2900 - IBM Fusion – 面向人工智能和容器的现代化基础设施.pdf

1、Orlando,FLOctober 69IBM TechXchange 20254479 AI Data EngineFrank LeeIBM StorageAI Data EngineAgenda010102More flexible deploymentsFueling AI FactoryFueling AI FactoryEnabling AI InferenceIBM TechXchange|2025 IBM CorporationCompute PodCompute Pod4 Scalable Units128 Compute Nodes1024 H100 GPUs82 TB of

2、 GPU Ram12,288 Physical Cores256 TB of RAM3481 TB NVME Local StorageScalable UnitScalable Unit32 Compute Nodes256 H100 GPUsIBM Blue VelaModular Deployment UnitsIBMs Purpose-Built Storage Solution for AI Factory IBM Storage Scale System 6000up to 1PB Usable Flash per node4310GB/s Read310GB/s Read155G

3、B/s Write155GB/s Write1+TB/s Read1+TB/s Read775GB/s Write775GB/s Write3+TB/s Read3+TB/s Read1.5TB/s Write1.5TB/s WritePer Rack!Per Rack!310 GB/s310 GB/s155 GB/s155 GB/sStorage Scale Storage Scale System 6000 System 6000 Data Engine for AI FactoryData Engine for AI FactoryAI FactoryAccelerationAccele

4、rationAbstractionAbstractionAutomationAutomationHPC ClusterHPCDGX/HGX Node with GPUtrainingGlobal Data PlatformAgenda010102More flexible deploymentsFueling AI FactoryEnabling AI InferenceEnabling AI InferenceIBM TechXchange|2025 IBM CorporationOver time,platforms Over time,platforms and methodsand m

5、ethodsF FindableindableA AccessibleccessibleI InteroperablenteroperableR ReusableeusableFAIRFAIR-able RDM(Research Data Management)able RDM(Research Data Management)Across tools,domains,fields,disciplines and institutionsFor large,distributed and unstructured dataBy Applications,workflows,AI modelsC

6、ATALOG1DISCOVER3INTEGRATE2ORCHESTRATE4metaEngine pilot MD AndersonCancer Center“finding needle in the haystack”202120172019metaOcean pilot Broad Institute“clean up data swamp”Launch of IBM Spectrum DiscoverIntroduction of Discover at U of Oslo for FAIR RDM2023Containerized Discov