TW201432566A - Expansion card of graphic processing unit and expanding method - Google Patents

Abstract

The present invention provides a expansion card of graphic processing unit which referred to as GPU. The expansion card includes two interfaces, communication chip, the first control unit and the second control unit. The interfaces are used to connect a server motherboard or connect a new GPU. The communication chip is responsible for communication and transmission data. The first control unit includes the request module which requests master GPU to distribute a slave address and the receiving module which receive the slave address from the master GPU. The second unit includes distribution address module which distributes a slave address, detection module which detects the number of all the GPU and distribution operation module is to balance operation percentage of all the GPU.

Description

Graphics processor expansion card and extension method

The invention relates to a GPU expansion card and an expansion method.

With the rise of cloud computing and the design of many GPU operations, because there are more than 256 stream processing in a GPU, many enterprise servers and data centers have adopted the design of GPU computing architecture, and exist in multiple hosts. And use the GPU to do some very complex operations. However, as the volume of business and computing needs increase, it is necessary to dynamically expand the GPU in real time, and is not limited by the bus (Pci Express, PCI-E) slot.

In view of the above, it is necessary to provide a graphics processor expansion card and an extension method, and the graphics processor expansion card is referred to as a GPU expansion card, which can dynamically and dynamically expand the GPU.

A GPU expansion card, comprising: an interface 1 and an interface 2, for connecting a server board or a new slave GPU; a communication chip for communicating and transmitting data with each serial GPU The control unit first, when the GPU expansion card is used to be triggered from the GPU, includes: a requesting module, wherein the slave GPU requests the primary GPU to allocate a sub-address through the communication chip, the sub-address For identifying the slave GPU, the GPU connected to the motherboard is marked as the main GPU; the receiving module is configured to receive the sub-address transmitted by the main GPU through the communication chip; and the control unit 2, when the GPU expansion card is used for the main The GPU will be triggered, including: an allocation address module for allocating a sub-address and passing it to the newly connected slave GPU; and a detection module for detecting the number of all GPUs connected in series; A computing module is allocated to balance the percentage of computing load allocated to all GPUs and passed through the communication chip to all serial slave GPUs.

A GPU extension method, the method comprising: a requesting step, when newly connecting a slave GPU, the slave GPU requests the master GPU to allocate a sub-address through the communication chip, where the sub-address is used to identify the slave GPU, and The GPU connected to the motherboard is marked as the main GPU; in the address allocation step, the main GPU allocates a sub-address and passes it to the newly connected slave GPU; in the receiving step, the newly connected slave GPU receives the main GPU and transmits it through the communication chip. Sub-address; detection step, the main GPU detects the number of all GPUs connected in series; the allocation operation steps, the main GPU balances the operating load percentage of all GPUs, and transmits them to all serial slave GPUs through the communication chip.

Compared with the prior art, the GPU expansion card and the extension method can directly load multiple GPUs through the GPU expansion card to share the computing load, balance the computing load percentage of all the GPUs through the main GPU, and are not subject to the bus bar ( Pci Express, PCI-E) slot limitations.

6. . . server

7. . . motherboard

30. . . Main GPU

40. . . GPU expansion card

42. . . Communication chip

43. . . Memory chip

44. . . Interface one

45. . . Interface two

46. . . microprocessor

48. . . Control unit one

400. . . Request module

401. . . Receiving module

41. . . Control unit two

410. . . Allocation address module

411. . . Detection module

412. . . Distribution computing module

1 is a diagram showing an application environment of a GPU expansion card of the present invention.

2 is a block diagram of a GPU expansion card of the present invention.

3 is a flow chart of a preferred embodiment of the GPU expansion method of the present invention.

As shown in FIG. 1, it is an application environment diagram of the GPU expansion card of the present invention. In this embodiment, the GPU expansion card is referred to as a GPU expansion card. The GPU expansion card 40 is applied to the GPU. When the main GPU 30 connected to the server 6 has a heavy computing burden, the GPU expansion card 40 can be instantly expanded. The GPU shares the computational pressure. The server 6 also includes a motherboard 7. Each of the GPU expansion cards 40 is also coupled to an external power source.

