线性驱动可插拔光器件 (LPO) 和线性接收光器件 (LRO) 的发展挑战和下一代机遇.pdf

1、Nan Ge,Amin YekaniMetaChallenges and Opportunities for Linear Drive Pluggable Optics(LPO)and Linear Receive Optics(LRO)AI Scaling Driving Optics GrowthSource:Saif Hasan,“Scaling Llama4 Training to 100K”,Scale Networking 2025Scale:Rapid growth in AI training scale substantially increases backend netw

2、ork demand,with estimated accelerator-to-optics ratios of 1:6 to 1:10.Tech Evolution:Adopting LPO/LRO can reduce I/O power by 30-50%and improve cost efficiency by 2030%,while 200G/L optics enable 2x density to meet accelerators growing bandwidth requirements.Path Forward:Pluggable optics innovations

3、 serve as practical intermediate steps before co-packaged optics can be broadly deployed.Scaling LlamasTraining Larger Model with EfficiencyLLAMA270B Parameters2T Tokens2K GPUsLLAMA3405B Parameters15.6T Tokens16K GPUsLLAMA4288B/2T Parameters30T Tokens32K GPUs*100K GPUs8x8xRetimed optics costs contin

4、ue to improvesupported by a healthy ecosystem and BOM optimization,while power efficiency is saturating as process node advancements reach diminishing returns.100G/L LPO has been positioned as a post-TTM optimization over the past two years,with significant learning gained to overcome implementation

5、 challenges.LRO shows promising potential for broader adoption in the 200G/L generation,with early results indicating scalability for large-scale data center deployments and expansion of the linear drive footprint.100G/L to 200G/L:Retimed vs Linear Power EvolutionEcosystem:From POC to DeploymentPOC

6、ValidationUse Case EvaluationEnabling TechnologiesMarket and StandardsDefine architecture requirementsEcosystem Engagement(100G/L,200G/L,Retimed,LPO,LRO)Testing and ValidationLPO MSA driving alignment,200G/L IEEE specsAdoption shift:Wait-and-see to pilot deploymentsDSF(disaggregated scheduled fabric