CN210639641U - Artificial intelligence training card, artificial intelligence server and special processing card - Google Patents
Artificial intelligence training card, artificial intelligence server and special processing card Download PDFInfo
- Publication number
- CN210639641U CN210639641U CN201920507498.9U CN201920507498U CN210639641U CN 210639641 U CN210639641 U CN 210639641U CN 201920507498 U CN201920507498 U CN 201920507498U CN 210639641 U CN210639641 U CN 210639641U
- Authority
- CN
- China
- Prior art keywords
- chip
- artificial intelligence
- training
- card
- interconnection
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Landscapes
- Power Sources (AREA)
Abstract
The application provides an artificial intelligence training card, an artificial intelligence server and a special processing card. The artificial intelligence training card comprises an interconnection chip, an artificial intelligence chip and a network chip. The artificial intelligence chip is connected with the interconnection chip, receives the training task through the interconnection chip and executes the training task. The network chip is connected with the artificial intelligence chip through the interconnection chip.
Description
Technical Field
The application relates to artificial intelligence, in particular to an artificial intelligence training card, an artificial intelligence server and a special processing card.
Background
Artificial Intelligence (AI) is a new technical science for studying and developing theories, methods, techniques and application systems for simulating, extending and expanding human Intelligence. The process of artificial intelligence comprises three stages of perception, decision and feedback, and the short board of any stage directly limits the popularization and application of the artificial intelligence technology. With the development of artificial intelligence technology, the artificial intelligence technology is gradually applied to various fields.
SUMMERY OF THE UTILITY MODEL
The application provides an artificial intelligence training card, an artificial intelligence server and a special processing card, which can realize the training of large-scale models.
One aspect of the present application provides an artificial intelligence training card comprising: interconnecting the chips; the artificial intelligence chip is connected with the interconnection chip, receives a training task through the interconnection chip and executes the training task; and the network chip is connected with the artificial intelligence chip through the interconnection chip.
Further, the interconnection chip includes a PCIE switch chip.
Further, the artificial intelligence training card comprises a main interface connected with the interconnection chip, wherein the main interface comprises a PCIE interface and receives a training task.
Furthermore, the artificial intelligence training card comprises a power interface and a power conversion module connected with the power interface, wherein the power conversion module is connected with the artificial intelligence chip, the interconnection chip and the network chip, and converts the electricity input through the power interface to provide the working voltage required by the artificial intelligence chip, the interconnection chip and the network chip.
Furthermore, the power conversion module comprises a first power converter, and the first power converter is connected with the power interface and the artificial intelligence chip and supplies power to the artificial intelligence chip.
Furthermore, the power conversion module comprises a second power converter, and the second power converter is connected with the power interface, connected with the interconnection chip and the network chip, and supplies power to the interconnection chip and the network chip.
Furthermore, the artificial intelligence training card comprises a main interface and a routing module connected with the main interface, wherein the routing module is connected with the artificial intelligence chip, the interconnection chip and the network chip.
Further, the routing module includes at least one of an I2C interface and a UART interface.
Further, the artificial intelligence training card comprises at least one of a temperature sensor, an analog-to-digital conversion module and an information storage module.
Another aspect of the present application provides an artificial intelligence server, comprising: an artificial intelligence training card; and the mainboard is provided with a controller, and the controller is connected with the artificial intelligence training card and issues a training task.
Another aspect of the present application provides a special purpose processing card, comprising: interconnecting the chips; the processing chip is connected with the interconnection chip, receives a processing task through the interconnection chip and executes the processing task; and the network chip is connected with the processing chip through the interconnection chip.
The artificial intelligence chip receives the training task through the interconnection chip and communicates with other artificial intelligence units through the network chip to realize cluster type training, so that the artificial intelligence server adopting the artificial intelligence training card can realize the training of large-scale models.