As shown in FIG. 2, it is an architectural diagram of the GPU expansion card of the present invention. The GPU expansion card 40 includes a control unit 48, a control unit 22, a communication chip 42, a memory chip 43, an interface 44, an interface 25, and a microprocessor 46. The control unit one 48 further includes a request module 400 and a receiving module 401. The control unit 22 includes a distribution address module 410, a detection module 411, and a distribution operation module 412.

The functions of the interface 44 and the interface 2 are the same as those used to connect to the motherboard or extend down from the GPU to connect the new slave GPU. When the interface 44 or the interface 25 of the GPU expansion card 40 is directly connected to the bus (Pci Express, PCI-E) interface of the motherboard 7, the GPU where the GPU is located is the main GPU 30, and is responsible for the server. 6 Communication and data transmission. The number of the main GPUs 30 is one. The other slave GPUs are not connected to the busbars of the motherboard 7, and are connected to each other through the interface 44 or the interface 25 of the GPU expansion card 40. The manner of serial connection can be through a cable or other connection.

The communication chip 42 is used for communication and transmission of data between the serially connected slave GPU and the main GPU 30.

The request module 400 in the control unit 48 is configured to, when newly connecting a slave GPU, transmit a preset request signal from the GPU n-1 to the first order through the serial bus I2C to allocate a sub-address, until The preset request signal is passed to the main GPU. The sub-address is used to identify the slave GPU, so that the main GPU can distinguish the respective slave GPUs and manage the computation load of each slave GPU. As shown in FIG. 1, each slave GPU serially connected to the main GPU is assigned a sub-address, the slave CPU 1 corresponds to the sub-address 1, and so on, and the slave CPU n-1 corresponds to the sub-address n-1.

The control unit 22 is triggered when the GPU is the main GPU, and the control unit 22 is configured to: after receiving the preset request signal transmitted from the newly connected GPU n, the allocation address module 410 allocates a subaddress to the newly concatenated slave GPUn and passes it through the communication chip to the next stage slave GPU1, which then passes the subaddress to the slave GPU2 through the communication chip until it is passed to the slave GPUn. The receiving module 401 of the control unit 48 is configured to receive the sub-address transmitted from the main GPU. After receiving, the slave GPUn waits for the main GPU to allocate an arithmetic load.

The detection module 411 of the control unit 22 is configured to detect the number of all GPUs connected in series.

The allocation operation module 412 in the control unit 22 is configured to allocate the operating load percentage to all the connected slave GPUs according to a calculation method. The operation load percentage is the sum of the calculation amount undertaken by one GPU and the entire operation amount. The calculation method is to remove 100% of the number of GPUs, and the obtained value is the operation load percentage of each GPU.

The memory chip 43 is used to store a program segment of the data and GPU expansion method.

The microprocessor 46 is used to process the program data and the program segment of the GPU expansion method.

Referring to FIG. 3, it is a flow chart of a preferred embodiment of the GPU expansion method of the present invention. The order of the steps in the flowchart may be changed according to different requirements, and some steps may be omitted.

In step S10, when the GPU computing load connected to the server 6 is heavy, a new slave GPU is serially connected to the interface 44 or the interface 45 of the expansion card of the GPU.

Step S11, the request module 400 in the control unit 48 of the GPUn sends a preset request signal from the GPUn-1 through the communication chip 42 to request the primary GPU to allocate a sub-address, and the request signal is re-transmitted from the GPUn-1. The wafer is passed to the slave GPUn-2 until it is passed to the master GPU.

Step S12, after the main GPU receives the request signal, the allocated address module 410 in the control unit 2 41 allocates a sub-address to the newly serialized GPUn and transmits the sub-address to the next-order slave GPU1 through the communication chip. The subaddress is then passed from GPU 1 to slave GPU 2 until it is passed to slave GPUn.

In step S13, the sub-address transmitted from the main GPU in sequence is received from the receiving module 401 of the control unit 48 of the GPUn.

In step S14, the detection module 411 in the control unit 411 of the main GPU detects the number of GPUs connected in series.

