Insights from Production Scheduled Ethernet Fabric in Large AI Training Clusters

"Pengfei Huo (Sr. Network Architect) - Bytedance
Xiguang - Henry Wu (Product Marketing) - Broadcom
The scale of xPU clusters is rapidly expanding from thousands to tens of thousands- and even aiming toward one million units- to meet increasing demands for AI computational power. Networking is vital- enabling scalability and optimal xPU utilization. Success hinges on network performance in congestion management- link failures recovery- load balancing- and noise isolation.
This presentation will examine how ByteDance uses a Scheduled Ethernet Fabric to tackle these challenges. We'll share performance metrics from ByteDance‚as inaugural Scheduled Fabric in production- supporting over a thousand xPUs. Further- we will discuss its operational aspects- compatibility- and differences from traditional non-Scheduled fabrics. We'll conclude with a proposal for further standardization and invite collaboration in the Scheduled Fabric ecosystem."