Drawings
FIG. 1 is a schematic diagram of a communication module of an embodiment of an artificial intelligence server according to the present application;
FIG. 2 is a schematic diagram illustrating power supplied to one embodiment of an artificial intelligence training card of the artificial intelligence server shown in FIG. 1;
FIG. 3 is a schematic diagram illustrating the administrative monitoring logic for one embodiment of the artificial intelligence training card of FIG. 1;
FIG. 4 is a schematic diagram illustrating an embodiment of an artificial intelligence training card according to the present application;
FIG. 5 is a schematic diagram illustrating a system cluster networking according to an embodiment of the artificial intelligence training card of the present application;
FIG. 6 is a diagram of an embodiment of a specialized processing card of the present application.
Detailed Description
Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus consistent with certain aspects of the present application, as detailed in the appended claims.
The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. Unless defined otherwise, technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention belongs. The use of "first," "second," and similar terms in the description and in the claims does not indicate any order, quantity, or importance, but rather is used to distinguish one element from another. Also, the use of the terms "a" or "an" and the like do not denote a limitation of quantity, but rather denote the presence of at least one. The word "plurality" or "a number" and the like mean two or more. The word "comprising" or "comprises", and the like, means that the element or item listed as preceding "comprising" or "includes" covers the element or item listed as following "comprising" or "includes" and its equivalents, and does not exclude other elements or items. The terms "connected" or "coupled" and the like are not restricted to physical or mechanical connections, but may include electrical connections, whether direct or indirect. 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 and all possible combinations of one or more of the associated listed items.
The artificial intelligence training card of the embodiment of the application comprises an interconnection chip, an artificial intelligence chip and a network chip. The artificial intelligence chip is connected with the interconnection chip, receives the training task through the interconnection chip and executes the training task. The network chip is connected with the artificial intelligence chip through the interconnection chip.
The artificial intelligence chip receives the training task through the interconnection chip and communicates with other artificial intelligence units through the network chip, such as another artificial intelligence training card, other artificial intelligence units of the artificial intelligence server or other artificial intelligence servers, so as to realize cluster training, and the artificial intelligence server adopting the artificial intelligence training card can realize the training of large-scale models. The artificial intelligence server can utilize a plurality of artificial intelligence training cards to respectively train a plurality of training tasks, so that the training of large-scale models can be realized by utilizing a sufficient number of artificial intelligence training cards, the concurrent massive training tasks are completed, the rapid iterative evolution is realized, and the calculation capacity of the artificial intelligence is improved.
Fig. 1 is a schematic diagram of a communication module of one embodiment of an artificial intelligence server 100. The artificial intelligence server 100 includes an artificial intelligence training card 200 and a main board 300. The artificial intelligence server 100 includes one or more artificial intelligence training cards 200. Only one artificial intelligence training card 200 is schematically shown. The artificial intelligence training card 200 may implement model training. The main board 300 is provided with a controller 301, and the controller 301 is connected with the artificial intelligence training card 200 to issue a training task. The controller 301 may include a central processor.
In applications requiring large-scale model training, the artificial intelligence server 100 includes multiple artificial intelligence training cards 200, and may even include a larger number of artificial intelligence training cards 200 to meet the requirements of large-scale model training. The controller 201 concurrently executes a plurality of training tasks, which may be assigned to a sufficient number of artificial intelligence training cards 200 to perform the training tasks individually.
The artificial intelligence training card 200 includes an interconnect chip 201, an artificial intelligence chip 202, and a network chip 203. The artificial intelligence chip 202 is connected with the interconnection chip 201, receives a training task through the interconnection chip 201, and executes the training task. The network chip 203 is connected with the artificial intelligence chip 202 through the interconnection chip 201.
The artificial intelligence chip 202 receives the training task through the interconnection chip 201 and communicates with other artificial intelligence units, such as other artificial intelligence units of the artificial intelligence server, another artificial intelligence training card, or other artificial intelligence servers, through the network chip 203 to implement cluster training, so that the artificial intelligence server 100 using the artificial intelligence training card 200 can implement large-scale model training. In one embodiment, the data trained by the artificial intelligence chip 202 may be sent to the next artificial intelligence training card, and the multiple artificial intelligence training cards may form an artificial intelligence training cluster. In some embodiments, the artificial intelligence server 100 may utilize a plurality of artificial intelligence training cards 200 to perform training of a plurality of training tasks, respectively, so that a sufficient number of artificial intelligence training cards 200 are utilized to implement training of large-scale models, complete concurrent massive training tasks, implement fast iterative evolution, and improve the computational power of artificial intelligence.
