Bharath Ramesh, Praful Lalchandani, & Jon Green | Networking for AI Summit
summary
Enterprise on‑prem AI stacks scale from 16 nodes to hundreds of thousands of GPUs with tested reference architectures, a Juniper lab that validates Ethernet vs InfiniBand, and a first‑to‑ship 64‑port 800‑gig switch. The network uses ROCEv2, GPUDirect RDMA, NVLink where available, DCQCN and RDMA‑aware load balancing to avoid congestion; Apstra + Mist deliver intent‑based automation and AIOps that achieve time compression from infinite to minutes for troubleshooting and job‑aware visibility.
AI-written summary · about
