU card type is about 843mm.

The main disadvantage of this design is that the 10GPU card module is stacked above the CPU, so that the CPU cannot use a high-power radiator, the structure and the air duct are complex in design, and therefore cannot use a high-performance and high-power CPU model.

(2) The rear-mounted 8 GPU large cards and 4 power modules are stacked and placed up and down on the right side of the rear window

According to the scheme, 4 power modules are all stacked to occupy the whole 4U height space on the right side of a rear window, 8 GPU large cards are placed in the rear window space, and the rest 2 GPU large cards are placed in the upper 2U space of a front window, namely, the upper side of a storage module.

The GPU card is still placed at the rear window position in the mode, the limitation on CPU power consumption is avoided, the length of the case is relatively controllable, and the total increase is only about 60-80 mm.

The disadvantage is that, because the right power module occupies the width, the rear window can only place 8 GPU cards side by side, and the rest 2 GPU cards need to be placed in the front window, so that the heat dissipation of the 2 GPU cards is very unfavorable, and GPU cards with higher heat dissipation requirements can not be supported basically. So this solution is only suitable for the scenario supporting a maximum of 8 GPU large cards.

(3) The storage module of the front window moves to the upper 2U position, and the OCSP mainboard moves to the front window

The scheme is to directly place 10GPU large cards side by side in a rear window, and the length of a case is controllable, so that improvement is performed on a front window.

Because the 2U space on the front window of most 4U 10GPU large card models at present is empty, only one honeycomb-shaped baffle is used for shielding, the functions of the upper 2U and the lower 2U of the front window are exchanged, the storage module is placed on the upper 2U, the space of the lower 2U is released, and the heat dissipation module can be moved towards the front window, namely the heat dissipation module and the storage module are stacked up and down, so that an OCSP board type mainboard and a power module are driven to move towards the front window, and the length of the whole chassis is shortened.

This can greatly optimize the chassis length, but has the disadvantage: because the heat dissipation module and part of the main board are positioned below the storage module, the heat dissipation module loses the possibility of convenient replacement, and if the whole structure of the storage module is required to be changed into flexible disassembly for the convenience of replacement, the structural part is required to be opened again, the development cost and the unit price cost are both greatly improved, the OCSP 2U assembly cannot be reused, and the overall development requirement of 4U 10GPU large card type cost priority is not met.

The characteristics of the three common schemes are shown in the following table.

Based on the original OCSP 2U model, the application increases the height of the chassis to 4U by a series of structural and electrical improvements, increases the power supply modules of the chassis to 4, controls the length of the chassis to be within 900mm, meets the space and power supply required by 10GPU large cards and the requirements of the cabinet installation depth, and simultaneously reserves the functions of 2 expansion card interfaces and 1 IO board interface at the rear part of the OCSP motherboard.

FIG. 3 is a schematic diagram of a 4U-height 10GPU large card type GPU server architecture based on an OCSP board type motherboard according to the present application.

Referring to fig. 3, the GPU server architecture implemented based on the OCSP motherboard includes: OCSP mainboard, GPU expansion board, first power backplate, second power backplate, power module, IO board, heat dissipation module, storage module.

The power supply module comprises four power supply modules, and the four power supply modules and the IO board are positioned on the rear window of the case; a heat dissipation module and a storage module are sequentially arranged on one side of the OCSP mainboard facing the front window of the chassis; a first expansion card connector 4, an IO board connector 5 and a second expansion card connector 6 are disposed on the side of the OCSP motherboard facing the rear window of the chassis.

Further, the OCSP motherboard is located in a first plane, in which the first power backplane is disposed adjacent to the second expansion card connector 6 in a second direction perpendicular to a first direction from the front window to the rear window and away from the second expansion card connector 6.

In the first direction, a second power back plate is provided close to the first expansion card connector 4, the IO board connector 5, the second expansion card connector 6, and the first power back plate.

That is, in the first direction from the front window to the rear window of the chassis, the storage module, the heat dissipation module, the OCSP motherboard, the first expansion card connector 4, the IO board connector 5, the second expansion card connector 6, and the first power back plane, the second power back plane, the power module, and the IO board are sequentially disposed.

One side of the first power backboard is electrically connected with the OCSP mainboard through the first power connector 1, the other side of the first power backboard is electrically connected with one side of the second power backboard through the second power connector 2, and the other side of the second power backboard is electrically connected with the power module through the third power connector 3.

That is, the first power back plate is not directly electrically connected to the power module, but is indirectly electrically connected to the power module through the second power back plate. In addition, four power supply modules inserted into the rear of the server are connected with the second power supply backboard in a golden finger mode, and an IO board at the rear of the server is connected with an IO interface of the main board by using a signal cable.

The GPU expansion board is positioned in a second plane parallel to the first plane above the OCSP motherboard. And the orthographic projection of the GPU expansion board on the first plane covers a part of the OCSP motherboard and all of the first power backboard, the second power backboard, the power module and the IO board.

