IBM：释放AI伺服器与机器学习潜力.pdf

釋放AI伺服器與機器學習潛力李永輝IBM大中華區硬體系統部首席技術官暨傑出工程師勢不可擋的AI市場“By2020,80%ofBigDataandAnalyticsdeploymentswillneeddistributedmicroanalyticsand40%ofallbusinessanalyticssoftwarewillincorporateprescriptiveanalyticsbuiltoncognitivecomputingfunctionality.BothofthesetrendsrequireadramaticincreaseinprocessingpowerthatcouldbeenabledbyGPUs.”IDC3DataComputeAlgorithms組成AI的三個元件各行各業的數據變換INCREASING DATA VARIETYSearch MarketingBehavioral TargetingDynamic FunnelsUser Generated ContentMobile WebSMS/MMSSentimentHD VideoSpeech To TextProduct/Service LogsSocial NetworkBusiness Data FeedsUser Click StreamSensorsInfotainment SystemsWearable DevicesCyberSecurity LogsConnectedVehiclesMachine DataIoT DataDynamic PricingPayment RecordPurchase DetailPurchase RecordSupport ContactsSegmentationOffer DetailsWeb LogsOffer HistoryA/B TestingBUSINESS PROCESSPETABYTESTERABYTESGIGABYTESEXABYTESZETTABYTESStreaming VideoNatural Language ProcessingWEBDIGITALAI釋放Data的價值Language UnderstandingLanguage TranslationSpeech TranscriptionFace RecognitionMachine ReasoningObject Detectionknowledge科技案例進展AI訓練的數據處理Define training taskPrepare training DataData Pre-processingDNN Model selectionConfigure the training hyper-parameterDNN Model TrainingStartPackage the new DNN model together with preprocessing into inference proc.DL training framework preparationApplication development with inference APITo build a team with deep learning expertise:2 months 1 year To prepare massive training data:10 man month(s)To train a new model:1 hour weekTo give a new inference result:1 secondCognitive System:Provide optimized SW+HW design,and tool chains to significantly enhanceProductivityPerformanceTime to market具備URLI的智能系統LearningDeep Learning/Machine LearningImage/Video/Voice RecognitionReasoningHigh Performance DataAnalytics Knowledge Representationand ReasoningUnderstandNatural Language Processing(NLP)Unstructured InformationManagementInteractiveText To Speech(TTS)/Speech to Text(STT)Question Answering TechnologyDataComputeAlgorithms組成AI的三個元件邁向 AI 之路Z14 伺服器LinuxONE 伺服器Power Systems UNIX 伺服器向外擴展伺服器（面向 NoSQL/Hadoop）CPU+GPU 伺服器（面向 AI/HPC）核心基礎架構下一代 AI 工作負載企業 AI 工作負載大資料工作負載任務關鍵工作負載IBM Systems 的解決方案專為顛覆當今最高級的資料應用而設計，其中不僅包括您目前所運行的任務關鍵應用，還包括下一代 AI 工作負載。彈性存儲伺服器(ESS)全新 POWER9 處理器專為認知業務而設計I/O Throughput Increases 61%366GB/sEnergy Efficiency Improves 45%*&Workload Optimized FrequencyImproved Thread Performance with SMT8/4,Shorter PipelineAccelerator:PCIe Gen4,CAPI 2.0,OpenCAPI,NVLINK 2.0&on-Chip SMP Interconnect&Off-Chip Accelerator EnablementOn-Chip AccelDDR4 InterfaceAccelerator SignalingAccelerator SignalingDDR4 InterfacePCIe SignalingSMP SignalingCoreL2L3 RegionCoreCoreL2L3 RegionCoreCoreL2L3 RegionCoreCoreL2L3 RegionCorePCIeCoreL2L3 RegionCoreCoreL2L3 RegionCoreCoreL2L3 RegionCoreCoreL2L3 RegionCoreCoreL2L3 RegionCoreCoreL2L3 RegionCoreCoreL2L3 RegionCoreCoreL2L3 RegionCore14nm finFET 半導體處理器80 億個電晶體、17 層金屬堆疊120 MB eDRAM、12 x 20 路關聯區域芯片上頻寬達 7 TB/sPOWER9 的創新加速技術CAPI2.0EnhancementOpenCAPICollaboration 一致性加速處理器介面 計算加速（機器學習、視頻、生物資訊）存儲加速（記憶體中資料庫、存儲演算法）網絡加速（壓縮、加密）PCIe Gen4IndustryFirst 2 倍輸送量 後向相容 I/O 週邊設備連接 實際 I/O 標準NVLINK 2.0 CPU-GPU 連接 最大輸送量高達 300 GB/秒 CPU-GPU/GPU-GPU 統一記憶體訪問 DL、HPC、GPU DB 的理想之選IBM與客戶、合作夥伴的POWER9創新案例BART SANO谷歌平臺副總裁谷歌對 IBM 能夠在最新POWER 技術研發上取得進展感到非常高興。