Henry Wu from Broadcom CSD discusses the Jericho3 AI Ethernet Fabric, a switch-based solution designed for AI workloads, highlighting its architecture consisting of GPUs connected via standard Ethernet to leaf and spine devices. The leaf devices, based on the Jericho family, handle switching, lookup, queuing, and scheduling, while the spine devices, based on the Ramon family, are power-efficient and optimized for AI workload interconnection without packet processing. The fabric is receiver-based, ensuring no congestion, and offers flexibility in placement, with consistent behavior regardless of whether leaf devices are in the same rack as GPUs or in a networking rack. Key innovations include perfect load balancing, a congestion-free mechanism, zero-impact failover, and scalability to 32K 800GB, with native support for multi-tenant cloud applications.
Sign in to continue reading, translating and more.
Continue