CN209265436U

CN209265436U - A kind of double GPU boards

Info

Publication number: CN209265436U
Application number: CN201920250061.1U
Authority: CN
Inventors: 梁超; 赵现普
Original assignee: Suzhou Wave Intelligent Technology Co Ltd
Current assignee: Suzhou Wave Intelligent Technology Co Ltd
Priority date: 2019-02-27
Filing date: 2019-02-27
Publication date: 2019-08-16
Anticipated expiration: 2029-02-27

Abstract

The utility model proposes a kind of double GPU boards, which includes the first GPU, the 2nd GPU, high speed connector, power connector and clock buffer, and the first GPU and the 2nd GPU have SXM2 interface, and pass through NVLINK connection between the two.Power connector gives P12V and P5V power delivery to the first GPU and the 2nd GPU, clock buffer converts the 100M clock signal that upstream service device mainboard device provides to the first clock signal and second clock signal, and the first clock signal is sent to the first GPU respectively, second clock signal is sent to the 2nd GPU.The utility model proposes a kind of double GPU boards connected to guarantee NVLINK signal quality between double GPU without using cable, two GPU with SXM2 interface are placed on the same GPU plate, guarantee two GPU electrification reset and work simultaneously.

Description

A kind of double GPU boards

Technical field

The utility model relates to server technology fields, specifically provide a kind of double GPU boards.

Background technique

In server technology field, with the rise of artificial intelligence, GPU (Graphics Processing Unit figure Processor) it will show that information carries out conversion driving required for computer system, and line scan signals, control are provided to display Embodiment of the advantage of correct display and the operation of display in high-performance calculation is more and more obvious.SXM2 (video-stream processor The high-performance processor module type that company NVIDIA is defined) GPU is NVIDIA company to improve process performance, breaks original A kind of high-performance CPU module for independently defining of PCIE video card specification.

Server collocation GPU is increasingly becoming mainstream, GPU of the generic server using PCIe mark card form, traditional GPU needs Server is provided outside standard PCIe slot, it is also necessary to provide external power cable mouth, use inconvenient.

Utility model content

For disadvantage mentioned above, the utility model embodiment proposes a kind of double GPU boards, using SXM2 interface, realizes single 2 GPU designs of plate, guarantee the signal quality of NVLINK.

The utility model embodiment provides a kind of double GPU boards, and the board includes the first GPU, the 2nd GPU, high speed Connector, power connector and clock buffer；

It is interconnected between first GPU and the 2nd GPU；

The high speed connector is communicated with each other by PCIe3.0 X16 and the first GPU；The high speed connector passes through PCIe3.0 X16 is communicated with each other with the 2nd GPU；

The power connector is connected with the first GPU and the 2nd GPU respectively by power supply line；

The input terminal of the clock buffer is connected with high speed connector；The output end of the clock buffer is respectively with One GPU and the 2nd GPU connection.

Further, the first GPU is the GPU with SXM2 interface.

Further, the 2nd GPU is the GPU with SXM2 interface.

Further, it is communicated with each other between the first GPU and the 2nd GPU by NVLINK.

Further, the power connector gives P12V and P5V power delivery to the first GPU and the 2nd GPU.

Further, the clock buffer converts the 100M PCIe CLOCK signal that upstream plant provides to first 100M PCIe CLOCK signal and the 2nd 100M PCIe CLOCK signal, and respectively by the first 100M PCIe CLOCK signal It is sent to the first GPU, the 2nd 100M PCIe CLOCK signal is sent to the 2nd GPU.

Further, the upstream plant is server master board.

Further, the server master board issue POWER enable signal by high speed connector respectively with SXM2 GPUA and SXM2 GPUB is communicated with each other.