POWER9 OpenCAPI 匯流排及大記憶體功能為穀歌資料中心的創新帶來了新的機會。IAN BUCKNVIDIA 總經理、副總裁加速計算IBM 致力於為客戶交付高性能的解決方案，例如適於NVIDIA Tesla V100 GPU 和 NVLink 的 POWER9 解決方案，旨在説明客戶加速深度學習工作負載。OpenPOWER 2018 高峰會文章LimeLight 是一家致力於為客戶（如 BBC 和 Marvel Comics）提供各種各樣的高效工具，説明他們提升數位內容流處理水準的公司；該公司表示，OpenPOWER（以及基於 POWER9 的 PCIe Gen4）幫助他們克服了 PCIe Gen3 與其他伺服器一同使用時常常會出現的瓶頸，不僅加快了流處理速度，還縮短了緩衝時間。https:/ POWER SYSTEMSAC922The best server of enterprise AIThe IBM Power Systems AC922 offers the fastest way to deploy deep learning frameworks and accelerated workloads with enterprise class support.Up to 3.8X reduction in AI model training for deep learning frameworks1.8Xbetter performance of accelerated databases新一代的 NVLINK GPU 加速NVLinkDDR4NVL100GB/sNVL100GB/s100GB/s100GB/s170GB/sV100 GPUV100 GPUV100 GPUIBM POWER9 CPUIBM POWER9 和 NVLink 2.0 有助於解決您在代碼方面的 PCI-E 瓶頸；相比參與測試的 x86 平臺的CUDA 主機設備頻寬，可將資料傳輸速度提升 5.6 倍。POWER9 是市場上唯一一款面向 NVLink 2.0 而推出的處理器，從 CPU 到 GPU 均是如此。Source-https:/ IBM Power AC922 POWER9 CORAL 系統 Summit：橡樹嶺國家實驗室(ORNL)計算速度超過 150 Peta FLOPS，將會成為全球速度最快的超級電腦之一 相比 Titan，僅需四分之一的節點，便可實現 5-10 倍的應用性能提升 3,500 x IBM Power AC922(POWER9+NVIDIA V100 GPU+Mellanox InfiniBand)+IBM ESS 存儲+IBM Spectrum Computing HPC Software Stack17最新業企版 1.5PowerAI:Enterprise Software DistributionBinary Package of Major Deep Learning Frameworks with Enterprise SupportAI Vision*/DL Impact:Tools for Ease of Dev.Graphical tools to Enhance Data Scientist Developer ExperienceDL Impact/DDL/LMS:Faster Training TimesPerformance Optimized for Single Node&Distributed Computing Scaling醫療保健領域檢查股票指數預測電力公司採用無人機檢測線路智能監控系統客戶案例DataComputeAlgorithms組成AI的三個元件股票指數趨勢預測RNN(LSTM)&linear regression ModelFutureExchange Trading CodeFuture ExchangeDescriptionclass=2,(up,down)10 Min Forecast accuracy(%)class=3,幅幅度度0.1%(up 0.1%,flat,down 0.1%)10 Min Forecast accuracy(%)if888CSI 300 Index(滬深300股指連續)81.66%78.85%rb888Steel Bar(螺紋鋼連續)81.94%76.68%cf888Cotton(棉花連續)81.24%74.71%Model Adoption to Future Exchange&Foreign ExchangePredict next 10mins，Accuracy 80%+Predict next 50mins，Accuracy 70%+ComputeIBM Power AC922Cognitive SystemsIBM POWER9 CPUNVIDIA TESLA V100 GPUAlgorithmIBM PowerAI&IBM Deep Learning ImpactTensorFlowSignal Model:RNN(LSTM)Programming:PythonDataStock Index Historical DataHighLowOpenCloseVolumeK-LineMACDDIF股票指數趨勢預測深度學習 Deep LearningAdvanced Driver Assistance Systems(ADAS)SystemAnti-glare on traffic signBlock out passing cars Block out front carBad Weather Condition AI的演進AI for Business TrendBroad AI Disruptive&PervasiveLearning from less dataInterpretable&ExplainableEthics&BiasContinuously Learn&AdaptAutomatically-Constructed ArchitectureNext-Gen Systems(Hybrid,Novel Devices&Material)Dynamic DataInformation Represented by Knowledge未來的 AINarrow AI-Initial Value CreationMassive human-curated training data setsBlack Box AITrain&DeployStatic Algorithm,Specific ArchitectureDeep Learning AccelerationSingle-Task,Single-Domain IntelligenceStatic DataInformation Represented by Data今天的 AI謝謝讓 IBM 與您攜手共同釋放 AI 潛力迎戰未來！