克服现实世界中数据与AI之间的挑战.pdf

1、1|2024 SNIA.All Rights Reserved.Overcome Real World Challenges between Data and AISteven Yuan CEO of StorageX.ai2|2024 SNIA.All Rights Reserved.AI2.0:AGI is driving next big cycle 3|2024 SNIA.All Rights Reserved.Data Volume20102020Scaling Issues:Compute+Network+Storage Bottleneck DataCPU&GPUThis Doe

2、snt Scale!Existing compute per data volumeIdeal Compute4|2024 SNIA.All Rights Reserved.Moving Compute-Return ResultsLower TrafficLower LatencyMoving Data Heavy TrafficLong LatencyThe Dumbbell effect causing high compute costData Centric Computing is a key solutionNetworkCPU/GPUNetworkStorage 5|2024

3、SNIA.All Rights Reserved.Learning Process of ChatGPT（1）Supervised learning（2）Reinforcement learning（3）Proximal Policy Optimization(PPO)DataDataDataData Movements+Compute6|2024 SNIA.All Rights Reserved.AutomationDigitizationArtificial Intelligence Automation tools to reduce labor cost use of data to

4、achieve the intended production goals,or operational efficiency.Use intelligence for self monitoring,automated problem identification,unmanned management etc.Key Factor Cost reduction Production Cycle Time Factory Capacity Key FactorUse digital tools for businessdecision making.Example:ERP,CIM,MES e

5、tc IoT devices Edge CloudKey FactorUse AI to achieve lighthouse factory:Higher Capacity Higher yield Faster problem solving Shorter cycle time More flexibilitySmart Factory:Intelligence needed for efficiency 7|2024 SNIA.All Rights Reserved.Network between Storage&ComputeCPUDRAMPCIe SwitchNetworkGPUG

6、PUGPUGPUCompute Node Big Data Set:Large amount of data generated from AOI machinery,Million files per day,could accumulate to PB level data per week.Large files:can be more than GB per file Faster cycle time:less than seconds for decision making Real World Challenges for Data and AI PCIe SwitchSSDSS