CN103970686A

CN103970686A - GPU (graphic processing unit) expansion card and expansion method

Info

Publication number: CN103970686A
Application number: CN201310045857.0A
Authority: CN
Inventors: 吴志偟
Original assignee: Hongfujin Precision Industry Shenzhen Co Ltd; Hon Hai Precision Industry Co Ltd
Current assignee: Hongfujin Precision Industry Shenzhen Co Ltd; Hon Hai Precision Industry Co Ltd
Priority date: 2013-02-05
Filing date: 2013-02-05
Publication date: 2014-08-06

Abstract

The invention discloses a GPU (graphic processing unit) expansion card. The GPU expansion card comprises a first interface, a second interface, a communication chip, a first control unit and a second control unit. The first interface and the second interface are used for being connected with a mainboard of a server or serially connected with a new auxiliary GPU; the communication chip is used for communication and data transmission among GPUs in serial connection; the first control unit can be triggered only when the GPU expansion card is used in the auxiliary GPU and comprises a request module and a receiving module, the request module is used for enabling the auxiliary GPU to request a main GPU to distribute a subaddress through the communication chip, and the receiving module is used for receiving the subaddress transmitted by the main GPU through the communication chip; the second control unit can be triggered when the GPU expansion card is used in the main GPU and comprises an address distribution module, a detection module and a distributive operation module, the address distribution module is used for distributing the subaddress and transmitting the subaddress to the new auxiliary GPU, the detection module is used for detecting the number of all GPUs in serial connection, and the distributive operation module is used for distributing operation load percentages of all GPUs in a balanced manner and transmitting to all auxiliary GPUs in serial connection through the communication chip.

Description

GPU expansion card and extended method

Technical field

The present invention relates to a kind of GPU expansion card and extended method.

Background technology

Along with the rise of high in the clouds computing and the design of many GPU computings, because a GPU the inside has 256 above stream processing, the server of many enterprise and data center have adopted the design of GPU computing framework, be present among multiple host simultaneously, and use GPU to do the computing that some complexities are very high.But along with the increase of portfolio and computing demand, need to remove dynamic real-time extension GPU, and not be subject to the restriction of bus (Pci Express, PCI-E) slot.

Summary of the invention

In view of above content, be necessary to provide a kind of GPU expanding system and method, dynamically real-time extension GPU.

A kind of GPU expansion card, it comprises: interface one and interface two, for one of the mainboard of connection server or serial connection new from GPU; Communication chip, for and the GPU of each serial connection between link up and communicate by letter and transmit data; Control module one, when this GPU expansion card is for being just triggered from GPU, comprising: request module, distribute a subaddressing from GPU by the main GPU of described communication chip request for this, described subaddressing is used for identifying this from GPU, and the GPU being connected with mainboard is marked as main GPU; Receiver module, the subaddressing passing over by communication chip for receiving main GPU; Control module two, when this GPU expansion card just can be triggered for main GPU, comprising: distribute address module, for distribute a subaddressing and pass to new serial connection from GPU; Detecting module, for detecting the quantity of all GPU of mutual serial connection; Distributive operation module, distributes the computing load number percent of all GPU for balance, and by communication chip pass to all serial connections from GPU.

A kind of GPU extended method, the method comprises: request step, be newly connected in series one during from GPU, should be from GPU by a subaddressing of the main GPU distribution of described communication chip request, described subaddressing is used for identifying this from GPU, and the GPU being connected with mainboard is marked as main GPU; Distribute address step, main GPU distribute a subaddressing and pass to new serial connection from GPU; Receiving step, what be newly connected in series receives from GPU the subaddressing that main GPU passes over by communication chip; Detecting step, the main GPU detecting quantity of all GPU of series connection mutually; Distributive operation step, main GPU balance is distributed the computing load number percent of all GPU, and by communication chip pass to all serial connections from GPU.

Compared to prior art, described GPU expansion card and extended method, can be connected in series in real time multiple GPU by GPU expansion card and share computing load, carrys out the computing load number percent of all GPU of balance by main GPU, and be not subject to the restriction of bus (Pci Express, PCI-E) slot.

Brief description of the drawings

Fig. 1 is the applied environment figure of GPU expansion card of the present invention.

Fig. 2 is the Organization Chart of GPU expansion card of the present invention.

Fig. 3 is the process flow diagram of the preferred embodiment of GPU extended method of the present invention.

