AI applications depends on large networks where hardware failure is likely to happen. By adding OCS between layers, it can reroute around failures connecting with redundant equipment, reducing the GPU downtime.
Today’s data centers are optimized for uniform all-to-all traffic, but real AI workloads show structured, changing traffic patterns. Replacement of the spine layer by an OCS layer enables on-demand topology changes dynamically adjusting to actual traffic demand, reduce multi-hop forwarding and optimizes for elephant flows.
