IBM GPU解决方案.ppt
GPU介绍,TSS System X STGHao Changjie,2,GPU产品IBM idataplex的GPU解决方案IBM Bladecenter的GPU解决方案,Tesla,Tesla C1060933 Gigaflop SP78 Gigaflop DP4GB 内存,Tesla C2050515 Gigaflop DP3 GB 内存ECC,7 倍峰值双精度性能,Tesla C2070515 Gigaflop DP6 GB 内存ECC,大型数据集,单精度价格性能,免责声明：性能规范可能会有变化,4,GPU Parallel Computing Developer Eco-System,Solution Providers,5,GPU介绍IBM idataplex的GPU解决方案IBM Bladecenter的GPU解决方案,6,Top ViewTop View,Standard 19”Rack,iDataPlex Rack,iDataPlex Rack design:Rack is rotated 90Half-depth,front-access servers.Low airflow impedance.Servers located side-by-side.Doubles server density in similar footprint.Greater cross-section for RDHX.100U Rack84U for Nodes,etc.16U(vertical pockets).Space-saving footprint1200mm wide x 600mm deep.,7,3U Chassis,2U Chassis,Web Server,Storage Drives&Options,I/O Tray,Storage Tray,HPC Server,Fast,Cool,Dense,Flexible.TCO without compromise!,8,4-2.5”SS SAS 6Gbps(or SATA,or 3.5”,or SSD),Infiniband DDR(or QDR,or 10GbE),NVIDIA Tesla M2050#1(or NVIDIA Tesla M1060,or FX3800,or Fusion IO,or),NVIDIA Tesla M2050#2(or NVIDIA Quadro FX3800,or Fusion IO,or),Server level valueEach server is individually serviceableEach GPU is individually replaceable6Gbps SAS drives and controller for maximum performanceService and support for server and GPU from IBM,dx360 M3 Refresh-Server GPU ConfigurationAnnounce May 18,2010,9,Increase Node Density with GPU Acceleration,1 TeraFlop,1 TeraFlop,Consolidate node density 8:1 for 72%less power consumption in flops/wattReduce acquisition costs by 65%Nearly 10 x more performance per server(1),8 dx360 M3 Servers(Xeon 5600),1 dx360 M3 server(Xeon 5600)2 NVIDIA Tesla M2050 GPUs,Based on 2.93GHz Intel with NVidia performance data,2010,10,10,IBM BladeCenter GPU Expansion Blade-Mechanical Design,Connecting to IBM HS22 blade,IBM BladeCenter GPU Expansion Blade(BGE),IBM HS22 blade,11,Increase Solution Density and reduce power costs,44 TeraFlops,44 TeraFlops,Consolidate footprint density 4:1Reduce power and cooling operating costs by X%Nearly 4x more performance per rack(1),336 dx360 M3 Servers(Xeon 5600 Processor)4 Racks,16 InfiniBand Switches,42 dx360 M3 Servers(Xeon 5600 Processor)1 Rack,2 InfiniBand Switches84 NVIDIA Tesla GPUs,12,NVIDIA M series Fermi vs.C series,M(Module)solutions differ from C(Card)solutions in several ways:Ms are passive;Cs are actively cooledMs dont have I/O connectors,Cs do(DVI)Ms can support OpenGL for rendering accelerated engine,Cs do not.Ms for data center products like iDataPlex and BladeCenter,Tesla S2050/S2070 1U System,Tesla C2050/C2070Workstation Board,Tesla M2050/M2070 Adapter,数据中心产品,工作站产品,13,Current,1Q10,2Q10,IBM System x iDataPlex Server&NVIDIA Graphics Roadmap,3Q10,4Q10,dx360 M2 Refresh 1 2 Socket Intel Nehalem-EP16 DIMM Slots(128GB Max)NVIDIA FX3800Fusion I/OVMWare support,dx360 M3Ann Mar,GA MarWestmere EnablementNVIDIA FX3800N+N Redundant PowerHard drive capacity increasesTrusted Platform ModuleFCoE via 10G CNA,dx360 M3 Refresh Announce May 18,SS Jul 13 slot riser for 2 GPU+HBASupport for 2 NVIDIA M1060 or M2050Pre-integrated 10Gb/IB models550W High Efficiency supplyHDD Simplification,600GB HDD Rack-level power capping,NVIDIA Tesla M2050GA May 2010 1 Fermi GPU3 GB Memory225 Watt Double Wide/Dual SlotPassive Cooling,iDataPlex Server Platform,NVIDIA GPU,NVIDIA Quadro FX3800 1 GB Memory107 WattSingle Wide/Single Slot,NVIDIA Tesla M2070GA August 2010 1 Fermi GPU6 GB Memory225 Watt Double Wide/Dual Slot Passive Cooling,NVIDIA Next Gen TeslaNext GenStage Passive Cooling,NVIDIA Tesla M1060In Production1 GT200 GPU4 GB Memory190 Watt Double Wide/Dual SlotPassive Cooling,dx360 M3 RefreshIntelligent Clusters 10BPlanned Ann October,SS Nov3 slot riser for 2 GPU+HBASupport for 2 NVIDIA M2070,14,GPU介绍IBM idataplex的GPU解决方案IBM Bladecenter的GPU解决方案,15,IBM BladeCenter GPU Expansion Blade-Mechanical Design,Top down view,NVIDIA Tesla M2070 GPU(or M2070Q),High speed I/O slot;connection point when stacking BGEs,Dual/redundant power&I/O connectors to the midplane,Feature CodeBGE+M2070 5090BGE+M2070Q A10R,16,IBM BladeCenter GPU Expansion Blade-Mechanical Design,Connected to IBM HS22 blade,IBM HS22 blade,IBM BladeCenter GPU Expansion Blade(BGE),17,BGE“Stacking”Example,IBM Blade Configuration,HS22,BGE#1,BGE#2,BGE#4,BGE#3,Ability to stack up to 4 BGEs per server,4 IBM blades+10 BGEs per BC-H Chassis(9U)3 or 1 GPUs/compute blade 40 GPUs in 42U rack,IBM GPU performance advantage in 42UGet up to 7%more GFlops of double precision performance in 42U!,Key,IBM CPU Blade,IBM GPU,
