激活数据以驱动人工智能：无需重建基础设施即可统一数据.pdf

1、1 2026-H 2026-HActivating Data for AI:Unifying Data Without Rebuilding InfrastructureFloyd ChristoffersonVP Product Marketing-HammerspaceHPC User ForumAustin,May 5-6 2 2026-HData underpins every AI solution Todays Enterprise Needs but it doesnt live in one placeUnified AccessAutomated PipelinesVisib

2、ility into Data73%of Enterprise data is siloed3 2026-HAI Exposes Limits of Centralized DataCloudData CenterEdgeObjectFileAI Changes How Data Must be UsedWhat Breaks At ScaleData stays where it isCompute moves to where its availableWorkloads are no longer fixedCopy sprawlIdle GPUsPower&cost wasteDepl

3、oyment delaysGPUsAI FactoriesInference ClustersAI is not just compute-constrained.It is data-constrained.4 2026-HThe Reality Has ChangedFrom HPC Built for Mod/Sim to AI+Hybrid WorkflowsAI+HYBRIDWORKLOADSTrainingInferenceAnalysisWhats Different Now Data is everywhere,not just in one place Identities

4、dont always match Workloads are dynamic and diverse GPUs need data fast and local Moving or copying creates friction,cost,and riskThe Model That Worked Centralized storage Consistent identity Stable access patterns High-throughput shared file systems Built for long-running jobsHPC ClusterSINGLE SITE

5、Single DomainSINGLE UID SPACEConsistent IdentitySimulationBatch OrientedPREDICTABLEWORKLOADSPARALLELFILE SYSTEMHigh-throughput Shared StorageHPC Was Built for Mod/SimCompute-Intensive.Predicable.Centered.The Emerging RealityData-Intensive.Dynamic.Distributed.DistributedGeographyMULTI-SITEMULTI-DOMAI

6、NAdministrativeBoundariesMULTI-UIDIdentityFragmentationDYNAMIC WORKFLOWSAI,Simulation,Inference,AnalysisCLOUDObject StoresDATA CENTERSMulti-SiteARCHIVEEDGE&REMOTELOCATIONSGPU NODE-LOCALNVMe/SSDPARALLEL FILE SYSTEMSThis is not just a Performance Problem.It Is a Data Orchestration Problem 5 2026-HData