NCP-AIN Tests Lossless AI Fabrics End to End — Not Generic Data Center Switching
A professional exam for engineers who run Spectrum-X Ethernet and InfiniBand fabrics for GPU clusters. Stop thinking in VLANs and ECMP; think in rails, subnet managers, congestion control and which tool sees which layer.
Check Your Readiness →Most candidates understand NVIDIA-Certified Professional: AI Networking (NCP-AIN) concepts — and still fail. This exam tests how you apply knowledge under pressure.
Spectrum-X and InfiniBand carry equal 30% weight, so you need both fabrics cold. Items show NVUE snippets, nv show or UFM output, or a symptom, and the distractors are real tools working at the wrong layer. NVIDIA says lab exercises are being added from November 1, so recheck the format.
Tune the ECMP hash to spread AI flows better
A few large synchronized flows collide under any hash. Enable per-packet adaptive routing on the switches with SuperNIC reordering on the hosts; both sides are required, and the slowest flow sets step time.
Read interface discard counters on each port
Counters only give totals. What Just Happened reports each drop with reason, port and flow, across L1, L2, router, tunnel, ACL and buffer groups. Only buffer drops indicate congestion; an ACL reason points to policy instead.
Conclude the remote host is unreachable on the fabric
ibping needs a responder: run ibping -S on the target first. Likewise ib_write_bw and ib_write_lat need a server side, and bandwidth tests never answer latency questions.
Put both tenants in one VRF on separate VLANs
A shared VRF shares a routing table, so tenants can route to each other. Give each tenant its own VRF and L3 VNI; type 5 routes carry tenant prefixes and type 2 carries MAC and host entries.
Deploy the RDMA shared device plugin through the GPU Operator
The Network Operator, configured by NicClusterPolicy, deploys the stack. SR-IOV virtual functions give per-pod isolation, Multus adds the secondary network, and IPoIBNetwork fits InfiniBand. The GPU Operator only adds GPUDirect RDMA peer memory.
PFC only pauses a priority; heavy pausing spreads congestion and causes head-of-line blocking. ECN marking plus SuperNIC congestion control handles sustained load, with PFC as the backstop. Many pause frames and few ECN marks means ECN thresholds need tuning.
LinkUp proves the physical layer works. Initialize means no subnet manager has assigned LIDs and routes. Start or fix the SM; for HA run standby SMs, since only one master per subnet is elected by highest priority, then lowest GUID.
The PKey high bit marks full membership. Full-to-full and full-to-limited traffic is allowed; limited-to-limited is blocked. The default partition is 0x7fff. In Kubernetes, ib-kubernetes applies PKeys to pods because namespaces do nothing on the fabric.
NVUE stages nv set and nv unset in a pending revision until nv config apply runs. Persistence to /etc/nvue.d/startup.yaml is automatic only with auto save on, the default from Cumulus Linux 5.9; otherwise nv config save is required.
In Spectrum-X the SuperNIC, typically one per GPU, owns east-west congestion control and packet reordering for adaptive routing. BlueField-3 in DPU mode sits on the north-south path for security, storage and infrastructure offload.
Strengthen weak areas with exam-style practice questions and detailed explanations.
Simulate the real exam experience and assess your readiness under timed conditions.
Review key concepts, objectives, and exam topics in one place.
Get personalized explanations, learning recommendations, and instant answers.
Follow a structured learning path designed to help you prepare efficiently.
NCP-AIN rewards picking the right fabric tool at the right layer. Check your Spectrum-X and InfiniBand readiness.