In short, the GPU expansion board is stacked above the OCSP motherboard, the first power backboard, the second power backboard, the IO board and the power module.

The first power backboard is arranged in a first plane; the second power back plate is arranged in a third plane perpendicular to the first plane and extending along the second direction.

In order to expand the 2U OCSP model which only supports two power modules originally into a 4U model which supports four power modules, the application is provided with two power backplates, namely a first power backplate and a second power backplate, as shown in figure 3, wherein the first power backplate is a board card which is horizontally arranged with the OCSP motherboard, and is connected with the OCSP motherboard towards the front window side of a case through two power interfaces of a first power connector 1, and is connected with the second power backplate towards the rear window side of the case through two power interfaces of a second power connector 2; the second power backboard is a board card which is vertically arranged with the OCSP motherboard, is connected with the first power backboard towards the front window direction of the case, and is respectively connected with the four power modules towards the rear window direction of the case through the four power interfaces of the third power connector 3.

So designed, the portion of the entire 4U 10GPU large card server that is increased in length over the OCSP 2U server is simply the longitudinal length of the first power backplate (about 80 mm) and the thickness of the second power backplate (no more than 20 mm), thereby controlling the total length of the 4U chassis to be within 883 mm.

In order to organically integrate the four power supply modules into the whole power supply system of the 4U 10GPU large card server, three power supply paths are arranged.

The first power supply path is formed by the power supply module, the second power supply backboard and the first power supply backboard, and a first power supply is provided for the GPU expansion board;

secondly, a second power supply path is formed by the power supply module, the second power supply backboard, the first power supply backboard and the OCSP mainboard, and a second power supply is provided for the GPU expansion board;

and thirdly, forming a third power supply path through the power supply module and the second power supply backboard, and providing a third power supply for the GPU expansion board.

The second power backboard is used as a board card which is only in contact with the four power modules in the whole machine and is responsible for acquiring power supply from the four power modules.

Fig. 4 is a schematic diagram of three power paths for supplying power to the GPU expansion board in the present application. The whole machine is internally provided with two main power utilization modules, namely a GPU expansion board and an OCSP motherboard, and a heat dissipation module and a storage module are powered from the OCSP motherboard. Wherein in three power-taking paths of the GPU expansion board shown in fig. 4, a represents a path of power supply directly from the first power back plate, and the path of power supply supplies power to the GPU expansion board by two groups of copper columns vertically welded on the first power back plate; b represents one path of power supply directly from an OCSP main board, 4 12pin power plugs supply power to the GPU expansion board in a power supply cable mode, and each 12pin power plug can provide 450W power to meet the requirements of 4 GPU large cards; c represents the power supply of a path directly from the second power backboard, 6 12pin power plugs supply power to the GPU expansion board in a power supply cable mode, and each 12pin power plug can provide 450W power to meet the requirements of 6 GPU large cards. To sum up, the a-way power supply is used for the conventional power supply requirement (75W for each slot) of 10 pc ie slots of the GPU expansion card, and the B-way and C-way power supplies together support the excess power supply requirement of 10GPU large cards.

Wherein, two sets of copper post of vertical welding on first power backplate in the first power supply route, every group has a 12V power copper post and a GND copper post. And the power supply copper column can be inserted into a power supply jack on the GPU expansion board in a pluggable manner, so that the GPU expansion board and the first power supply backboard can be flexibly separated.

When the power supply module is installed, the four power supply modules of the power supply module are all installed reversely and are electrically connected with one end, far away from the OCSP main board, of the second power supply backboard through the third power supply connector 3.

FIG. 5 is a side view of the power module, the second power back plate, and the OCSP motherboard contact locations.

Because the power supply modules all provide power for the whole machine through the golden finger positioned below the power supply modules, the golden finger positioned below the power supply modules is positioned on a similar horizontal plane with the main board, in order to reduce the length of the chassis as much as possible, the four power supply modules positioned at the rear window of the chassis are vertically and overturned and installed by adopting the design thought of staggered installation, and therefore the golden finger of the power supply module is higher than the main board.

Further, the orthographic projection of the third power connector 3 and the second power back plate on the first plane is partially overlapped with the OCSP motherboard.

That is, by mounting the power module upside down, a partial overlap of the power module, the second power back plate and the OCSP main plate is achieved, thereby reducing the length of the cabinet by about 15-20mm.

As described above, for the second power back plate, the vertical height of the second power back plate needs to be sacrificed appropriately for the purpose of overlapping the front projection of the second power back plate on the first plane with the edge portion of the OCSP main board at the connection with the power module, and is generally smaller than the height of the power module, i.e., 1U. At the same time, a part of the area is hollowed out below the second power supply backboard so as to ventilate and dissipate heat of the power supply module.