The interconnection chip 201 is connected with the controller 301, the artificial intelligence chip 202 and the network chip 203 to realize data transmission. In one embodiment, the interconnect chip 201 may enable the artificial intelligence chip 202 to switch between the controller 201 and the network chip 203, enable the artificial intelligence chip 202 to communicate data with the controller 201, or enable the artificial intelligence chip 202 to communicate data with the network chip 203. The interconnect chip 201 may be connected to the controller 301 and the artificial intelligence chip 202, and provide the training task issued by the controller 301 to the artificial intelligence chip 202. The interconnection chip 201 may connect the artificial intelligence chip 202 and the network chip 203, and transmit a training model generated by the artificial intelligence chip 202 after training to the network chip 203. In other embodiments, the interconnect chip 201 may connect the controller 301 and the network chip 203 to implement data transmission therebetween. The interconnect chip 201 may enable data transfer between any two of the controller 301, the artificial intelligence chip 202, and the network chip 203.
The artificial intelligence training card 200 includes a host interface 204 that interfaces with the interconnect chip 201. The main interface 204 is connected to the motherboard 300, and the interconnection chip 201 may be connected to the controller 301 through the main interface 204. The training task issued by the controller 301 is transmitted to the interconnection chip 201 through the main interface 204, and further transmitted to the artificial intelligence chip 202. In one embodiment, the main interface 204 comprises a PCIE interface that receives training tasks. The PCIE switch chip is connected to the PCIE interface, and further connected to the controller 301. In one example, the transmission rate between the PCIE interface and the PCIE switch chip, the transmission rate between the PCIE switch chip and the artificial intelligence chip 202, and the transmission rate between the PCIE switch chip and the network chip 203 may all be PCIE X16.
The artificial intelligence chip 202 may be used for intelligent computing such as model training. The artificial intelligence chip 202 executes model training according to the training task, generates a training model after the training is completed, and the training model can be used in subsequent intelligent calculation. The artificial intelligence chip 202 may be an artificial intelligence chip of Alibara, but is not limited to an artificial intelligence chip of Alibara.
In one example, the network chip 203 may be a 200G network chip, such as a Mellanox connection CX6-DX model chip, which can implement high bandwidth communication. The artificial intelligence training card 200 comprises an optical module 205, and the optical module 205 is connected with the network chip 203 and is connected with an external network so as to communicate with other artificial intelligence units. In some embodiments, the artificial intelligence training card 200 may receive externally transmitted signals through the optical module 205 and the network chip 203.
FIG. 2 is a schematic diagram of the power supply of one embodiment of an artificial smart training card 200. The artificial intelligence training card 200 comprises a power interface 206 and a power conversion module 207 connected with the power interface 206, wherein the power conversion module 207 is connected with the artificial intelligence chip 202, the interconnection chip 201 and the network chip 203, converts electricity input through the power interface 206 and provides working voltage required by the artificial intelligence chip 202, the interconnection chip 201 and the network chip 203. The power interface 206 is connected to a power source. In one embodiment, the artificial intelligence server 100 can include a power strip 400, and the power strip 400 can receive alternating current, such as mains power, and convert the alternating current to direct current (e.g., 48 vdc) to the power interface 206. The power conversion module 207 may reduce the voltage of the dc received by the power interface 206 and provide the reduced voltage to the artificial intelligence chip 202, the interconnect chip 201, and the network chip 203.
In some embodiments, the power conversion module 207 includes a first power converter 208, and the first power converter 208 is coupled to the power interface 206 and the artificial intelligence chip 202 to provide power to the artificial intelligence chip 202. In one embodiment, the first power converter 208 may convert the voltage of the direct current received by the power interface 206 into a voltage suitable for the artificial intelligence chip 202.
