NCP-AIN Exam Domains: Complete Breakdown of All 6 Domains (2026)

NVIDIA publishes the NCP-AIN domain weights, and they tell you where the exam actually lives: 60% on the two fabric technologies, 20% on troubleshooting them, and the last 20% split across automation, design, and Kubernetes. This article walks all six domains topic by topic, with the depth the exam expects for each.

If you are new to the certification, the complete NCP-AIN guide covers the exam basics first. This breakdown is the study companion: work through it domain by domain, and use the 6-week plan to schedule it.

Domain 1: NVIDIA Spectrum Networking (30%)

This domain covers Ethernet fabrics built on Spectrum switches, and its center of gravity is one question: how do you make Ethernet lossless enough for RDMA?

RoCE and the lossless machinery. RoCE v2 (RDMA over Converged Ethernet) lets NICs move data directly between server memory across a routed Ethernet network. RDMA transport assumes the network will not drop packets, so the exam expects you to know the three mechanisms that deliver that guarantee and how they interact:

• PFC (Priority Flow Control) pauses individual traffic classes hop by hop when a receive buffer fills, preventing drops at the cost of head-of-line blocking risk
• ECN (Explicit Congestion Notification) marks packets when queues cross a threshold, signaling congestion before buffers overflow
• DCQCN is the congestion control algorithm that consumes those ECN marks: the receiver reflects them to the sender as CNPs (congestion notification packets), and the sender reduces its rate

Know the failure modes: PFC without ECN leads to pause storms and congestion spreading; ECN thresholds set too high mean PFC fires first and throughput collapses. Exam scenarios frequently describe symptoms and ask which parameter is misconfigured.

QoS on Spectrum switches. Traffic classification, DSCP-to-switch-priority mapping, buffer allocation per traffic class, and keeping RoCE traffic in its own lossless class while control traffic stays lossy.

BGP-EVPN multi-tenancy. Spectrum fabrics use standard leaf-spine BGP underlays with EVPN overlays (VXLAN encapsulation) to isolate tenants. If you have EVPN experience from enterprise networking, this transfers directly; the AI angle is keeping overlay design from interfering with lossless behavior.

Spectrum-X platform. Spectrum-X pairs Spectrum-4 switches with BlueField-3 SuperNICs into a platform tuned for AI: per-flow adaptive routing on the switch, out-of-order packet handling on the SuperNIC, and end-to-end telemetry for congestion control. The exam expects you to know what the platform adds over standalone Spectrum switches and where DOCA (NVIDIA's SDK for programming BlueField devices) fits.

NetQ telemetry. NetQ validates fabric state continuously: BGP sessions, interface health, EVPN consistency, and RoCE counters across the fleet. Know what NetQ checks proactively versus what WJH captures reactively.

Domain 2: NVIDIA InfiniBand Networking (30%)

InfiniBand is a different world from Ethernet, and the exam treats it as such. If your background is Ethernet, expect the least skill transfer here.

Fabric fundamentals. InfiniBand is lossless by design through credit-based flow control: a sender transmits only when the receiver has advertised buffer credits, so congestion produces backpressure instead of drops. Addressing comes from the subnet manager, which discovers the topology, assigns LIDs (local identifiers) to every port, and computes forwarding tables for every switch. No subnet manager, no fabric.

UFM (Unified Fabric Manager). UFM is NVIDIA's management platform for InfiniBand: fabric discovery, health monitoring, congestion telemetry, and lifecycle operations, wrapping the subnet manager with an enterprise control plane. Exam questions cover UFM deployment models, high-availability pairs, and which UFM view answers which operational question.

PKey partitioning. Partition Keys isolate tenants on a shared fabric, conceptually similar to VLANs: each port holds full or limited membership in one or more partitions, enforced at the fabric level. Multi-tenant isolation scenarios on the exam resolve to PKey configuration.

Adaptive routing and SHARP. Adaptive routing lets switches spread traffic across equal-cost paths based on live congestion instead of static hashing. SHARP (Scalable Hierarchical Aggregation and Reduction Protocol) moves collective operations into the switches themselves: the fabric computes the reduction as data flows through, cutting all-reduce traffic roughly in half. Know what each feature does, when to enable it, and what NCCL gains from it.

QoS and high availability. InfiniBand QoS runs on service levels (SLs) mapped to virtual lanes (VLs), giving traffic classes separate buffering on every link. HA questions cover redundant subnet managers, UFM failover, and dual-rail host connectivity.

