大规模部署QLC：平衡密度、功耗、性能和可靠性.pdf

1、Ben Di QualSenior Director Product ManagementPure StorageDeploying QLC at Scale Balancing Density,Power,Performance,and Reliability32025 Pure StorageAgendaStorage ChallengesCapacityPerformanceData Center Resources42025 Pure StorageAgendaStorage ChallengesCapacityPerformanceData Center ResourcesWhy Q

2、LC?DensityPerformancePower per PB52025 Pure StorageAgendaStorage ChallengesCapacityPerformanceData Center ResourcesWhy QLC?DensityPerformancePower per PBOvercoming Adoption ChallengesStandard Form FactorEnduranceError RatesMemory Utilization62025 Pure StorageStorage Continues to GrowTraditional Work

3、loadsMostly HDDLow acquisition cost per TBMajority of the data is at rest72025 Pure StorageStorage Continues to GrowIDC assumes 7.9ZB in 2021 at 19%CAGRTraditional WorkloadsMostly HDDLow acquisition cost per TBMajority of the data is at restEmerging WorkloadsMore of the data is activeIncreasing dema

4、nds on performance82025 Pure StoragePerformance Demands are AcceleratingEstimate based on industry papers and analysis with tokens converted to EB values.Assumes approaches to training remain similarDemands on PerformanceAmount of hot data is increasingPerformance per TB is accelerating92025 Pure St

5、oragePerformance Demands are AcceleratingGPT3Compute Data Scaling&ImagesFrontier LLMs&Synthetic DataVideoMainstream Multi-modalDemands on PerformanceAmount of hot data is increasingPerformance per TB is acceleratingEstimate based on industry papers and analysis with tokens converted to EB values.Ass

6、umes approaches to training remain similar102025 Pure StoragePerformance Demands are AcceleratingLBMs and Specialized Modalities3D,robotics,Long contextWider spread Multi-ModalDemands on PerformanceAmount of hot data is increasingPerformance per TB is acceleratingEstimate based on industry papers an