Cisco
2026-05-07
Cisco-AMD Benchmark Shifts AI Fabric Control from GPU to SmartNIC and Switch
Summary
Key Takeaways
Cisco and AMD have published a deep technical blog validating the deterministic performance of their AI fabric architecture. The core components include Cisco N9364E-SG2 switches (based on Silicon One G200, 51.2Tbps throughput, 64 ports of 800GbE), AMD Pensando Pollara 400 smartNICs (400Gbps), AMD Instinct MI300X GPUs, and the AMD ROCm software stack.
Tests used two Clos topologies (2×2 and 4×2), IBPerf for RDMA performance, and MLPerf for real workloads. Key KPIs focused on the delta between P01 (1st percentile) and P99 (99th percentile) bandwidth. In single-hop, bisectional, and incast (31:1 pattern) tests, both P01 and P99 bandwidth remained tightly clustered near 400Gbps line rate, proving stability under extreme congestion like all-to-all communication.
Cisco highlights Nexus Dashboard for Day-0 to Day-N operations. The solution is already deployed at G42's large-scale AI cluster. The results demonstrate that precise tuning of ECN and DCQCN ensures maximum GPU utilization and minimized job completion time (JCT).
Why It Matters
This blog is a strategic move by Cisco to build a containment circle around Nvidia's InfiniBand and Spectrum-X using Pensando smartNICs and Silicon One switches. The control point shifts from GPU compute to the network's congestion control algorithms and load-balancing logic. By deep-linking Nexus Dashboard with Pollara NICs, Cisco aims to lock users into a Cisco+AMD ecosystem, stripping architectural flexibility to adopt white-box switches or Nvidia NICs.
The blog downplays the tail latency optimization details of the Pensando Pollara programmable NIC. While average bandwidth is excellent under 31:1 incast, PFC and ECN threshold tuning is highly expert-dependent. Any topology or traffic pattern change could reintroduce Head-of-Line Blocking. Also omitted is performance degradation in cross-DC or WAN scenarios, and the centralized control plane (Nexus Dashboard) may become a latency bottleneck for monitoring data collection at scale (>1000 nodes).
PRO Decision
【Vendors (Arista, Nvidia, White-box camp)】Counter the Cisco-AMD joint solution by immediately publishing comparative benchmarks using SONiC or OpenFlow white-box switches with Nvidia BlueField-3/4 NICs, emphasizing adaptive PFC/ECN tuning under dynamic topology changes and mixed traffic (AI + traditional workloads) to prove the operational flexibility of open architectures.
【Enterprises (CIO/Architects)】Conduct a zero-trust technical audit: demand independent third-party test reports on Nexus Dashboard's control plane latency and data collection throughput at >1000 nodes. Also, evaluate whether firmware upgrades on Pensando Pollara NICs cause network outages, and request detailed cross-vendor interoperability test results (e.g., mixed deployment with Mellanox ConnectX-7 NICs) to avoid single-component lock-in.
【Investors】Recognize this partnership as Cisco's defensive containment of Nvidia in the AI networking market. Short-term bullish for Cisco, but long-term watch for Pensando's R&D amortization costs and market share erosion by Nvidia Spectrum-X. Monitor if AMD opens Pensando technology to other switch vendors (e.g., Juniper) to reduce Cisco's bargaining power.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)