光互连：赋能人工智能基础设施的未来.pdf

《光互连：赋能人工智能基础设施的未来.pdf》由会员分享，可在线阅读，更多相关《光互连：赋能人工智能基础设施的未来.pdf（17页珍藏版）》请在三个皮匠报告上搜索。

1、Roy RubensteinConsultant,LightCountingComputing Meets Optical Communications:A ConfluenceComputing Meets Optical Communications:A ConfluenceRoy Rubenstein,Consultant,LightCountingSPECIAL FOCUS01:Optics:The big picture02:Developments:Scale up and scale out03:Barriers:what matters-expanded 04:LightCou

2、ntings take05:Reflections06:Call to actionOutline“What we are seeing is a new renaissance for new architectures”Amin Vahdat,Google Fellow and VP of Engineering,Executive Forum,OFC 2025 Context:Computational demand is growing 10 x a year.for each of the last 8 years.Networking has become the bottlene

3、ck to achieve the scale required.Optics is playing a key role here:pluggables,co-packaged optics,and optical circuit switches.An inflection point:optics is getting integrated into the semiconductor world,driven by the dramatic growth in computing demands.01:Photonics renaissance moment?01:Optics:In

4、a state of fluxGPU&OpticsData center:Inside&campusmm,cm,m,10m,100m2km,10km20km80-120kmCoherent pluggables100s-1000s kmULH,submarine100ZR400ZR,800ZR,1600ZR3200ZR?Coherent embeddedCoherent-litePluggables800G,1.6TPath to 3.2TMulticore transceiversOCSesOptical I/OCPOCPCmicroLedsVCSELsTHz wireOCSesData c

5、enter:Metro to WAN Hollow-core fibermulticore fiber1.6TKeyOverlap push direction To meet the huge scaling challenges across all levels of compute(xPU,memory,and I/O)requires scaling up and scaling out the computing architectures:Scale-up:copper for nowpartitioning the workload across multiple xPUs w

6、ith the goal of achieving linear scaling(to appear as one large virtual xPU)Classical definition:getting an upgraded box to replace the existing oneScale-out:optics To achieve an architecture of 10,000s and 100,000s of xPUs(typically using InfiniBand or Ethernet,optical transceivers,co-packaged opti