The effect provided in utility model content is only the effect of embodiment, rather than whole effects that utility model is all Fruit, a technical solution in above-mentioned technical proposal have the following advantages that or the utility model has the advantages that

The utility model embodiment proposes a kind of double GPU boards, which includes the first GPU, the 2nd GPU, connects at a high speed Connect device, power connector and clock buffer, the first GPU and the 2nd GPU are the GPU with SXM2 interface, and the first GPU and It is communicated with each other between 2nd GPU by NVLINK.Power connector gives P12V and P5V power delivery to the first GPU and second GPU, clock buffer convert the 100M PCIe CLOCK signal that upstream service device mainboard device provides to the first 100M PCIe CLOCK signal and the 2nd 100M PCIe CLOCK signal, and be respectively sent to the first 100M PCIe CLOCK signal First GPU, the 2nd 100M PCIe CLOCK signal are sent to the 2nd GPU.The utility model proposes a kind of double GPU boards, be Guarantee NVLINK signal quality between double GPU, connected without using cable, two GPU with SXM2 interface is placed on same On a GPU plate, guarantee two GPU electrification reset and work simultaneously.The utility model proposes board can be multiplexed, meeting In the case that interface defines, the quantity of board can increase, and realize flexible configuration.

Detailed description of the invention

In order to illustrate the embodiment of the utility model or the technical proposal in the existing technology more clearly, below to embodiment or The required attached drawing used, which is done, in description of the prior art simply introduces, it should be apparent that, for those of ordinary skill in the art Speech, without creative efforts, is also possible to obtain other drawings based on these drawings.

Fig. 1 is a kind of structural schematic diagram for double GPU boards that the utility model embodiment 1 proposes；

Fig. 2 is a kind of system connection signal of the double GPU board extension multiplexings proposed based on the utility model embodiment 1 Figure.

Specific embodiment

The following will be combined with the drawings in the embodiments of the present invention, carries out the technical scheme in the embodiment of the utility model Clearly and completely describe, it is clear that the described embodiments are only a part of the embodiments of the utility model, rather than whole Embodiment.Based on the embodiments of the present invention, those of ordinary skill in the art are without making creative work Every other embodiment obtained, fall within the protection scope of the utility model.

In the description of the present invention, it should be understood that term " longitudinal direction ", " transverse direction ", "upper", "lower", " preceding ", The orientation or positional relationship of the instructions such as " rear ", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outside" be based on Orientation or positional relationship shown in the drawings is merely for convenience of description the utility model, rather than the dress of indication or suggestion meaning It sets or element must have a particular orientation, be constructed and operated in a specific orientation, therefore should not be understood as to the utility model Limitation.

Embodiment 1

The utility model embodiment 1 provides a kind of double GPU boards, and double GPU boards include the first GPU, the 2nd GPU, height Fast connector, power connector and clock buffer, wherein the first GPU be GPU, the 2nd GPU with SXM2 interface be with It is communicated with each other between the GPU of SXM2 interface, the first GPU and the 2nd GPU by NVLINK.Power connector is by P12V and P5V Power delivery gives the first GPU and the 2nd GPU, the 100M PCIe that clock buffer provides upstream service device mainboard device CLOCK signal is converted to the first 100M PCIe CLOCK signal and the 2nd 100M PCIe CLOCK signal, and respectively by first 100M PCIe CLOCK signal is sent to the first GPU, and the 2nd 100M PCIe CLOCK signal is sent to the 2nd GPU.

A kind of structural schematic diagram of the double GPU boards proposed as shown in Figure 1 for the utility model embodiment 1.In a GPU Above CARD, it is integrated with the GPU A with SXM2 interface and GPU B, SXM2 GPU A and the SXM2 GPU with SXM2 interface Pass through NVLINK interconnected communication between B.Power connector on CPU CARD by power supply line respectively with SXM2 GPU A and SXM2 GPU B is interconnected, and SXM2 GPU A and SXM2 GPU B mainly is defeated by P12V and P5V fax.High speed connector It is communicated with each other by PCIe3.0 X16 and SXM2 GPU A；High speed connector is mutual by PCIe3.0 X16 and SXM2 GPU B Communication；There are also clock buffer on CPU CARD, for keeping clock synchronous.Clock buffer provides upstream plant 100M PCIe CLOCK signal convert to the first 100M PCIe CLOCK signal and the 2nd 100M PCIe CLOCK signal, And the first 100M PCIe CLOCK signal is sent to the first GPU respectively, the 2nd 100M PCIe CLOCK signal is sent to Two GPU.It controls signal POWER Enable and the first 100M PCIe CLOCK signal, the 2nd 100M PCIe CLOCK signal is protected It is synchronous to hold clock.

