CN108897706A - Accelerator interface - Google Patents

Accelerator interface Download PDF

Info

Publication number
CN108897706A
CN108897706A CN201810440616.9A CN201810440616A CN108897706A CN 108897706 A CN108897706 A CN 108897706A CN 201810440616 A CN201810440616 A CN 201810440616A CN 108897706 A CN108897706 A CN 108897706A
Authority
CN
China
Prior art keywords
accelerator
ddr
storage location
internal storage
server
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.)
Granted
Application number
CN201810440616.9A
Other languages
Chinese (zh)
Other versions
CN108897706B (en
Inventor
林琦
杨艳萍
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Rongxin Micro Technology Co ltd
Original Assignee
Beijing Weimi Technology Development Co ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Beijing Weimi Technology Development Co ltd filed Critical Beijing Weimi Technology Development Co ltd
Priority to CN201810440616.9A priority Critical patent/CN108897706B/en
Publication of CN108897706A publication Critical patent/CN108897706A/en
Application granted granted Critical
Publication of CN108897706B publication Critical patent/CN108897706B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F13/00Interconnection of, or transfer of information or other signals between, memories, input/output devices or central processing units
    • G06F13/14Handling requests for interconnection or transfer
    • G06F13/16Handling requests for interconnection or transfer for access to memory bus

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Multi Processors (AREA)
  • Advance Control (AREA)

Abstract

The invention relates to an accelerator interface, which solves the technical problems of low applicability, low transmission rate and high delay, and is connected between a server and an accelerator by adopting the accelerator interface; the accelerator interface is a DDR interface which comprises a DDR storage module and a high-speed selector switch for controlling the sharing or disconnection of DDR memory units, the accelerator interface adopts the technical scheme that DDR addresses and command signal sequences are adopted to complete data communication between a server and an accelerator, the problem is solved well, and the accelerator interface can be used for connecting the accelerator and the server or the server.

