Benchmarking scale-out AI fabrics with Cisco N9000 and AMD Pensando Pollara 400 NICs
dailyHardware
Summary
Cisco and AMD published a validated scale-out AI fabric reference using Cisco N9364E-SG2 (N9000 family) switches, AMD Instinct MI300X GPUs and AMD Pensando Pollara 400G NICs. They benchmarked 2x2 and 4x2 Clos topologies (up to 128 GPUs across 16 servers) with IBPerf and MLPerf, finding P01 and P99 bandwidths close to 400 Gbps and a small P01–P99 spread. The design pairs 800G optics and Nexus Dashboard telemetry with validated switch/NIC settings to maintain deterministic throughput and reduce GPU stalls as clusters scale.
Why it matters
Provides concrete, validated evidence that fabric-level engineering with specific switch+NIC configs can deliver near-line-rate, predictable throughput and reduce costly GPU idle time in scale-out AI deployments.