探索 Azure 的高级 AI 基础架构实现安全可持续增长.pdf

1、Exploring Azures Advanced AI Infrastructure2Kelvin HuangHW Dev DirectorAzure Hardware System and InfrastructureMicrosoftExploring Azures Advanced AI infrastructureAI CLUSTERS3Outline4321Microsoft Azure Infrastructure overviewTechnology Pathfinding OpportunitiesDisaggregated Infrastructure Changes Fu

2、eling the Next Wave of AI PlatformsCall to Action4Source:Morgan Stanley Growth,Tech Industry Is Building Out Compute At Scale to Support AI5Azure:Supercomputing Breakthroughs400 GB/s NDR InfiniBand proven HPC interconnect1st for cloud20,000 cores MPI job1st for cloud200 Gb/s HDR InfiniBand with adap

3、tive routing1st for cloudMay 2010,000 V100 GPUTop5-scale supercomputer for OpenAI(CPU+GPU)1st for cloud1 TB/sec Parallel File System1st for cloud1.6 TB/s InfiniBand NVIDIA A100 clusters1st for cloudHBv3 Milan GA in Azure at AMD Launch1st for cloud or on-prem10-year Supercomputing-as-a-Service for Me

4、t Office1st for cloud#10 supercomputer+4 more in Top40 1st for cloud#1 cloud on MLPerf benchmark#2 overall1st for cloudHBv4 Genoa-x in Azure at AMD Launch1st for cloud or on-prem2019AMD EPYC Rome InfiniBand HPC clusters1st for cloud80,000 cores MPI job1st for cloud 12X more than any other cloud 2020

5、202120222023#1 cloud on MLPerf benchmark#2 overall Again!1st for cloudNov 2314,400 H100 GPU#3 supercomputer1st for cloudMay 2430 x Supercomputers6Microsoft is continuing to make substantial investments in datacenterinfrastructure to support the growing demand for AI and cloud services.This year,were

6、 investing$80 billion toward the development of AI-enabled datacenters around the world.70+Azure regions400+Datacenters worldwide600k+KM of network190+Network PoPs7Sustainability34GW140KW900MWMicrosoft 2030 GoalsZero WasteWater PositiveCarbon NegativeMeasuring energy and water efficiency for Microso