RevolutionizingtheDatacenter
Join theConversation#OpenPOWERSummit
ConTuttoDr.ThomasRoewer,IBMT.J.Watson
ResearchCenter
Join theConversation#OpenPOWERSummit
SummaryConTuttoisanFPGA-basedmemorycardthatplugsintotheDMImemoryslotofanIBMPower8server.§ Differentmemorytechnologiescanbeusedin
Powersystem• DRAM• MRAM• NVDIMM
§ Highestbandwidth&lowestlatencyFPGAattachpointinanycomputersystem
§ Near-memoryacceleration(JanvanLunteren)
26/14/16
P8MemorySub-System
6/14/16 3
§ 8DMIlinksavailableonaP8Dual-Chip-Module
§ DifferentialMemoryInterface(DMI)high-speedlinksconnecttoamemorybufferASICs
§ 4memorycontrollerspermemorybufferASIC
§ ASICandDRAMchipsareco-locatedonacustomDIMM(CDIMM)
§ 32memorycontrollersavailabletoP8
Power8DCM
Mem
ory
Buffer
Mem
ory
Buffer
Mem
ory
Buffer
Mem
ory
Buffer
Mem
oryBuffer
Mem
oryBuffer
Mem
oryBuffer
Mem
oryBuffer
DMI-Links
P8MemorySub-SystemwithConTutto
6/14/16 4
§ BuiltanFPGA-basedcardthatplugsintotheDMIslot
§ EnablesregularsystemoperationwithanymixofCDIMMsandConTuttocardspopulated
§ FullcompatibilitywithDMIprotocol
§ MemorycontrollersimplementedinfabriclogicandindependentofDMIprotocollogic
§ Flexiblesystemarchitectureenableseasyimplementationofadditionalfeatures
Power8DCM
Mem
ory
Buffer
Mem
ory
Buffer
Mem
ory
Buffer
Mem
ory
Buffer
Mem
oryBuffer
Mem
oryBuffer
Mem
oryBuffer
FPGA
ConTuttoCard
6/14/16 5
§ Intendedtobeanexperimentationandproto-typingvehicle
§ Cardcharacteristics• 10signallayers• 10power/groundlayers
§ PlugcompatiblewithCDIMM,but2.5''higher-- andDIMMsaddwidth
§ LargeAlteraFPGAwithcapacityforadditionalfunctionincorporated
§ CFAM-S(connectiontoserviceprocessor)enablessystemintegration
DMIConnector
Stratix V5SGXA9
DIMMConnectors
PCIe x4
TCAM
JTAG USB
CFAM-S
Logicdesign§ Support forallDMIcommands
implemented§ SomeMemoryBufferASIC
featuresnotavailableininitialversionofConTutto:• NoDMIbussparingorfail-over• NoL4cache• Severalperformance features
notimplemented§ 32:1muxratiobetweenDMI
linkandFPGAfabriclogic§ Usingabout40%ofavailable
logicresourcesinStratix VA9device
§ Avalonbusfordesignmodularity, i.e.plug-and-playmemorycontroller
6/14/16 6
TxPhyDMI
RxPhyDMI
21@8Gbit/sec 14@8Gbit/sec
32x14@250MHz32x21@250MHz
DMILinkProtocolLogicReadMaster WriteMaster
AlteraAvalonBus
MC0 MC1PCIe TCAM
ResultsandOutlook
§ SeeConTuttoattheIBMOpenPower booth§ BoottoLinuxusingDRAM,MRAMandNVDIMMs
pluggedintoConTutto§ 270ns- 500nsadditionalround-triplatencywith
initialConTuttoimplementationcomparedto§ Severalfollow-onresearchprojectsatIBMandat
externalindustrypartnersongoing
6/14/16 7
Acknowledgments
§ ThisworkwaspartiallyfundedbytheDepartmentofEnergyundertheFastForward-2contract.
§ Teammembers:• AdamMcPadden,BharatSukhwani,Kyu-hyoun Kim,
ChuckHaymes,DeanSanner,Sameh Asaad,DanDreps,Hillery Hunterandmanymanymore…....
6/14/16 8