In some embodiments, the power conversion module 207 includes a second power converter 209, and the second power converter 209 is connected to the power interface 206 and is connected to the interconnect chip 201 and the network chip 203 to supply power to the interconnect chip 201 and the network chip 203. The second power converter 209 can convert the voltage of the direct current power received by the power interface 206 into a voltage suitable for the interconnect chip 201 and the network chip 203. The voltage converted by the second power converter 209 may be different from the voltage converted by the first power converter 208. In other embodiments, the power conversion module 207 is designed according to the voltages required by the artificial intelligence chip 202, the interconnection chip 201 and the network chip 203, and is not limited to the illustrated embodiment.
In some embodiments, the power conversion module 207 may also provide power to other components of the artificial smart training card 200. In some embodiments, the artificial intelligence training card 200 includes at least one of a temperature sensor 210, an analog-to-digital conversion module 211, and an information storage module 212. In the illustrated embodiment, the artificial intelligence training card 200 includes a temperature sensor 210, an analog-to-digital conversion module 211, and an information storage module 212. The power conversion module 207 may provide power to the temperature sensor 210, the analog-to-digital conversion module 211, and the information storage module 212. In the illustrated embodiment, the second power converter 209 may provide power to the temperature sensor 210, the analog-to-digital conversion module 211, and the information storage module 212.
The temperature sensor 210 may sense the temperature of the artificial smart training card 200. The analog-to-digital conversion module 211 may convert an analog signal into a digital signal, for example, an analog signal of the temperature sensor 210 may be converted into a digital signal. The information storage module 212 may be configured to record information of the board manufacturer, the factory date, the SN code, and the like of the artificial intelligence training card 200.
In one embodiment, the artificial intelligence training card 200 receives a standby control (STBY) power, such as a 3.3V _ STBY power, from the motherboard 300 via the host interface 204 as a power to monitor STBY status management of the artificial intelligence training card 200.
FIG. 3 is a diagram illustrating the administrative monitoring logic for one embodiment of the artificial smart training card 200. In some embodiments, the artificial intelligence training card 200 includes a routing module 213 coupled to the host interface 204, the routing module 213 coupled to the artificial intelligence chip 202, the interconnect chip 201, and the network chip 203. The routing module 213 may be connected to the motherboard 300 via the host interface 204 and may be connected to the controller 301. The routing module 213 can include a routing chip that can select a channel. In some embodiments, the routing module 213 can be coupled to at least one of the temperature sensor 210, the analog-to-digital conversion module 211, and the information storage module 212.
In one embodiment, the routing module 213 includes at least one of an I2C (Inter-Integrated Circuit) interface and a UART (Universal Asynchronous Receiver/Transmitter) interface. In the illustrated embodiment, the routing module 213 includes an I2C interface 214 and a UART interface 215. The routing module 213 may include an I2C/UART routing chip. The routing module 213 is connected to the artificial intelligence chip 202 and the network chip 203 through the UART interface 215. The UART interface 215 may be a debug (debug) interface of the artificial intelligence chip 202 and the network chip 203, and may be connected to the motherboard 300 or an external debug device through the main interface 204. When the artificial intelligence chip 202 and the network chip 203 have faults, the fault information can be acquired through the UART interface 215 and debugging can be performed.
The routing module 213 is connected to the artificial intelligence chip 202, the network chip 203 and the interconnect chip 201 via an I2C interface 214. The working state of the artificial intelligence training card 200 can be monitored through the I2C interface 214, for example, register values of the artificial intelligence chip 202, the network card chip 203 and the interconnection chip 201 are obtained, and information in the above chips can be refreshed through the I2C interface 214. In one embodiment, the routing module 213 is also coupled to the temperature sensor 210 via an I2C interface 214, and the value of the temperature sensor 210 can be read via an I2C interface 214. In one embodiment, the routing module 213 is further connected to the analog-to-digital conversion module 211 via an I2C interface 214, and the value of the analog-to-digital conversion module 211 can be obtained via an I2C interface 214. In one embodiment, the routing module 213 is also coupled to the information storage module 212 via the I2C interface 214, and information in the information storage module 212 can be read via the I2C interface 214.
