Enabling Scalable Infrastructure: Alibaba's High-Performance 102.4T Ethernet Switch

Author: Jing (Alibaba) Collection: 2025-OCP-Global-Summit-Networking

Summary

This talk presents Alibaba's HPN 8.0 (High Performance Network 8.0), the next-generation AI fabric architecture centered on a 102.4T Ethernet switch, and the detailed hardware design of two switch SKUs — "Simba" (128×800G, code name) and "Kimba" (64×1.6T, code name). The talk continues Alibaba's multi-year progression: HPN 7.0 (deployed with 51.2T switches starting September 2023, described at SigComm) → HPN 8.0 (102.4T switch as cornerstone). HPN 8.0 converges training and inference clusters in a single architecture with an inter-DC scale-out layer using ZR optics. The Universal Network Platform (UNP) is a key operational innovation: a chassis-level standardization with Removable Switch Modules (RSMs) that allows swapping switch SKUs and ASICs (Broadcom, Nvidia, Cisco) without changing the chassis or control module, enabling fast platform iteration across dozens of SKUs in three years.

The Simba switch is a 4RU phyless (no flex wire, pure PCB) design with 128 OSFP800 ports arranged in stacked configuration, SONiC inside, supporting OSFP800 DR8/VR8, LPO, and LRO optics. The Kimba switch uses the same 4RU chassis but with 64×OSFP1.6T ports. Liquid cooling for switches is addressed via a novel method: separating the cage and connector, then inserting a large cold plate between two stacked optic layers to cool groups of optics at once — contributed to OSFP MSA Rev 5.2. A TH5-based liquid-cooled switch has been running for ~10 months. LPO (Linear Pluggable Optics) deployment findings show significantly better link stability than DSP-based (FRO) optics: fewer link flaps across millions of optics in production, attributed to elimination of DSP tuning decision points, though LPO introduces manufacturing challenges and interop limitations (LPO-to-LPO stable, LPO-to-FRO sometimes problematic).