近GPU存储：加速存储以扩展AI工作负载的要求.pdf

1、Ta-Yu Wu,MetaEhsan K.Ardestani,MetaNear-GPU Storage:Requirements for Accelerating Storage to Scale AI WorkloadsNear-GPU Storage:Requirements for Accelerating Storage to Scale AI WorkloadsTa-Yu Wu,MetaEhsan K.Ardestani,MetaOCP SPECIAL FOCUS:ARTIFICIAL INTELLIGENCE(AI)Enabling Efficient AI Scaling wit

2、h StorageGPU PerfGrowthNetworkingLarge ClusterPowerBudgetsEnabling Efficient AI Scaling with StorageAI Deployments TodayGPU ServerGPUMemoryScaling AI Deployments for the FutureGPU ServerGPUMemoryStorageModelGrowthRapid SWChangeDataQualityAI Development Contributing FactorsHardware-Related Contributi

3、ng FactorsFocus Areas for GPU Initiated Storage Workloads01020304Performance EfficiencyMeasure performance efficiency in IOPS per dollar/watt in additional to TB per dollar/wattBalance in PerformanceDevices should have high sustained and peak performance for both reads and writesDevice FlexibilityFl

4、exible device architecture allowing different capacity and performance profiles to be deployed at customers site enables rapid alignment with latest workload iteration.Usable IOPSFocus on IOPS vs.Latency as a key performance metricNear GPU Storage Focus AreasNear-GPU Storage:A New Landing Zone for G

5、PU Initiated WorkloadsMemoryNear-GPU StorageCompute StorageWarm StorageCold Storage-Capacity per watt/dollar focus-IO Size are 4K-Topology/Technology:TLC Flash locally attached-New storage tier focused on performance per watt/dollar-IO Sizes are between 512B-4K-Capacity per watt/dollar focus-IO Size

6、 are 4K to MBs-Topology/Technology:TLC/QLC/HDD remote access-Capacity per watt/dollar focus-Large block access-Topology/Technology:HDD remote access-Performance per watt/dollar focus-Word size transfers(=64B)-Topology/Technology:DDR locally attachedIOPS per Watt|DollarCapacity per Watt|DollarAI Appl