Main element symbol description

Server	6
		Mainboard	7
Main GPU	30
		GPU expansion card	40
Communication chip	42
		Storage chip	43
Interface one	44
		Interface two	45
Microprocessor	46
		Control module one	48
Request module	400
		Receiver module	401
Control module two	41
		Distribute address module	410
Detecting module	411
		Distributive operation module	412

Following embodiment further illustrates the present invention in connection with above-mentioned accompanying drawing.

Embodiment

As shown in Figure 1, be the applied environment figure of GPU expansion card of the present invention.In the present embodiment, GPU expansion card 40 is applied in GPU, in the time that the main GPU30 computational burden being connected with server 6 is heavier, can share computing pressure from GPU by GPU expansion card 40 real-time extension.Described server 6 also comprises mainboard 7.Described each GPU expansion card 40 is also connected with external power source.

As shown in Figure 2, be the Organization Chart of GPU expansion card of the present invention.Described GPU expansion card 40 comprises control module 1, control module 2 41, communication chip 42, storage chip 43, interface 1, interface 2 45 and microprocessor 46.Described control module 1 also comprises request module 400 and receiver module 401.Described control module 2 41 comprises distribution address module 410, detecting module 411 and distributive operation module 412.

The function of described interface 1 and described interface 2 45 is the same, is all for connecting mainboard or when downwards expansion is from GPU, can be connected in series new for GPU.In the time that the interface 1 of GPU expansion card 40 or interface 2 45 are directly connected with bus (Pci Express, the PCI-E) interface of mainboard 7, the GPU at this GPU expansion card place is main GPU30, is responsible for communicating and data transmission with server 6.The quantity of described main GPU30 is one, and not being connected with the bus of mainboard 7 from GPU of other, is connected in series each other by interface 1 or the interface 2 45 of GPU expansion card 40.The mode of described serial connection can be passed through winding displacement or other connected mode.

Described communication chip 42 is linked up and is communicated by letter and transmit data with main GPU30 from GPU for each serial connection.

Request module 400 in described control module 1 is for when one of new serial connection is during from GPUn, by universal serial bus I2C upwards single order transmit default request signal request from GPUn-1 and distribute a subaddressing, until this default request signal is passed to main GPU.Described subaddressing, is convenient to main GPU and is distinguished each from GPU and manage each computing load from GPU from GPU for mark.As shown in Figure 1, each serial connection with main GPU all distributes a subaddressing from GPU, from CPU1 corresponding be subaddressing 1, the like, from CPUn-1 corresponding be subaddressing n-1.

Described control module 2 41 just can be triggered in the time that GPU is main GPU, described control module 2 41 is for receiving from the default request signal passing over from GPUn of new serial connection, described distribution address module 410 distribute a subaddressing to new serial connection from GPUn and by communication chip pass to single order from GPU1, again this subaddressing is passed to from GPU2 by communication chip from GPU1, until be delivered to from GPUn.The receiver module 401 of described control module 1 is for receiving the subaddressing passing over from main GPU.After receiving, describedly wait for the load of main GPU distributive operation from GPUn.

Detecting module 411 in described control module 2 41 is for detecting the quantity of all GPU of mutual series connection.

Distributive operation module 412 in described control module 2 41 for according to a kind of computing method give all series connection from GPU distributive operation percentage load.Described computing load number percent is the summation that operand that a GPU bears accounts for whole operand.Described computing method are that the numerical value obtaining is the computing load number percent of each GPU by the described GPU quantity of 100% removal.

Described storage chip 43 is for storing the program segment of data computation and GPU extended method.

Described microprocessor 46 is for the treatment of the program segment of operational data and GPU extended method.

Consulting shown in Fig. 3, is the process flow diagram of the preferred embodiment of GPU extended method of the present invention.According to different demands, in this process flow diagram, the order of step can change, and some step can be omitted.

Step S10, in the time that the GPU computing load being connected in series with server 6 is heavier, on the interface 1 of the expansion card of this GPU or interface 2 45, be connected in series one new for GPUn.

Step S11, request module 400 from the control module 1 of GPUn is distributed a subaddressing by communication chip 42 to send the main GPU of default request signal request from GPUn-1, from GPUn-1, this request signal is passed to from GPUn-2 by communication chip again, until pass to main GPU.

