Optical CXL for disaggregated compute architectures.pdf

1、Ron Swartzentruber,LightelligenceOptical CXL for Disaggregated Compute Architectures Ron Swartzentruber,Director of Product Management,LightelligenceOptical CXL for Disaggregated Compute Architectures ARTIFICIAL INTELLIGENCE&OPTICSAIOPTICSMemory centric shift in the datacenterThe growth of AI and us

2、e of Large Language ModelsThe popularity of CXL and the need for optical connectivityOPT inference case study showing the benefits of optical CXL AgendaDisaggregation is the Future for the Datacenter Lightelligence,IncAI Trends Lightelligence,IncOptical Interconnect LatencyConsistent,100s of nsVaria

3、ble,many 100s of s Lightelligence,IncCXL:Emerging Standard for Disaggregation Lightelligence,IncOptical CXL is Superior for Scaling Lightelligence,IncOptical CXL in the Datacenter Lightelligence,IncCase Study:LLM Inference Lightelligence,IncLLM Model List Lightelligence,Inc Lightelligence,IncResults

4、Inference Output Lightelligence,IncDecode Throughput Improvement Lightelligence,IncSummary of Results Lightelligence,IncOptics used in Case Study Lightelligence,IncFor PhtowaveTM product information:https:/www.lightelligence.ai/OCP Marketplace Link:https:/www.opencompute.org/productsJoin the FTI Sho

5、rt Reach Optical Interconnect workstream:https:/www.opencompute.org/projects/short-reach-optical-interconnectAI HW-SW CoDesign workstream:https:/www.opencompute.org/projects/ai-hw-sw-codesignFor more on CXL(Compute Express Link)https:/computeexpresslink.org/For more information Lightelligence,IncEnd

6、notesHardware configurationSuper Micro ServerAMD EPYC 9124 16-Core CPUSamsung DDR5 4800 MT/sMEM0 size:256GBMEM1 size:256GBBandwidth:307GB/sNvidia GPUGen4x16,DMEM size:24GBBandwidth:32GB/sSamsung NVMEGen4x4,MEM size:1.92TBBandwidth:8GB/sMicr