The process of double GPU card work is to detect GPU CARD first after the uplink port device server booting of double GPU cards Whether be, detect PRSNT_N signal due to be connected to be pulled low on GPU CARD after judge GPU CARD at present in server Under, Power Enable is then become into high potential, opens the power connector above GPU CARD, to SXM2 GPU A and SXM2 GPU B power supply, after the completion of SXM2 GPU A and SXM2 GPU B is powered on, SXM2 GPU A and SXM2 GPU B difference POWER GOOD signal is fed back simultaneously by high-speed signal connector to server, and server confirms GPU after receiving this signal CARD has already powered on completion, RESET signal can be discharged, while to the release of PLTFORM RESET signal, by GPU's RESET signal release guarantees that GPU timing meets its requirement, works normally.SXM2 GPU A and SXM2 GPU B is needed 100M PCIe CLOCK is then to be converted into 2 after a clock buffer after providing 100M PCIe CLOCK by server Group CLOCK, is conveyed to each SXM2 GPU A and SXM2 GPU B respectively.

For the utility model proposes a kind of double GPU boards, two SXM2 GPU are concentrated on a board, guarantee The signal quality of NVLINK, double GPU boards can be multiplexed, as long as in the case where uplink PCIe port meets, it can be by double GPU Board according to actual needs the case where increase.It is a kind of double GPU plates proposed based on the utility model embodiment 1 as shown in Figure 2 The system connection schematic diagram of card extension multiplexing.

GPU CARD A is connected to system uplink PCIe port, GPU CARD B is connected to system uplink PCIe port, GPU CARD C is connected to system uplink PCIe port, GPU CARD X is connected to system uplink PCIe port.

Although specification and drawings and examples have been carried out detailed description to the utility model creation, ability Field technique personnel should be appreciated that still to create the utility model and be modified or replaced equivalently；And all are not departed from The technical solution and its improvement for the spirit and scope that the utility model is created are encompassed by the guarantor that the utility model creates patent It protects in range.

Claims

1. a kind of double GPU boards, which is characterized in that the board includes the first GPU, the 2nd GPU, high speed connector, power supply company Connect device and clock buffer；

It is interconnected between first GPU and the 2nd GPU；

The input terminal of the clock buffer is connected with high speed connector；The output end of the clock buffer is respectively with first GPU and the 2nd GPU connection.

2. a kind of double GPU boards according to claim 1, which is characterized in that the first GPU is with SXM2 interface GPU。

3. a kind of double GPU boards according to claim 1, which is characterized in that the 2nd GPU is with SXM2 interface GPU。

4. a kind of double GPU boards according to claim 1, which is characterized in that lead between the first GPU and the 2nd GPU NVLINK is crossed to communicate with each other.

5. a kind of double GPU boards according to claim 1, which is characterized in that the power connector is by P12V and P5V Power delivery gives the first GPU and the 2nd GPU.

6. a kind of double GPU boards according to claim 1, which is characterized in that the clock buffer mentions upstream plant The 100M PCIe CLOCK signal of confession is converted to the first 100M PCIe CLOCK signal and the 2nd 100M PCIe CLOCK letter Number, and the first 100M PCIe CLOCK signal is sent to the first GPU respectively, the 2nd 100M PCIe CLOCK signal is sent to 2nd GPU.

7. a kind of double GPU boards according to claim 6, which is characterized in that the upstream plant is server master board.

8. a kind of double GPU boards according to claim 7, which is characterized in that the server master board issues POWER Enable signal is communicated with each other with SXM2 GPUA and SXM2 GPUB respectively by high speed connector.