Step S12, main GPU receives after this request signal, distribution address module in control module 2 41 410 give this new serial connection from GPUn distribute a subaddressing and by communication chip pass to single order from GPU1, again this subaddressing is passed to from GPU2 from GPU1, until be delivered to from GPUn.

Step S13, receives from the receiver module 401 of the control module 1 of GPUn the subaddressing passing over successively from main GPU.

Step S14, the detecting module 411 detectings quantity of the GPU of series connection mutually in the control module 2 41 of main GPU.

Step S15, the distributive operation module 412 in the control module 2 41 of main GPU30 is given the GPU3 balance distributive operation percentage load of all series connection according to a kind of computing method, and passes to each GPU.

In the present embodiment, for example, in the time that the quantity of the GPU of serial connection is four mutually, now the computing load number percent of each GPU is 25%.When having, any one GPU computing load is higher, can communicate request branching operation with main GPU and be loaded to GPU on the low side, until the computing load number percent of each GPU is 25%.

By described step S10 to S15, can realize and run into computing full load as GPU, can remove to be connected in series in real time multiple GPU by interface 1 on GPU expansion card 40 or interface 2 45 and share computing load, and carry out the computing load number percent of all GPU of balance by main GPU.

Finally it should be noted that, above embodiment is only unrestricted in order to technical scheme of the present invention to be described, although the present invention is had been described in detail with reference to preferred embodiment, those of ordinary skill in the art is to be understood that, can modify or be equal to replacement technical scheme of the present invention, and not depart from the spirit and scope of technical solution of the present invention.

Claims

1. a GPU expansion card, is characterized in that, described GPU expansion card comprises:

Interface one and interface two, for one of the mainboard of connection server or serial connection new from GPU;

Communication chip, for and the GPU of each serial connection between link up and communicate by letter and transmit data;

Control module one is triggered in the time that this GPU expansion card is used for from GPU, comprising:

Request module, distributes a subaddressing from GPU by the main GPU of described communication chip request for this, and described subaddressing is used for identifying this from GPU, and described main GPU is the GPU being connected with mainboard;

Receiver module, the subaddressing passing over by communication chip for receiving main GPU;

Control module two is triggered in the time that this GPU expansion card is used for main GPU, comprising:

Distribute address module, for distribute a subaddressing and pass to new serial connection from GPU;

Detecting module, for detecting the quantity of all GPU of mutual serial connection;

Distributive operation module, distributes the computing load number percent of all GPU for balance, and by communication chip pass to all serial connections from GPU.

2. GPU expansion card as claimed in claim 1, is characterized in that, described GPU expansion card is connected with external power source.

3. GPU expansion card as claimed in claim 1, it is characterized in that, when this GPU expansion card is when transmitting request from GPU to main GPU, first transmit request to coupled upper single order from GPU, described connected transmit until be delivered to main GPU to the GPU of single order it again from GPU.

4. GPU expansion card as claimed in claim 1, it is characterized in that, when this GPU expansion card is for from GPU, main GPU is during to this from GPU transmission of signal, first pass to be connected with main GPU from GPU, described connected from GPU transmitting until be delivered to this from GPU from GPU to single order it again.

5. a GPU extended method that utilizes the GPU expansion card described in claim 1, is characterized in that, the method comprises:

Request step, is newly connected in series one during from GPU, should be from GPU by a subaddressing of the main GPU distribution of described communication chip request, and described subaddressing is used for identifying this from GPU, and described main GPU is the GPU being connected with mainboard;

Distribute address step, main GPU distribute a subaddressing and pass to new serial connection from GPU;

Receiving step, what be newly connected in series receives from GPU the subaddressing that main GPU passes over by communication chip;

Detecting step, the main GPU detecting quantity of all GPU of series connection mutually;

Distributive operation step, main GPU balance is distributed the computing load number percent of all GPU, and by communication chip pass to all serial connections from GPU.

6. GPU extended method as claimed in claim 5, is characterized in that, described GPU expansion card is connected with external power source.

7. GPU extended method as claimed in claim 5, it is characterized in that, new serial connection transmit request from GPU to main GPU time, first transmit request to coupled upper single order from GPU, described connected transmit until be delivered to main GPU to the GPU of single order it again from GPU.

8. GPU extended method as claimed in claim 5, it is characterized in that, main GPU to receipt message from GPU transmission of signal time, first pass to be connected with main GPU from GPU, described connected from GPU again to single order it from GPU transmit until be delivered to described receipt message from GPU.