QuestionQ46

VerifiedArchitect the solution

A customer is deploying a cluster of twelve servers for AI workloads. Each server has four NVIDIA GPUs connected through NVLink. Each server supports 400 GbE connectivity on as many as six ports; however, you may recommend fewer ports. You have been asked to design a network for the servers’ backend AI interconnect.

Based on HPE Networking recommendations for this workload, how many CX 9300-32D switches should you propose?

Explanation

A rail-optimized AI backend network dedicates one leaf switch per GPU/NIC rail so that each GPU's traffic stays isolated on its own non-blocking fabric; with four NVLink-connected GPUs per server, four rail switches are required regardless of how many additional 400GbE ports the NICs could support, and each CX 9300-32D's 32 ports comfortably accommodate all twelve servers' single downlink per rail without needing a spine layer.

Community Discussion

No comments yet. Be the first to start the discussion!