CN209560410U - A kind of expansible AI acceleration system structure based on OCP standard - Google Patents

Abstract

The expansible AI acceleration system structure based on OCP standard that the utility model discloses a kind of, including acceleration system cabinet, the HGX-2 mainboard of two carrying 8GPU, Nvlink interconnection board for the interconnection of two pieces of HGX-2 mainboards, Fan Transit Unit for being interconnected between multiple fan mould groups, power supply interface and the main pinboard for carrying BMC managing chip, the cabinet is divided into upper layer and lower layer, two pieces of HGX-2 mainboards are separately positioned on the upper layer and lower layer of cabinet, two pieces of HGX-2 mainboards are interconnected in the rear end of cabinet by Nvlink interconnection board, the Fan Transit Unit parallel with Nvlink interconnection board is equipped with behind Nvlink interconnection board, the Fan Transit Unit is equipped with fan management device and several fan mould groups, the fan management device and several fan mould group electricity Connection, the fan management device are electrically connected with the main pinboard；The front end of two pieces of HGX-2 mainboards is connected separately with one piece of main pinboard.

Description

A kind of expansible AI acceleration system structure based on OCP standard

Technical field

The utility model relates to AI acceleration system construction applications, specifically a kind of expansible AI based on OCP standard Acceleration system structure.

Background technique

OCP standard is initiated by Facebook, is dedicated to pushing open hardware technology standard, is improved data center's efficiency and spirit Activity, either density, availability, flexibility and scalability, this platform is compared with traditional machine posture acceleration system structure, tool There is the advantage for being more readily maintained, managing and disposing.

For AI acceleration system, CPU module is one of vital component, in the prior art, in May, 2018, NVIDIA is announced to release computing platform HGX-2 of new generation in the GTC conference of Taiwan, is possessed up to 16 pieces of top calculating and block Tesla V100；However, not occurring the expansible AI acceleration system structure for being directed to OCP standard also at present.

Utility model content

The utility model in order to overcome the deficiencies of the above existing technologies, provides a kind of based on the expansible of OCP standard AI acceleration system structure meets OCP standard criterion, the cabinet of compatible OCP standard, except the expansible connection service of PCIE expansion interface Except device or NIC card, this system is without other external interfaces, system apex outlet, the interconnection of two pieces of back-end realization of HGX-2 mainboard And heat dissipation, it is reasonable in design, there are preferable extension elasticity and good market prospects.

The technical solution adopted by the utility model to solve its technical problems is that:

A kind of expansible AI acceleration system structure based on OCP standard, including acceleration system cabinet, the cabinet are divided into Lower two layers, further include two carrying 8GPU HGX-2 mainboard, for the interconnection of two pieces of HGX-2 mainboards Nvlink interconnection board, be used for Fan Transit Unit, power supply interface and the main pinboard for carrying BMC managing chip interconnected between multiple fan mould groups, the cabinet are high Spending is 6U, depth 800mm, and two pieces of HGX-2 mainboards are separately positioned on the upper layer and lower layer of cabinet, two pieces of HGX-2 masters Plate is interconnected in the rear end of cabinet by Nvlink interconnection board, and the Nvlink interconnection board is arranged in vertical direction, and Nvlink is mutual The Fan Transit Unit parallel with Nvlink interconnection board is equipped with behind yoke plate, the Nvlink interconnection board and the Fan Transit Unit are logical Cross I sprig connection, the Fan Transit Unit be equipped with fan management device and several fan mould groups, the fan management device with Several fan mould group electrical connections, the fan management device are electrically connected with the main pinboard；Two pieces of HGX-2 masters The front end of plate is connected separately with one piece of main pinboard, and several 2.5 inch hards and several are equipped on every piece of main pinboard PCIE expansion interface；The power supply interface is electrically connected with main pinboard, two pieces of HGX-2 mainboards and fan mould group, the fan Mould group and 2.5 inch hard quantity are respectively eight.

Power supply interface is electrically connected with the copper bar of cabinet, and HGX-2 mainboard can be directly electric with the copper bar of cabinet when being inserted into cabinet Connection, to realize the electrical connection of HGX-2 mainboard and power supply interface.

Preferably, the PCIE expansion interface is four, and two of them PCIE expansion interface supports signal switching card, in addition Two PCIE expansion interfaces support 100G NIC card.

Preferably, between the HGX-2 mainboard and the Nvlink interconnection board, the HGX-2 mainboard and the main pinboard Connected by ExaMAX high speed connector.

Preferably, the outside of acceleration system cabinet rear end is arranged in the fan mould group.

Preferably, power-assisted handle is equipped with HGX-2 mainboard connecting pin opposite side on the main pinboard.

