accelerator for power - openpower...

21
Revolutionizing the Datacenter Accelerator For POWER Eric Deng, VP Semptian, China Join the Conversation #OpenPOWERSummit

Upload: others

Post on 04-Feb-2021

0 views

Category:

Documents


0 download

TRANSCRIPT

  • RevolutionizingtheDatacenter

    Join theConversation#OpenPOWERSummit

    AcceleratorForPOWEREric Deng,VP

    Semptian,China

    Join theConversation#OpenPOWERSummit

  • Who is Semptian

    23/25/16

    was founded in 2003, a leading solution and serviceprovider in big data acquisition and analyzing in China.

    HQ in Shenzhen, China.

    Go to public in 2014 (NEEQ: 831196)

  • SemptianOverview

    33/25/16

    Our Vision

    Data creates value, changes work, lifestyle, study and entertainment. Data makes life easier, quicker and happier.

    60+

    260

    $40M

    80%

    260+ employees50% engineers30% salesR&D expense 15%+

    2015 Growth

    Revenue in 2015

    60+ patents10+ growth per year

  • Our Goal

    43/25/16

    To provide board accelerator for POWER with FPGA-basedheterogeneous computing

    Memory

    PowerProcessor UltraScaleFPGA

    PCIe/CAPI

    Massivelyparallelacceleratorü ComparedwithCPU,Performanceper

    Wattincreasedby22to25timesü ComparedwithCPU,Latencyreduce

    to 1/50ü HighIOintegration(PCIe,DDR4,etc.)

    * FromtheHotCloud 2013document“Achieving10Gbpsline-ratekey-valuestoreswithFPGAs”

  • SemptianNSAAccelerator

    53/25/16NSA:Network&ServiceAccelerator

    Technical Stack

    POWER Server

    OS

    APPs

    Service

    FPGA

    NSA

    IPXilinx, 3rd party

    IPSemptian

    We’re here

  • MakePOWERmorePowerful-CAPIFPGA

    6

    IBMSuppliedPOWERServiceLayer

    CustomerAccelerator

    Function(AFU)

    FPGACAPI

    NSA

  • CAPIadvantages

    73/25/16

    lEasier,MoreNaturalProgrammingModel–Traditionalthreadlevelprogramming–LonglatencyofI/Otypicallyrequiresrestructuringofapplication

    lEnablesApplicationsNotPossibleonI/O–Pointerchasing,etc…

    lVirtualAddressing &DataCaching–SharedMemory–Lowerlatencyforhighlyreferenceddata

  • NSAAcceleratorFeatures

    83/25/16

    HighPerformanceLowPower

    SDA

    HardwareLooseCoupling

    UpgradeOnline

    3rd PartyIPProtection

  • HighPerformance&LowPower

    93/25/16

    0

    20

    40

    60

    80

    100

    FPGA CPU

    30

    95

    Power

    Power(W)

    0

    2

    4

    6

    8

    10

    HadoopEC GZIP ImageResizing

    Performance

    CPU FPGA

    CPU FPGA Speedup

    HadoopEC 400MB/s 3GB/s 7.5x

    GZIP 300MB/s 3GB/s 10x

    ImageResizing 7sec/1kpcs 0.69sec/1kpcs 10x

  • HardwareLooseCoupling

    103/25/16

    01Additional rackspace: NO

    02PCIe 3.0in normal serverchassis

    03Additional cooling: NO

    04Externalpowersupply: NO

  • SoftwareDefinedAccelerator

    113/25/16

    LibrariesCompiler Debugger Profiler

    OpenCL,C,C++Code

    CAPI/PCIe

    SoftwareengineercandevelopFPGA

  • AccelerationsUpgradeOnline

    123/25/16

    SAFEimage

    AFUimage

    Update online

    Power8 Processor NSA FPGA Accelerator Flash

    Configuration

    Updatefrom theHostCPU,noneedtheJTAGCableYoucanchangetheAFUsasyouneedEventheupdatefailed,FPGAcanloadSafeimage

  • 3rd Party IP Protection

    133/25/16

    UniqueID

    Initial Verification

    FPGA Security Chip

    • The dedicated security chip contains an unique ID for every board.• Verification can be executed on initialization to enableworkingonly on

    the licensed board.

    ID-bindedImage

  • End-to-End delivery for you

    143/25/16

    Cards

    Chassis

    Boxes

    Over 12 years experience in COTS and customized products

  • End-to-End delivery for you

    153/25/16

    • High efficiency and resilience supply• 10K+ cards and appliances per year

  • ErasureCodeforHadoopHDFS

    163/25/16

    ReplicationStrategy:300%StoragecapacityrequiredPower,cooling, costchallenge

    RSCode(10,4)

    Hadoop3.0

    ErasureCodeRS(10,4):• 140%Storagecapacityrequired: 50% cost

    reduction• Power,cooling improved• DatareliabilityimprovedCodingperformancechallenges• Java:~8MB/s/core• optimizedassemblecode~100MB/s/core

  • ErasureCodeforHDFS

    173/25/16

    IBM Power 822L withSemptian NSA

    Performance– 7X~10Xperformanceimprove: speedupto3GB/s– CPU offloading: equivalentto20~30CPUcores

    capability– CAPIinterfaceenableeasyprogramandlowlatency

    Scalability– Using1ormorePCIe/CAPIslots

    Testedonplatform:– IBMPower8server822L– RedPower serverfromZoomTech

    RedPower withSemptian NSA

  • ImageResizing

    183/25/16

    Rawimage

    resizing1

    resizing2

    resizing3

    Commonapplicationto adapt to various terminals

  • ImageResizing

    193/25/16

    § Advantage:• Noextramaintainingcost.• 3~10.7Xprocessingspeed(1000

    imagesprocesstimeis0.65Svs.7.0SbydualE5-2630)

    • Sameworkflowprocess• ReleaseCPUpower

    § Cost:• OneNSA-120 card• Immigrationcost(~0)——from

    currentImageMagick solutiontoCIPsolution (100%compatibletoImageMagick)

    • HW: NSA-120• SW: CIP (Compatible to ImageMagick

    and GraphicsMagick)

    device Watt Speed Qty USD/year

    DualXEON E5-2630Server 500 1 10.7 5000

    Server +NSA 500+15 10.7 1 482

    Severalcyclestodecodeonepixel

    CPU Onecycledecodes64pixels

    NSASaveEnergy

    >10x

    Notes:u Assume electricity cost: $0.1068/kwh, 24x7 operation

    10.7xSpeedup:1000 imagesprocesstimeis0.65Svs.7.0SbydualE5-2630

    u Data based on Energy Information Administration US, accesstime June 18, 2014.

  • Contact

    203/25/16

    Our booth:#1013

    SemptianTechnologies Co.,Ltd

    Website:http://www.semptian.com

    E-mail: [email protected]

    Tel: 0086-0755-86656060Fax: 0086-0755-86656090

  • 213/25/16

    www.semptian.com

    Thank You!