Ultra Ethernet 1.0 的技术方面.pdf

1、J Metz,Ph.DTechnical Director,AMDChair,Ultra Ethernet ConsortiumUltra Ethernet 1.0-TechnicalsOCP SPECIAL FOCUS:ARTIFICIAL INTELLIGENCE(AI)As performantas a supercomputing interconnectAs ubiquitousand cost-effectiveas EthernetAs scalableas a cloud data centerTHE NEW ERANEEDS A NEW NETWORK 2025 Ultra

2、Ethernet Consortium.All Rights Reserved.Many working groups1 specification,many layersReleased 1.0 June,2025Ultra Ethernet Activities 2025 Ultra Ethernet Consortium.All Rights Reserved.Ever-increasing Memory Bandwidth needsEver-increasing Memory Capacity needs(Near)Instantaneous Data Access(Exabytes

3、+)The AI Monster4How to keep AI infrastructure fast,scalable and efficient as models grow exponentially?2025 Ultra Ethernet Consortium.All Rights Reserved.UEC Standing on the Shoulders of GiantsBackwards-compatibleUses libfabric as its north-bound APIDesigned to integrate into existing frameworks wh

4、ere libfabricis commonly utilizedKey driving force is in the Ultra Ethernet Transport(UET)Software APILibfabrics 2.0 with extensionsNew Transport LayerMulti-pathing Packet sprayingOrdered(ROD)and un-ordered(RUD)Lossy(no PFC)or LosslessCongestion Control:Enhanced Tx and new Rx TrimmingIn Network Coll

5、ectiveNetwork LayerIP v4/v6ECNData Link LayerNegotiation LLDPLink Level Retry-LLRHeader Efficiency ImprovementsPhysical LayerIEEE Compliant 100G SignalingAI and HPC Profiles 2025 Ultra Ethernet Consortium.All Rights Reserved.More important conceptsSES exchanges messagesPDS exchanges packets over Pac

6、ket Delivery Contexts(PDCs)PDCs are ephemeral connections and have multiple reliability modesROD,RUD,RUDI,UUDCMS allows packets spraying across the fabricThe fabric can be lossless or best-effortLink and PHY reliability features aim to eliminate retransmissionsMore System View an