The beneficial effects of the utility model are:

1, the utility model provides the AI acceleration system structure based on OCP and HGX-2 mainboard, has enhanced scalability, machine Case height is 6U, depth 800mm, meets OCP standard criterion, the cabinet of compatible OCP standard, and intermodule is designed using plug, Management is maintained easily, HGX-2 mainboard is electrically connected by copper bar with power supply interface, and installation operation is convenient；Fan management device and master turn Fishplate bar electrical connection, several fan mould groups of Collaborative Control；By connecting Nvlink interconnection board and Fan Transit Unit I sprig, protect Card Nvlink interconnection board is closely connect with Fan Transit Unit, improves the heat dissipation effect to Nvlink interconnection board and HGX-2 mainboard, Nvlink interconnection board is vertically arranged with Fan Transit Unit, is reduced and is occupied chassis space；Several 2.5 inches are equipped on main pinboard Hard disk provides enough memory spaces for system；In addition to the expansible connection server of PCIE expansion interface or NIC card, this is System is closed without other external interfaces, system apex outlet, the interconnection and heat dissipation of two pieces of back-end realization of HGX-2 mainboard, structure design Reason, can satisfy the acceleration demand of different user, have preferable extension elasticity and good market prospects.

2, the PCIE expansion interface of two support signal switching cards with the signal of other servers for connecting, two supports The PCIE expansion interface of 100G NIC card has preferable extension elasticity for extending NIC card.

3, ExaMAX high speed connector stable connection, data transmission bauds are fast.

4, the outside of acceleration system cabinet rear end is arranged in fan mould group, is not take up enclosure interior.

5, since the connection of ExaMAX high speed connector is relatively firm, inserting for main pinboard is facilitated by the setting of power-assisted handle It pulls out.

Detailed description of the invention

Fig. 1 is the structural schematic diagram of the utility model；

Fig. 2 is the top view of the utility model；

Fig. 3 is the side perspective view of the utility model；

Fig. 4 is the rearview of the utility model；

Fig. 5 is the structural schematic diagram of Nvlink interconnection board and fan management plate；

Fig. 6 is the structural schematic diagram of main pinboard；

In figure: 1, cabinet, 2, HGX-2 mainboard, 3, Nvlink interconnection board, 4, Fan Transit Unit, 41, fan mould group, 42, wind Fan manager, 5, power supply interface, 6, main pinboard, 61,2.5 inch hards, 62, PCIE expansion interface, 63, power-assisted handle, 7, ExaMAX high speed connector.

Specific embodiment

The utility model in order to better understand explains in detail the embodiments of the present invention with reference to the accompanying drawing.

As shown in figures 1 to 6, a kind of expansible AI acceleration system structure based on OCP standard, including acceleration system cabinet 1, It further includes the HGX-2 mainboard 2 of two carrying 8GPU for upper layer and lower layer that the cabinet 1, which divides, is interconnected for two pieces of HGX-2 mainboards 2 Nvlink interconnection board 3, the Fan Transit Unit 4 for interconnecting between multiple fan mould groups 41, power supply interface 5 and carrying BMC manage core The main pinboard 6 of piece, 1 height of cabinet are 6U, depth 800mm, and two pieces of HGX-2 mainboards 2 are separately positioned on cabinet 1 Upper layer and lower layer, two pieces of HGX-2 mainboards 2 are interconnected in the rear end of cabinet 1 by Nvlink interconnection board 3, and the Nvlink is mutual Yoke plate 3 is arranged in vertical direction, and the Fan Transit Unit 4 parallel with Nvlink interconnection board 3 is equipped with behind Nvlink interconnection board 3, The Nvlink interconnection board 3 is connect with the Fan Transit Unit 4 by I sprig, and the Fan Transit Unit 4 is equipped with fan management Device 42 and several fan mould groups 41, the fan management device 42 is electrically connected with several fan mould groups 41, described Fan management device 42 is connect with the main pinboard electricity 6；The front end of two pieces of HGX-2 mainboards 2 is connected separately with one piece of master and turns Fishplate bar 6 is equipped with several 2.5 inch hards 61 and several PCIE expansion interfaces 62 on every piece of main pinboard 6；The power supply Interface 5 is electrically connected with 6, two pieces of HGX-2 mainboards 2 of main pinboard and fan mould group 41, and the fan mould group 41 and 2.5 inches are hard 61 quantity of disk is respectively eight.

Power supply interface 5 is electrically connected with the copper bar of cabinet 1, the copper that HGX-2 mainboard 2 can directly with cabinet 1 when being inserted into cabinet 1 Row's electrical connection, to realize the electrical connection of HGX-2 mainboard 2 and power supply interface 5.

The utility model provides the AI acceleration system structure based on OCP and HGX-2 mainboard 2, has enhanced scalability, cabinet Height is 6U, depth 800mm, meets OCP standard criterion, the cabinet of compatible OCP standard, intermodule is using plug design, side Just maintenance management, HGX-2 mainboard 2 are electrically connected by copper bar with power supply interface 5, and installation operation is convenient；

Fan management device 42 is electrically connected with main pinboard 6, several fan mould groups 41 of Collaborative Control；By the way that Nvlink is interconnected Plate 3 is connect with Fan Transit Unit 4 with I sprig, guarantees that Nvlink interconnection board 3 is closely connect with Fan Transit Unit 4, raising pair The heat dissipation effect of Nvlink interconnection board 3 and HGX-2 mainboard 2, Nvlink interconnection board 3 are vertically arranged with Fan Transit Unit 3, are reduced Occupy chassis space；It is equipped with several 2.5 inch hards 61 on main pinboard 6, provides enough memory spaces for system；

In addition to the expansible connection server of PCIE expansion interface 62 or NIC card, this system is without other external interfaces, system Apex outlet, the interconnection and heat dissipation of two pieces of back-end realization of HGX-2 mainboard 2 are reasonable in design, can satisfy different user Acceleration demand has preferable extension elasticity and good market prospects.

