Which network design approach should be used to build a high-performance, nonblocking AI/ML data center...

A rail-optimized topology is specifically engineered for large GPU clusters where collective communication performance and minimum inter-GPU latency are critical. In a rail-optimized design, corresponding GPU/NIC interfaces across compute nodes attach to the same rail switch. This allows communication between GPUs of the same rank to remain on a single leaf or rail switch whenever possible, eliminating unnecessary spine traversal and reducing network hops.

Cisco's AI/ML networking guidance gives a directly applicable design example: a rail-optimized, nonblocking architecture connecting 512 GPUs in 64 nodes, with eight GPUs per node . The design uses eight rail/leaf switches and four spine switches and maintains nonblocking bandwidth while optimizing GPU collective communication. Cisco explicitly states that rail-optimized design is recommended for NVIDIA GPU environments.

A conventional leaf-spine architecture can certainly provide high performance and nonblocking bandwidth, but option C is more specific to the requirement for the lowest possible GPU-to-GPU latency at this scale. Flat Layer 2 does not provide the required scalable fabric architecture, while a traditional three-tier model introduces additional hops and potential oversubscription.

Study Guide Reference: 2.0 AI Infrastructure Components and Architecture — 2.1 Evaluate network deployment based on bandwidth, latency, redundancy, scalability, and security.