FIG. 4 is a schematic diagram illustrating an embodiment of an artificial smart training card 200. The artificial intelligence training card 200 includes a circuit board 216, such as a printed circuit board, which may be substantially rectangular in shape. In one example, the circuit board 216 may have a length of 250mm and a width of 112mm, but is not limited thereto. The artificial intelligence chip 202, the network chip 203 and the interconnect chip 201 may be distributed on a circuit board 216. The main interface 204 is located on the side of the circuit board 216. The primary interface 204 may be a conductive contact interface (or gold finger) attached to the circuit board 216 and partially extending outside the circuit board 216. The artificial intelligence training card 200 may also include other elements, not shown in FIG. 4.
FIG. 5 is a system cluster networking diagram of one embodiment of an artificial smart training card 200. The plurality of artificial intelligence training cards 200 are applied to one artificial intelligence server 100, and are one server node. The server 100 may include a card conversion unit 101 connected to a plurality of artificial intelligence training cards 200. In one example, the plurality of artificial intelligence training cards 200 may be divided into a plurality of groups, each group including a plurality of artificial intelligence training cards 200 divided into a plurality of stages. For example, in the figure, the server 100 includes 16 artificial intelligence training cards 200, which are divided into four groups, each group includes four levels, and the card conversion unit 101 can switch between the groups.
A plurality of artificial intelligence servers 100 can form a primary training cluster 500, and the plurality of artificial intelligence servers 100 can communicate with each other. The training cluster 500 includes 16 artificial intelligence servers, but is not limited thereto. A plurality of primary training clusters 500 may form a larger secondary training cluster 600, and the plurality of primary training clusters 500 may communicate with each other.
Fig. 6 is a block diagram of an embodiment of a specialized processing card 700 of the present application. The specialized processing card 700 includes an interconnect chip 701, a processing chip 702, and a network chip 703. The processing chip 702 is connected to the interconnect chip 701, receives a processing task through the interconnect chip 701, and executes the processing task. The network chip 703 is connected to the processing chip 702 through the interconnect chip 701. The processing chip 702 receives processing tasks through the interconnect chip 701 and communicates with other specialized processing cards or devices through the network chip 703.
In some embodiments, the processing chip 702 may be used in a training phase of a deep learning neural network, or may be used in an inference phase. The processing chip 702 may be an artificial intelligence chip. Interconnect chip 701 is similar to interconnect chip 201 described above. The network chip 703 is similar to the network chip 203 described above. The interconnect chip 701 may be connected to a motherboard via a host interface 704. The network chip 703 may communicate with other special purpose processing cards or devices through the optical module 705. The processing chip 702 and its modules are described similarly to the artificial intelligence training card 100 described above.
The above description is only exemplary of the present application and should not be taken as limiting the present application, as any modification, equivalent replacement, or improvement made within the spirit and principle of the present application should be included in the scope of protection of the present application.
Claims (11)
1. An artificial intelligence training card comprising:
interconnecting the chips;
the artificial intelligence chip is connected with the interconnection chip, receives a training task through the interconnection chip and executes the training task; and
and the network chip is connected with the artificial intelligence chip through the interconnection chip.
2. The artificial intelligence training card of claim 1, wherein: the interconnection chip comprises a PCIE switching chip.
3. The artificial intelligence training card of claim 2, wherein: the artificial intelligence training card comprises a main interface connected with the interconnection chip, wherein the main interface comprises a PCIE interface and receives a training task.
4. The artificial intelligence training card of claim 1, wherein: the artificial intelligence training card comprises a power interface and a power conversion module connected with the power interface, wherein the power conversion module is connected with the artificial intelligence chip, the interconnection chip and the network chip convert electricity input through the power interface to provide working voltage required by the artificial intelligence chip, the interconnection chip and the network chip.
