Scaling AI: The Case For 100T Switching

Author: Matt Roman, Product Marketing Lead, Celestica Collection: 2025 OCP Global Summit - Expo Hall Sessions

Summary

Celestica, a Toronto-headquartered technology company with approximately $10 billion in annual revenue, 27,000+ employees across 40+ locations in 16 countries, presents the transition from 51.2T to 102.4T switching silicon and its new DS6000/DS6001 switch platforms. These platforms are based on the Broadcom Tomahawk 6 switch chip with 224 Gbps SerDes technology, delivering 64 ports of 1.6T Ethernet (equivalently 128x800GbE, 256x400GbE, or 512x200GbE). The DS6000 is a 3RU air-cooled 19-inch system, while the DS6001 is a 2OU hybrid liquid-cooled system in OCP ORv3 form factor with DC busbar power.

The 100T switching fabric enables collapsing three-tier networks to two-tier architectures, yielding a 67% reduction in latency and 67% fewer optics (since a three-tier design requires 67% more optics). At 200G per GPU connectivity, a single 100T switch supports 512 GPUs in a non-blocking single-stage fabric for scale-up. For scale-out, the two-tier architecture can support over 100,000 GPUs. The DS6001's hybrid liquid cooling uses cold plates on the Tomahawk 6 ASIC and front-panel optical interfaces (the two largest heat sources), while lower-power subsystems like the BMC CPU module remain air-cooled. Both platforms run SONiC open-source NOS or customer-proprietary network operating systems.