In step S15, the distribution operation module 412 in the control unit XX of the main GPU 30 balances the operation load percentages to all the GPUs 3 connected in series according to a calculation method, and transmits them to the respective GPUs.

In this embodiment, for example, when the number of GPUs connected in series is four, the calculation load percentage of each GPU at this time is 25%. When any GPU computing load is too high, it will communicate with the main GPU to request the transfer operation load to the lower GPU until the operating load percentage of each GPU is 25%.

Through the steps S10 to S15, when the GPU encounters the full load of the operation, the GPU can be connected to the GPU through the interface 44 or the interface 45 of the GPU expansion card 40 to share the computing load and pass through the main GPU. Balance the operating load percentage of all GPUs.

It should be noted that the above embodiments are only for explaining the technical solutions of the present invention and are not intended to be limiting, and the present invention will be described in detail with reference to the preferred embodiments. Modifications or equivalents are made without departing from the spirit and scope of the invention.

40. . . GPU expansion card

42. . . Communication chip

43. . . Memory chip

44. . . Interface one

45. . . Interface two

48. . . Control unit one

46. . . microprocessor

400. . . Request module

401. . . Receiving module

41. . . Control unit two

410. . . Allocation address module

411. . . Detection module

412. . . Distribution computing module

Claims (8)

A graphics processor expansion card, the graphics processor expansion card comprising:
Interface 1 and interface 2, used to connect to the server's motherboard or serially connect a new slave GPU;
a communication chip for communicating and transmitting data with each GPU connected in series;
Control unit one, when the graphics processor expansion card is used to be triggered from the GPU, includes:
a requesting module, wherein the slave GPU requests the primary GPU to allocate a sub-address through the communication chip, the sub-address is used to identify the slave GPU, and the master GPU is a GPU connected to the motherboard;
a receiving module, configured to receive a sub-address transmitted by the main GPU through the communication chip;
Control unit 2, when the graphics processor expansion card is used in the main GPU, is triggered, including:
Assigning an address module for allocating a sub-address and passing it to the newly concatenated slave GPU;
a detection module for detecting the number of all GPUs connected in series;
A computing module is allocated to balance the percentage of computing load allocated to all GPUs and passed through the communication chip to all serial slave GPUs. According to the graphics processor expansion card of claim 1, the graphics processor expansion card is connected to an external power source. According to the graphics processor expansion card of claim 1, when the graphics processor expansion card is used to transfer a request from the GPU to the main GPU, the request is first transmitted to the upper-order slave GPU connected thereto, the connected From the GPU to its upper-order GPU pass until it is passed to the main GPU. According to the graphics processor expansion card of claim 1, when the graphics processor expansion card is used in the slave GPU, when the master GPU transmits a signal to the slave GPU, it is first transferred to the slave GPU connected to the master GPU. The connected slave GPU passes to the next-stage slave GPU until it is passed to the slave GPU. An extension method of a graphics processor expansion card according to claim 1 of the patent application scope, the method comprising:
a requesting step, when the slave GPU is newly connected in series, the slave GPU requests the master GPU to allocate a sub-address through the communication chip, where the sub-address is used to identify the slave GPU, and the master GPU is connected to the motherboard GPU;
In the address allocation step, the primary GPU allocates a sub-address and passes it to the newly concatenated slave GPU;
Receiving step, the newly connected slave GPU receives the sub-address transmitted by the main GPU through the communication chip;
In the detecting step, the main GPU detects the number of all GPUs connected in series;
In the allocation operation step, the main GPU balances the operating load percentage of all GPUs and passes them to all serial slave GPUs through the communication chip. According to the expansion method of claim 5, the graphics processor expansion card is connected to an external power source. According to the extension method of claim 5, when the graphics processor expansion card is used to transfer a request from the GPU to the main GPU, the request is first transmitted to the GPU connected to the upper stage, and the connected slave GPU Pass to the upper-order GPU until it is passed to the main GPU. According to the expansion method of claim 5, when the graphics processor expansion card is used in the slave GPU, when the master GPU transmits a signal to the slave GPU, it is first transferred to the slave GPU connected to the master GPU, and the connected slave The GPU then passes to the next stage of the slave GPU until it is passed to the slave GPU.