Description

A kind of accelerator interfaces
Technical field
The present invention relates to accelerator fields, and in particular to a kind of accelerator interfaces.
Background technique
Artificial intelligence needs to carry out machine learning, is quickly calculated with intelligent algorithm.General purpose computer is difficult to meet, and needs Various types of accelerators.Accelerator requires have high transmission rates and low-latency.Graphics processor(Graphics Processing Unit, abbreviation GPU)And field programmable gate array(Field-Programmable Gate Array, letter Claim FPGA)It is most important algorithm accelerator.
The interface that current above-mentioned accelerator and server use is bus interface PCI express, bus interface PCI The maximum bandwidth of express is 16MB/s.Also, the PCI express of high bandwidth is generally used at high-end server central Manage device(Central Processing Unit, abbreviation CPU).Therefore, there are high performance accelerators for existing accelerator interfaces Interface suitability is poor, and that there are transmission rates is low, postpones high technical problem.Therefore it provides one kind can be adapted to extensively it is various The high-performance accelerator interfaces of server and server are with regard to necessary.
Summary of the invention
The technical problem to be solved by the present invention is to applicability existing in the prior art is low, transmission rate is low, and delay is high The technical issues of.A kind of new accelerator interfaces are provided, which has applicability wide, and transmission rate is high, postpones low The characteristics of.
In order to solve the above technical problems, the technical solutions adopted are as follows:
A kind of accelerator interfaces, the accelerator interfaces are connected between server and accelerator;The accelerator interfaces are DDR Interface, ddr interface include DDR memory module, and the high speed COMS bus exchange switch that control DDR internal storage location is shared or disconnected, described Accelerator interfaces complete the data communication of server and accelerator using the address DDR and command signal sequence.
In above scheme, for optimization, further, the circuit that switches at high speed is high-speed analog switch.
Further, the DDR internal storage location is dynamic random access memory or dual inline memory module.
Further, the capacity of the dynamic random access memory is 4GB-8GB.
Further, the data communication of the server and accelerator is both-way communication, the memory size of DDR internal storage location For the communal space.
Further, the DDR internal storage location is shared need to execute the following steps:
Step 1, server executes acceleration task, and task data is written;
Step 2, Acceleration of starting device interface, accelerator knot carry out the confirmation of State tracking logic commands match, fit through, execute Step 3, it fails to match then repeats step 2;
Step 3, idle DDR internal storage location switching is waited;
Step 4, switch Accelerator control DDR internal storage location;
Step 5, acceleration task is handled;
Step 6, switching server controls DDR internal storage location, and server real-time query accelerates task status, and task status is not complete Cheng Ze returns to step 2, and task status is shared to complete then to terminate DDR internal storage location.
Further, the State tracking logic commands match, which confirms, includes:
Step A, DDR internal storage location reads the order of column address strobe time delay, tracking mode;
Step B is tracked and is successfully then executed output matching confirmation instruction;
Step C, tracking are completed, and column address strobe time delay command parameter is restored, and execute State tracking logic next time Commands match.
Beneficial effects of the present invention:
Effect one, the present invention is used between the ddr interface access server and accelerator of high bandwidth, and realizes ddr interface Memory sharing, provide more high bandwidth, the lower accelerator interfaces of delay;It can provide 64 DDR4 SODIMM, 26.6GB/ S), delay time also reducible 10 times of reduction;
Effect two, ddr interface exist in all universal cpus, and applicability than before only deposit by high performance accelerator interfaces It is more powerful in high-performance CPU.
Detailed description of the invention
Present invention will be further explained below with reference to the attached drawings and examples.
Fig. 1, the connection schematic diagram of accelerator interfaces.
Fig. 2, order tracking and reception schematic diagram.
Fig. 3, DDR internal storage location share schematic diagram.
The interface diagram of Fig. 4, DDR4.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to embodiments, to the present invention It is further elaborated.It should be appreciated that described herein, specific examples are only used to explain the present invention, is not used to limit The fixed present invention.
Embodiment 1
The present embodiment provides a kind of accelerator interfaces, such as Fig. 1, the accelerator interfaces are connected between server and accelerator; The accelerator interfaces are ddr interface, and ddr interface includes DDR memory module, and control DDR internal storage location is shared or disconnects High speed COMS bus exchange switch, the accelerator interfaces complete the number of server and accelerator using the address DDR and command signal sequence According to communication.
The accelerator of the present embodiment uses FPGA, the alternative DRAM of DDR internal storage location.Remaining is similar to the present embodiment, no longer Description.DDR3 or DDR4 can be used in ddr interface.
Since ddr interface does not interrupt interaction mechanism, the present embodiment uses the address DDR, command signal special sequence It does server to communicate with accelerator, this special sequence is the unique sequences and original using the CAS LATENCY setting in MR0 There is the normal work of DRAM not conflict, the setting of original CAS LATENCY can be restored to after the sequence ends.Utilize LOAD It is as follows that accelerator switching start command is done in order:
DDR3:MR0[6:4] [2] CAS Latency sequence:5,14,6,13,9.
DDR4:MR0 [12] [6] [4] [2] CAS Latency sequence:10,32,12,30,24.
In addition, the clock of server and accelerator can not be completely the same, therefore, switching time will meet DDR3/DDR4 Code requirement.
Wherein, the circuit that switches at high speed is high-speed analog switch.The parameter of high-speed analog switch is primarily upon high speed mould The switch speed of quasi- switch.Very quickly, the present embodiment is using SGM4717 high speed for the switch speed of existing analog switch Dual-channel analog switch.
In detail, the DDR internal storage location is dynamic random access memory or dual inline memory module.
Wherein, the dynamic random access memory, the i.e. capacity of DRAM are 4GB-8GB.
Wherein, the data communication of the server and accelerator is both-way communication, and the memory size of DDR internal storage location is total Enjoy space.
In addition, the DDR internal storage location is shared need to execute the following steps such as Fig. 3:
Step 1, server executes acceleration task, and task data is written;
Step 2, Acceleration of starting device interface, accelerator knot carry out the confirmation of State tracking logic commands match, fit through, execute Step 3, it fails to match then repeats step 2;
Step 3, idle DDR internal storage location switching is waited;
Step 4, switch Accelerator control DDR internal storage location;
Step 5, acceleration task is handled;
Step 6, switching server controls DDR internal storage location, and server real-time query accelerates task status, and task status is not complete Cheng Ze returns to step 2, and task status is shared to complete then to terminate DDR internal storage location.
Wherein, such as Fig. 2, the State tracking logic commands match, which confirms, includes:
Step A, DDR internal storage location reads the order of column address strobe time delay, and tracking mode successively tracks each state;
Step B is tracked and is successfully then executed output matching confirmation instruction;
Step C, tracking are completed, and column address strobe time delay command parameter is restored, and execute State tracking logic next time Commands match.
Wherein, the Interface status of DDR4 such as Fig. 4.
Although the illustrative specific embodiment of the present invention is described above, in order to the technology of the art Personnel are it will be appreciated that the present invention, but the present invention is not limited only to the range of specific embodiment, to the common skill of the art For art personnel, as long as long as various change the attached claims limit and determine spirit and scope of the invention in, one The innovation and creation using present inventive concept are cut in the column of protection.

