利用 1.6T 网络大规模驱动 AI：开放式 100T 交换机将如何重新定义 2026 年及以后的数据中心（Celestica 出品）.pdf

1、2025 OCP Global Summit Executive Session|October 2025|CelesticaTareq Bustami,VP Market Technology,CelesticaHasan Siraj,Head of Software Products,BroadcomDriving AI at Scale with 1.6T Networking2025 OCP GLOBAL SUMMIT2025 OCP Global Summit Executive Session|October 2025|Celestica2Driving AI at Scale w

2、ith 1.6T Networking:How Open 100T Switches will Redefine Data Centers in 2026 and BeyondWELCOME TO THE CELESTICA AND BROADCOM PRESENTATIONHasan SirajHead of Software Products and EcosystemTareq BustamiVP Market Technology2025 OCP Global Summit Executive Session|October 2025|Celestica1 MillionA I A C

3、 C E L E R A T O R SNew AI Models 100X+Scale Distributed Computing2025 OCP Global Summit Executive Session|October 2025|CelesticaA Very,Very Large Distributed Computing SystemXPU ClusterXPU Network for AI workloadsCPUOptimized for Serial TaskGPUOptimized for Parallel TaskTENS OF THOUSANDS CORESMULTI

4、PLE CORESTHOUSANDS CORES2025 OCP Global Summit Executive Session|October 2025|CelesticaThe NetworkIs the Computer2025 OCP Global Summit Executive Session|October 2025|CelesticaAI Fabrics:The Nervous System of AI Infrastructure6SCALE-OUTSCALE-UPAcross Data CentersIn RackAcross RacksAcross RacksSpineL

5、eafSpineLeafData Center 02Data Center 012025 OCP Global Summit Executive Session|October 2025|CelesticaKey Challenges in Deploying Large Scale AI Clusters7Release Cadence2 YEARS1 YEARSerDes SpeedsCoolingOptics100G200GAIRLIQUID800G1.6T2025 OCP Global Summit Executive Session|October 2025|CelesticaOpe

6、n Standards2025 OCP Global Summit Executive Session|October 2025|CelesticaEthernet Scale-up:9High Performance,Open,Existing SpecificationsExecute at your own paceFreedom to innovate/implementPush vs Pull memory accessOrdering modelLoad BalancingSchedulingEthernet for Scale-up Networking(ESUN)Focus A