Inspur GPU Server - 株式会社キング・テック

1 InspurGPU Server Inspur AI Computing Platform3 GPU Server4 GPU Server8 GPU Server16 GPU ServerNF5280M4 (2 CPU + 3 GPU)NF5280M5(2 CPU + 4 GPU)GPU Node (2U 4 GPU Only)NF5288M5 (2 CPU + 8 GPU)SR GPUBOX (16 P40 GPU Only) CSP Deep Learning Real-time transcoding VOD Server HPC Heterogeneous Computing HPC Cluster CSP Multi GPU, GPU P2P, RDMA supports High capacity local storage Highperformance-priceratio HPC Multi GPU/MIC, GPU P2P, RDMA supports 100G Ethernet, InfiniBand Server target marketPurposeH/W Requirements NVLink is a high-speed interconnect that replaces PCI Express to provide up to 12X faster data sharing between support on UnifiedMemoryPool3 NVIDIA P100 (Pascal)NVLink High-speed interconnectNVIDIA GPUD irect Peer-to-Peer (P2P) CommunicationGPUD irect Advantages Accelerated communication with network and storage devices Peer-to-Peer Transfers between GPUs Peer-to-Peer memory access RDMANF5288M4NF5288M5 CPU1. 2xIntel Xeon Processors E5-2600V4series2. DDR4 DIMM per nodesupport DDR4-2400 memory 16 DDR4 DIMM per node4 Apache Pass supportsPCIE I/O8 x x16 (x8 link) or 4 x x16 (x16 link) and 1 x x241 x x8 mezzanine RAID2 x x16 HHHL Front slotStorageSupport 8 x SAS/SATA/SSDS torage Controller.

2 8 x (8639)2 x PCIe & SATA on boardGPU SupportSupporting up to 4 x GPU and MIC Accelerator CardsUp to 8 x 300W GPU/SXM2 System FanRedundant Hot swap System Fan, Air CoolingRedundant Hot swap System Fan, Air cooling or Hybrid 1620/2000W PSU 80plus Platinum2x 3000 WPSU 80plus TitaniumNF5288M4 VSNF5288M5 2U GPU Server for HPC and Machine Learning 2 x SKL EP Processors, TDPupto 165W, SupportSKL-FSKUs Support8 XeonPhi/GPU ina2 Uchassis BothPCIeAIC& SXM2 GPU aresupported 8 storagebays GPUTDP upto300W 3000W 1+1 PSU TitaniumSXM2 ConfigurationPCIeAIC GPU Server for PurleyFront viewRear view8 PCIe 16 HHHL slot2 3000w PSU4 10G Ethernet2 C20 Power connector4 PCIe 16 HHHL slot(only for SXM2 configuration)Rear I/OFront I/ONF5288M5 GPU Server for Purley8 SXM2 NVIDIA GPU4 PCIe 16 HHHL slotLiquid cooling connector(Optional)2 Skylake, 165W TDP5 Redundant dual rotor fan16 DDR4 2400 DIMM2 Front PCIe 16 expansionSXM2 NVIDIA GPU configurationNF5288M5 SXM2 Liquid SolutionWateroutLQ2 WaterinLQ4 CPU0 CPU1 RAIDUPIF ront PCIe x16 PCIe switch96-laneRear PCIe x168x PCIe x16 Front PCIe x16 PCIe switch96-lanePCIe switch96-laneGPU0 GPU2 GPU1 GPU3 GPU4 GPU6 GPU5 GPU7 PCIeNVLINKRear PCIe x16 Rear PCIe x168 SXM2 GPU Topologic on NF5288M5 PCIe AIC GPU configuration8 PCIe dual slot PCIe 16 For XEON Phi/GPU2 Skylake, 165W TDP5 Redundant dual rotor fan16 DDR4 2400 DIMMC oprocessor maintain handle2 Front PCIe 16 Card per group designFlexible topologic in 8 PCIe configurationCPU0 CPU1 RAIDPCIe 8 UPIHHHLPCIe 16 CPU1 RAIDPCIe 8 UPISWSWCPU0 CPU1 RAIDPCIe 8 UPISWSWS limline PortPCIe 16 HHHLPCIe 16 HHHL PCIe 16 HHHLPCIe 16 Slimline PortPCIe 16 Slimline PortPCIe 16 Slimline PortPCIe 16 Slimline PortPCIe 16 Slimline PortPCIe 16 SWSWCPU0 Proposal C More expansibility.