Therefore, in the present application, the second power backboard is disposed at one end close to the OCSP motherboard, and includes a first area and a second area, and the second area is hollowed out.

However, this hollowing out process can result in further restrictions on the size of the PCB of the second power back plate, thereby affecting its function as a power back plate for the transfer of large currents required.

In order to solve the problem, in the third plane, the surface copper bars are arranged in the position, which is perpendicular to the first direction and the second direction and is far away from the third direction formed by the OCSP main board, and is close to the second area for hollowing.

Specifically, the surface copper bar is located below the third power connector 3, above the second region, and has the same length as the second region in the second direction.

Fig. 6 shows the relative placement of the hollowed out area and the surface mount copper bars of the second power back plate.

Through the arrangement, the overcurrent capacity of the second power supply backboard can reach 80A, and the problem of insufficient overcurrent capacity caused by PCB stenosis can be effectively solved.

In order to preserve the signal functions of the first expansion card connector 4, the IO board connector 5 and the second expansion card connector 6 on the OCSP motherboard, a space of about 15mm is reserved between the power module and the first expansion card connector 4 and the IO board connector 5 of the OCSP motherboard, so that signals of the first expansion card connector 4 and the IO board connector 5 which are blocked by the power module are led out by using a 4c+ cable. The signals of the IO board connector 5 are led out to the IO board of the rear window by using a 4c+ cable, and the signals of the first expansion card connector 4 are led out and then can be used for connecting the pcie equipment. Since the second expansion card connector 6 is not blocked by the power module, no special treatment is done.

Referring to fig. 5 again, a first space is provided between the first expansion card connector 4, the IO board connector 5 and the power module, and the cable terminals of the first expansion card connector 4 and the IO board connector 5 are arranged in the first space.

The cable terminal comprises a main body part and an extension part, wherein the main body part is positioned in the first interval, and the extension part penetrates through the hollowed-out area of the second power backboard and extends towards the front window direction of the case, so that the signal cable can be smoothly led out.

Because the four power modules are respectively arranged on the left side and the right side of the rear window of the chassis in pairs, the IO board is arranged in the middle of the rear window of the chassis, and the second expansion card connector 6 opposite to the IO board is not blocked, the cable terminal with the extension part does not need to be additionally arranged.

In general, the application uses OCSP board type main board which accords with Intel standard to develop 4U 10GPU large card server with length within 883mm, can avoid the research and development investment of main board, and has numerous suppliers of OCSP board type main board and controllable main board cost.

Specifically, the application develops the 4U 10GPU large card server based on the 2U OCSP model, fully reserves the structural parts multiplexing the front window part, including the front storage module structural parts, the heat dissipation module structural parts and the like, only needs to develop the GPU expansion board and two power back boards which are not complex, still reserves all interface functions under the condition of limited length, has low development difficulty, and can well control the batch single machine cost due to more multiplexing parts.

While the application has been described in detail in connection with the general description and the specific embodiments thereof, modifications and improvements may be made thereto. The above description is only of the preferred embodiments of the present application and is not intended to limit the scope of the application, but other variations and modifications can be made by those skilled in the art without departing from the spirit and scope of the application.

Claims

1. A GPU server architecture implemented based on an OCSP motherboard, comprising: OCSP motherboard, GPU expansion board, first power back board, second power back board, power module, IO board, heat dissipation module, memory module;

the method is characterized in that:

2. The GPU server architecture of claim 1, wherein the GPU expansion board is located in a second plane parallel to the first plane above the OCSP motherboard;

3. The OCSP motherboard-based GPU server architecture of claim 1, wherein the first power backplane is disposed within the first plane;

4. The GPU server architecture implemented based on an OCSP motherboard of claim 1, wherein a first power supply path is formed by the power module, the second power backplane, and the first power backplane, providing a first power supply to the GPU expansion board;

5. The GPU server architecture of claim 4, wherein two sets of copper posts are soldered vertically on the first power backplane in the first power path, and the GPU expansion board is inserted in a pluggable manner to supply power.

6. The GPU server architecture of claim 1, wherein the four power modules of the power module are all installed upside down and electrically connected to an end of the second power backplane remote from the OCSP motherboard through the third power connector.

7. The GPU server architecture of claim 6, wherein the orthographic projection of the third power connector and the second power backplane on the first plane partially overlaps the OCSP motherboard.

8. The GPU server architecture of claim 3, wherein an end of the second power backplane near the OCSP motherboard comprises a first region and a second region, and wherein the second region is hollowed out.

9. The GPU server architecture of claim 8, wherein in the third plane, a copper bar is disposed in close proximity to the second area in a third direction perpendicular to the first direction and the second direction and away from the OCSP motherboard;

10. The GPU server architecture of claim 1, wherein a first pitch is provided between the first expansion card connector, the IO board connector, and the power module, and wherein cable terminals of the first expansion card connector, the IO board connector are disposed within the first pitch.