5. The artificial intelligence training card of claim 4, wherein: the power conversion module comprises a first power converter, and the first power converter is connected with the power interface and the artificial intelligence chip and supplies power to the artificial intelligence chip.
6. The artificial intelligence training card of claim 4, wherein: the power conversion module comprises a second power converter, the second power converter is connected with the power interface and the interconnection chip and the network chip and supplies power to the interconnection chip and the network chip.
7. The artificial intelligence training card of claim 1, wherein: the artificial intelligence training card comprises a main interface and a route selection module connected with the main interface, and the route selection module is connected with the artificial intelligence chip, the interconnection chip and the network chip.
8. The artificial intelligence training card of claim 7, wherein: the routing module includes at least one of an I2C interface and a UART interface.
9. The artificial intelligence training card of claim 1, wherein: the artificial intelligence training card comprises at least one of a temperature sensor, an analog-to-digital conversion module and an information storage module.
10. An artificial intelligence server, comprising:
the artificial intelligence training card of any one of claims 1-9; and
the main board is provided with a controller, and the controller is connected with the artificial intelligence training card and issues a training task.
11. An application specific processing card comprising:
interconnecting the chips;
the processing chip is connected with the interconnection chip, receives a processing task through the interconnection chip and executes the processing task; and
and the network chip is connected with the processing chip through the interconnection chip.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201920507498.9U CN210639641U (en) | 2019-04-15 | 2019-04-15 | Artificial intelligence training card, artificial intelligence server and special processing card |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201920507498.9U CN210639641U (en) | 2019-04-15 | 2019-04-15 | Artificial intelligence training card, artificial intelligence server and special processing card |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN210639641U true CN210639641U (en) | 2020-05-29 |
Family
ID=70794336
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201920507498.9U Active CN210639641U (en) | 2019-04-15 | 2019-04-15 | Artificial intelligence training card, artificial intelligence server and special processing card |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN210639641U (en) |
-
2019
- 2019-04-15 CN CN201920507498.9U patent/CN210639641U/en active Active
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113872796B (en) | Server and node equipment information acquisition method, device, equipment and medium thereof | |
| CN103138971A (en) | Server cabinet system | |
| CN104407680A (en) | Management system based on overall cabinet nodes | |
| CN113238802A (en) | Interrupt distributor, data processing chip, interrupt distribution method and data processing method | |
| CN103729319A (en) | Equipment system based on serial bus and data transmission method | |
| CN104678809A (en) | Universal sensor control equipment and system | |
| CN210639641U (en) | Artificial intelligence training card, artificial intelligence server and special processing card | |
| CN219695682U (en) | Edge device for industrial site | |
| CN112000545A (en) | Graphics processor board card and graphics processor management method | |
| CN205263790U (en) | Display control board | |
| CN113568847B (en) | Network card and processor interconnection device and server | |
| CN103095739A (en) | Cabinet server system and node communication method thereof | |
| CN109240972A (en) | A kind of GPU board and the VPX signal processing cabinet using the board | |
| CN112769968B (en) | Circuit breaker, cabinet, system, address acquisition method and equipment | |
| CN100447714C (en) | Online switching interfaces of peripheral devices in multitask mode computer | |
| CN204423022U (en) | Can general purpose transducer opertaing device and system | |
| CN210351233U (en) | Circuit breaker, cabinet and system | |
| CN111400240A (en) | Circuit board and artificial intelligence edge computing embedded controller | |
| CN216286659U (en) | High-speed data processing equipment based on homemade VPX framework | |
| CN114840461B (en) | Expansion device of server and server | |
| CN219246072U (en) | Multi-core computing board card, computer and intelligent computing system | |
| CN217955105U (en) | Bearing board card, computer host equipment and computer | |
| CN217181527U (en) | CPU prototype PCIe interface hot plug function verification platform | |
| CN218734654U (en) | Extensible switch | |
| CN216623138U (en) | Industrial control computer |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| GR01 | Patent grant | ||
| GR01 | Patent grant |