Claims (7)

1. a kind of accelerator interfaces, it is characterised in that:The accelerator interfaces are connected between server and accelerator;It is described to add Fast device interface is ddr interface, and ddr interface includes DDR memory module, and the high speed that control DDR internal storage location is shared or disconnected Switching switch, the accelerator interfaces complete the data communication of server and accelerator using the address DDR and command signal sequence.
2. accelerator interfaces according to claim 1, it is characterised in that:The circuit that switches at high speed is opened for High Speed Analog It closes.
3. accelerator interfaces according to claim 1, it is characterised in that:The DDR internal storage location is dynamic randon access Memory or dual inline memory module.
4. accelerator interfaces according to claim 3, it is characterised in that:The capacity of the dynamic random access memory is 4GB-8GB。
5. accelerator interfaces according to claim 1, it is characterised in that:The data communication of the server and accelerator is Both-way communication, the memory size of DDR internal storage location are the communal space.
6. accelerator interfaces according to claim 1, it is characterised in that:The DDR internal storage location is shared need to execute it is following Step:
Step 1, server executes acceleration task, and task data is written;
Step 2, Acceleration of starting device interface, accelerator knot carry out the confirmation of State tracking logic commands match, fit through, execute Step 3, it fails to match then repeats step 2;
Step 3, idle DDR internal storage location switching is waited;
Step 4, switch Accelerator control DDR internal storage location;
Step 5, acceleration task is handled;
Step 6, switching server controls DDR internal storage location, and server real-time query accelerates task status, and task status is not complete Cheng Ze returns to step 2, and task status is shared to complete then to terminate DDR internal storage location.
7. accelerator interfaces according to claim 6, it is characterised in that:The State tracking logic commands match confirmation packet It includes:
Step A, DDR internal storage location reads the order of column address strobe time delay, tracking mode;
Step B is tracked and is successfully then executed output matching confirmation instruction;
Step C, tracking are completed, and column address strobe time delay command parameter is restored, and execute State tracking logic next time Commands match.
CN201810440616.9A 2018-05-10 2018-05-10 Accelerator interface Active CN108897706B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810440616.9A CN108897706B (en) 2018-05-10 2018-05-10 Accelerator interface

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810440616.9A CN108897706B (en) 2018-05-10 2018-05-10 Accelerator interface

Publications (2)

Publication Number Publication Date
CN108897706A true CN108897706A (en) 2018-11-27
CN108897706B CN108897706B (en) 2021-07-23

Family

ID=64342723

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810440616.9A Active CN108897706B (en) 2018-05-10 2018-05-10 Accelerator interface

Country Status (1)

Country Link
CN (1) CN108897706B (en)

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20020174290A1 (en) * 2001-05-15 2002-11-21 Wu Kun Ho Memory accelerator, acceleration method and associated interface card and motherboard
US20110010472A1 (en) * 2008-02-27 2011-01-13 Se Jin Kang Graphic accelerator and graphic accelerating method
CN104820657A (en) * 2015-05-14 2015-08-05 西安电子科技大学 Inter-core communication method and parallel programming model based on embedded heterogeneous multi-core processor
CN105677595A (en) * 2016-01-21 2016-06-15 方一信息科技(上海)有限公司 FPGA method achieving computation speedup and PCIESSD storage simultaneously
CN106371807A (en) * 2016-08-30 2017-02-01 华为技术有限公司 Method and device for extending processor instruction set
CN106814662A (en) * 2015-11-30 2017-06-09 三星电子株式会社 The method of Accelerator control device and control accelerator logic
CN108874684A (en) * 2018-05-31 2018-11-23 北京赫芯斯信息技术有限公司 Split the NVDIMM interface data read-write equipment of CACHE caching

