According to a Cisco blog post, Cisco benchmarked a scale-out AI fabric built from Cisco N9000 Series switches, AMD Instinct MI300X GPUs, and AMD Pensando Pollara 400 NICs. Cisco reports testing two Clos topologies (2×2 and 4×2) using Cisco N9364E-SG2 switches with 51.2 Tbps throughput and 64 800 GbE ports, 128 GPUs (16 servers x 8 GPUs), 128 Pollara 400 NICs, and Cisco 800G OSFP optics. The tests used the AMD ROCm software stack and Cisco Nexus Dashboard for visibility. Cisco frames the results as reducing job completion time and improving GPU utilization by addressing network-induced stalls. Editorial analysis: for practitioners, the report underscores that fabric bandwidth, congestion resilience, and telemetry are central variables when scaling multi-GPU training clusters.