The PCIE expansion interface 62 is four, and two of them PCIE expansion interface 62 supports signal switching card, in addition two A PCIE expansion interface 62 supports 100G NIC card, the PCIE expansion interface 62 of two support signal switching cards to be used for and other clothes The signal connection of business device, the PCIE expansion interface 62 of two support 100G NIC cards have preferably extension for extending NIC card Elasticity.

It is logical between the HGX-2 mainboard 2 and the Nvlink interconnection board 3, the HGX-2 mainboard 2 and the main pinboard 6 It crosses ExaMAX high speed connector 7 to connect, stable connection, data transmission bauds is fast.

The outside of 1 rear end of acceleration system cabinet is arranged in the fan mould group 41, is not take up 1 inner space of cabinet.

It is equipped with power-assisted handle 63 with 2 connecting pin opposite side of HGX-2 mainboard on the main pinboard 6, due to ExaMAX high Fast connector 7 connects relatively firm, facilitates the plug of main pinboard 6 by the setting of power-assisted handle 63.

The system when in use, can connect or expand with server as desired by the PCIE expansion interface 62 of system front end NIC card is opened up, in addition to the expansible connection server of PCIE expansion interface 62 or NIC card, this system is without other external interfaces, system Apex outlet, the interconnection and heat dissipation of two pieces of back-end realization of HGX-2 mainboard 2, power supply interface 5 are electrically connected with the copper bar of cabinet 1, HGX-2 mainboard 2 can be directly electrically connected with the copper bar of cabinet 1 when being inserted into cabinet 1, to realize HGX-2 mainboard 2 and power supply interface 5 electrical connection, this system structure is reasonable in design, can satisfy the acceleration demand of different user, has preferable extension bullet Property and good market prospects.

For the ordinary skill in the art, introduction according to the present utility model is not departing from the utility model Principle and spirit in the case where, the changes, modifications that are carried out to embodiment, replacement and variant still fall within the utility model Within protection scope.

Claims (6)

1. a kind of expansible AI acceleration system structure based on OCP standard, including acceleration system cabinet, the cabinet is divided into up and down Two layers, which is characterized in that further include two carrying 8GPU HGX-2 mainboard, for two pieces of HGX-2 mainboards interconnection Nvlink it is mutual Yoke plate, the Fan Transit Unit for being interconnected between multiple fan mould groups, power supply interface and the main pinboard for carrying BMC managing chip, Two pieces of HGX-2 mainboards are separately positioned on the upper layer and lower layer of cabinet, and two pieces of HGX-2 mainboards pass through in the rear end of cabinet Nvlink interconnection board interconnection, the Nvlink interconnection board vertical direction be arranged, and behind Nvlink interconnection board be equipped with The parallel Fan Transit Unit of Nvlink interconnection board, the Nvlink interconnection board are connect with the Fan Transit Unit, and the fan turns Fishplate bar is equipped with fan management device and several fan mould groups, and the fan management device is electrically connected with several fan mould groups It connects, the fan management device is electrically connected with the main pinboard；The front end of two pieces of HGX-2 mainboards is connected separately with one piece of master Pinboard is equipped with several 2.5 inch hards and several PCIE expansion interfaces on every piece of main pinboard；The power supply interface It is electrically connected with main pinboard, two pieces of HGX-2 mainboards and fan mould group.

2. a kind of expansible AI acceleration system structure based on OCP standard as described in claim 1, which is characterized in that described PCIE expansion interface is four, and two of them PCIE expansion interface supports signal switching card, other two PCIE expansion interface branch Hold 100G NIC card.

3. a kind of expansible AI acceleration system structure based on OCP standard as described in claim 1, which is characterized in that described It is connected by ExaMAX high speed between HGX-2 mainboard and the Nvlink interconnection board, the HGX-2 mainboard and the main pinboard Device connection.

4. a kind of expansible AI acceleration system structure based on OCP standard as described in claim 1, which is characterized in that described The outside of acceleration system cabinet rear end is arranged in fan mould group.

5. a kind of expansible AI acceleration system structure based on OCP standard as described in claim 1, which is characterized in that described Power-assisted handle is equipped with HGX-2 mainboard connecting pin opposite side on main pinboard.

6. a kind of expansible AI acceleration system structure based on OCP standard as described in claim 1, which is characterized in that described Nvlink interconnection board is connect with the Fan Transit Unit by I sprig.