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20020174290A1 (en) * 2001-05-15 2002-11-21 Wu Kun Ho Memory accelerator, acceleration method and associated interface card and motherboard
US20110010472A1 (en) * 2008-02-27 2011-01-13 Se Jin Kang Graphic accelerator and graphic accelerating method
CN104820657A (en) * 2015-05-14 2015-08-05 西安电子科技大学 Inter-core communication method and parallel programming model based on embedded heterogeneous multi-core processor
CN106814662A (en) * 2015-11-30 2017-06-09 三星电子株式会社 The method of Accelerator control device and control accelerator logic
CN105677595A (en) * 2016-01-21 2016-06-15 方一信息科技(上海)有限公司 FPGA method achieving computation speedup and PCIESSD storage simultaneously
CN106371807A (en) * 2016-08-30 2017-02-01 华为技术有限公司 Method and device for extending processor instruction set
CN108874684A (en) * 2018-05-31 2018-11-23 北京赫芯斯信息技术有限公司 Split the NVDIMM interface data read-write equipment of CACHE caching

Also Published As

Publication number Publication date
CN108897706B (en) 2021-07-23

Similar Documents

Publication Publication Date Title
US6378055B1 (en) Memory accessing and controlling method
US9183169B2 (en) SAS expander based persistent connections
US20150262633A1 (en) Dual-port ddr4-dimms of sdram and nvram for ssd-blades and multi-cpu servers
CN103180817B (en) Memory expansion unit and server
CN103154920B (en) Unified I/O adapter
KR101051506B1 (en) Method and memory controller for scalable multichannel memory access
US10255220B2 (en) Dynamic termination scheme for memory communication
US9946664B2 (en) Socket interposer having a multi-modal I/O interface
CN106844048B (en) Distributed memory sharing method and system based on hardware characteristics
CN105988970B (en) The processor and chip of shared storing data
KR20110089321A (en) Method and system for improving serial port memory communication latency and reliability
EP1652058A4 (en) Switch/network adapter port incorporating selectively accessible shared memory resources
KR20080047998A (en) Apparatus and method for switching an apparatus to a power saving mode
CN114662136B (en) PCIE (peripheral component interface express) channel-based high-speed encryption and decryption system and method for multi-algorithm IP (Internet protocol) core
US20070204091A1 (en) Single Bus Command for Transferring Data in a Processing System
CN112506823B (en) FPGA data reading and writing method, device, equipment and readable storage medium
CN109564562B (en) Big data operation acceleration system and chip
US20130227210A1 (en) Memory, memory controllers, and methods for dynamically switching a data masking/data bus inversion input
JP3516431B2 (en) I / O traffic transmission over processor bus
US10990307B2 (en) System and method for providing a configurable storage media interface
US9202541B2 (en) Semiconductor apparatus configured to reduce data processing performance
US5535333A (en) Adapter for interleaving second data with first data already transferred between first device and second device without having to arbitrate for ownership of communications channel
US6633927B1 (en) Device and method to minimize data latency and maximize data throughput using multiple data valid signals
CN108897706A (en) Accelerator interface
CN206133528U (en) RapidIO and SATA switching controller

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
TA01 Transfer of patent application right
TA01 Transfer of patent application right

Effective date of registration: 20210624

Address after: 100085 room 065, 1st floor, building 15, chuangkezhen community, Haidian District, Beijing

Applicant after: Beijing Rongxin Micro Technology Co.,Ltd.

Address before: 100044 room 705, 7th floor, 12 Zhongguancun South Street, Haidian District, Beijing

Applicant before: BEIJING WEIMI TECHNOLOGY DEVELOPMENT Co.,Ltd.

GR01 Patent grant
GR01 Patent grant
TR01 Transfer of patent right
TR01 Transfer of patent right

Effective date of registration: 20211111

Address after: 215000 room 217a, building B, science and technology entrepreneurship Park, Zhangjiagang Free Trade Zone, Suzhou City, Jiangsu Province

Patentee after: Rongxin micro (Suzhou) Electronics Co., Ltd

Address before: Room 106, 1f, building 15, Chuangke town community, Haidian District, Beijing 100085

Patentee before: Beijing Rongxin Micro Technology Co., Ltd

TR01 Transfer of patent right
TR01 Transfer of patent right

Effective date of registration: 20220525

Address after: 100000 room 106, 1f, building 15, chuangkexiaozhen community, Haidian District, Beijing 065

Patentee after: Beijing Rongxin Micro Technology Co.,Ltd.

Address before: 215000 room 217a, building B, science and technology entrepreneurship Park, Zhangjiagang Free Trade Zone, Suzhou City, Jiangsu Province

Patentee before: Rongxin micro (Suzhou) Electronics Co.,Ltd.