e-Tube AI互连——跨越铜崖避免光链路损耗扩展AI集群.pdf

1、Dr.Sean ParkCo-Founder and CEOPoint2 Technology Inc.e-Tube AI Interconnect-Leaping the Copper Cliff And Avoid the Optical Penalties for AI Cluster Scale UpAI Scale Up Bottlenecked by Cable InterconnectToday:Nvidia NVL7272 GPUs2 mi.of copper cableTomorrowOne rack-576 GPUs 1632x data bandwidthCopper c

2、ant scale at the speed of AI computeMarket leaders and startups pursuing alternatives for 224G and beyondAI interconnect success factors:Energy efficiency,reliability,scalability,manufacturability,and costAI Cluster Scale-Up Interconnect Targets MetricTargetEnergy Efficiency 5 pJ/bCable Reach 10mCos

3、t$0.10/Gbps$0.05/GbpsBeach Front Density 1 Tb/mmLatencyNear zeroPresentation Starts HereThe Copper CliffToo thickToo shortNot scalableCopper and Optical Barriers Limiting Short-Reach,Terabit-level AI Compute Scale UpThe Optical PenaltiesToo ExpensiveToo power hungryReliability Issues LatencyIntroduc

4、ing e-Tube:New Category of InterconnectRF Transmitter and Receiver ICs for Data Transmission over Plastic Dielectric Waveguide(RF Over PlasticRF Over Plastic)Power3X LowerLatency1000X LowerCost3X LowerCable Length10X Coppervs.OpticalCablingFully Integrated Si solutionRF ICs built with matured and st

5、andard foundry process technologiesCable and connectorization built from copper manufacturing techniquese-Tube flat frequency response and fixed bandwidth across the spectrum Scale like Silicon:Scale in speed using the same e-Tube core construction for 224G/448G/beyondHow e-Tube WorksElectrical Doma

6、in(Copper)Electrical Domain(Copper)RF Domain(eRF Domain(e-Tube over Plastic Waveguide)Tube over Plastic Waveguide)Cable GenerationSERDES Data RateModulationDifferential Pairs(TX&RX)Dual Carrier FrequenciesChannel BandwidthModulation RF Data Ratee-Tube Cores(TX&RX)1.6T224GPAM416(8 Tx,8 Rx)90GHz/225GH