3 High CPU to GPU bandwidth RAID mezzanine 2 HHHL PCIe 16 in front 8 B High ratio on Xeon Phi/GPU vs CPU RAID mezzanine 2 HHHL PCIe 16 in front or 1xHHHL PCIex16 + or A All GPUs in same domain RAID mezzanine 2 HHHL PCIe 16 in front 8 Supports Different GPU CardsSupports 8 NVIDIA Tesla P100 GPU Cards Built-in NVIDIA NVLink Supports 8 NVIDIA Tesla P100 P40 P4 GPU Cards PCIeinterface GX4 GPU BOX4*GPUs2*1600 WPower SupplyNVMe SSD ExpansionEfficientThermalFanPCI-e*16 ExpansionPCI-e Switch ChipPCI-eExpansionMotherboardGX4 GPUresourcedecouplingandpooling8*GPUs 16*GPUsScale outScale upPartition design of CPU Server and GPU Box Flexibletopology & highscalabilityEfficientdatacommunicatio ns & highTCOrevenueGX4 Flexible GPUs TopologyPCIe switchGPU 0 GPU 2 GPU 1 GPU 3 CPU ServerCPU0 CPU1 UPIPCIe switchGPU 0 GPU 2 GPU 1 GPU 3 PCIe switchGPU 0 GPU 2 GPU 1 GPU 3 CPU ServerCPU0 CPU1 UPIPCIe switchGPU 0 GPU 2 GPU 1 GPU 3 PCIe switchGPU 0 GPU 2 GPU 1 GPU 3 CPU ServerCPU0 CPU1 UPIPCIe switchGPU 0 GPU 2 GPU 1 GPU 3 BalancedPubliccloudserviceSmall-scalemod eltrainingCommonDeepLearningmodeltrainin gCascadedDeepLearningmodeltrainingP2 Pfunction enhanced TCOR evenue TraditionGPUGPUGPUGPUCPUCPUGPUGPUGPUGPUC PUCPUGPUGPUGPUGPUCPUCPUGPUGPUGPUGPUCPUCP UIB SwitchLargescaleI/O Redundancy 4setsofCPU + Memory +Storage 4* IBcards 1*IBswitch 16*GPUsTraditionGPUC lusterFrameworkPurchaseCostHighTCOB enefit GPU BOX16-CARD IN ONE SYSTEMGPU communication needsnonetworkprotocolconversionreduce50 %+ I/O redundancy Comparedwithtraditionframework,Purchasec ostreduceby$15,000+.

4 1 setsofCPU + Memory+Storage 0* IBcards 0*IBswitch 16*GPUsPurchaseCostLower I/O RedundancyPCIe switchGPU 0 GPU 2 GPU 1 GPU 3 CPU ServerCPU0 CPU1 UPIPCIe switchGPU 0 GPU 2 GPU 1 GPU 3 PCIe switchGPU 0 GPU 2 GPU 1 GPU 3 PCIe switchGPU 0 GPU 2 GPU 1 GPU 3GX4 SupportwithfullrangeofPCIE acceleratorsNVIDIA Tesla P100 NVIDIA Tesla P40 NVIDIA Tesla P4 Fast Data SwappingIn GPU MemoryLarge Amountof Training DataHigher Efficiencyfor DL InferenceSupportvariousGPU,FPGA,KNLandot herPCIE cards,andreserveNVMe poolingfunctionHigherprice/performancera tiofor HPCI ntel KNLB etterTCO for InferenceFPGAGX4 GPUS pecificationsGPU BOX Specifications ModelNumberSF0204P1 GPU4*PCIeP100/P40/P4/ (WithoutGPU)PCIeSupport1standardPCI-e*16 slot 4mini PCI-e*4 cableI/ORJ45managementport,serialportPow erSupply1600w 1+1redundantpowersupplyOutletRear-endout letHeadnode SpecificationsModel NumberNF5280M5 CPUS upport2*IntelNext Generation Processer Platform -SkylakeMEM1. 24 x DDR4 DIMMand 12 x Apache Pass; 2. Support RDIMM, LRDIMM, NVdimm3.

5 Support 2400, 2666 * 12 + *4 , * 24 + *4+ *4,including6frontNVMePCI-eSupportup to 4 GPU BOX