Domain 3: Troubleshooting Tools (20%)

The troubleshooting domain tests diagnostic sequences: given a symptom, which tool do you reach for, and what does its output tell you to do next?

WJH (What Just Happened). Spectrum switches record every dropped packet with the reason: buffer overflow, ACL drop, forwarding miss, MTU violation. WJH turns "something is dropping traffic" into "this port dropped these flows for this reason." It is the reactive complement to NetQ's proactive validation, and the exam loves scenarios where WJH output points to the root cause.

InfiniBand diagnostics. The command-line family the exam references most:

• ibstat / ibstatus for local HCA state, link width, and speed
• ibnetdiscover to walk and map the fabric topology
• ibdiagnet for a full-fabric health sweep: bad links, symbol errors, routing issues, counter anomalies
• perftest tools (ib_write_bw, ib_send_lat) for point-to-point bandwidth and latency measurement

Know the workflow, from confirming a healthy local port outward to a fabric-wide sweep, and which counters (symbol errors, link downs, port receive errors) indicate cable problems versus congestion.

UFM diagnostics. Fabric-wide health from the management plane: congestion maps, error counters over time, and event correlation across links.

Cable and link diagnostics. Transceiver and DAC/AOC issues produce signature patterns: symbol errors climbing on one port, links negotiating below expected speed, or flapping under load. The exam expects you to distinguish a physical-layer problem from a congestion problem from the counters alone.

Domain 4: Automation & Configuration (10%)

NVUE. NVUE is Cumulus Linux's declarative configuration system: nv set stages changes, nv config apply activates them, nv config save persists them, and revisions can be diffed and rolled back. The exam tests the workflow, including how to preview and validate a staged change before applying it to production switches.

Ansible at fleet scale. Rolling one QoS change to 64 leaves by hand invites drift, so the exam covers Ansible-driven automation against NVUE's API, idempotent playbooks, and configuration templating across a fabric.

ZTP (Zero Touch Provisioning). New switches boot, fetch a script from DHCP-provided options, install their OS image, and pull their config with no console cable involved. Know the ZTP sequence and where it fits in fabric bring-up.

Domain 5: AI Data Center Design & Optimization (5%)

Small weight, but the concepts here also inform scenario stems in the two big domains.

Rail-optimized topology. In a rail-optimized design, GPU N of every server connects to the same leaf switch (rail N). Collective operations between matching GPU ranks then traverse exactly one switch, which is the traffic pattern NCCL generates during all-reduce. This is the reference design for AI pods and the topology the exam assumes unless stated otherwise.

Fat-tree and oversubscription. Classic non-blocking CLOS design: total uplink bandwidth per leaf equals total downlink bandwidth, giving 1:1 oversubscription for the compute fabric. Storage and management networks tolerate oversubscription; the GPU fabric does not.

Bandwidth planning. Per-GPU network bandwidth (one 400Gb/s NIC per GPU in current reference architectures), separate fabrics for compute, storage, and management, and the math for how many leaf and spine switches a given GPU count requires.

Domain 6: Kubernetes Integration (5%)

NVIDIA Network Operator. The Network Operator does for the network stack what the GPU Operator does for compute: it deploys and manages the RDMA device plugin, SR-IOV components, secondary network CNIs, and NIC drivers as Kubernetes-native components. The two operators are designed to run together for GPUDirect RDMA workloads.

Multus and SR-IOV. Kubernetes gives every pod exactly one network interface by default. Multus is a meta-CNI that attaches additional interfaces, and SR-IOV (Single Root I/O Virtualization) carves a physical NIC into virtual functions that pods use directly at near-line rate. The pattern the exam tests: primary CNI for pod traffic, plus a Multus-attached SR-IOV virtual function carrying RDMA for training.

RDMA in Kubernetes. How workloads request RDMA resources through the device plugin, and why GPUDirect RDMA (NIC-to-GPU-memory transfers that bypass the host CPU) depends on both operators being deployed.

Turning the Breakdown into a Score

Two habits convert this domain map into exam points. First, study in domain-weight order and hold yourself to per-domain targets rather than an overall feeling of readiness. Second, verify each domain with weighted practice questions: Preporato's NCP-AIN practice exams mirror this exact six-domain split across 7 full-length tests and 420 explained questions, and the per-domain score tracking shows which of these sections still owes you points.

For the schedule that puts it all together, continue with the 6-week study plan, and keep the cheat sheet open for review passes.

Sources:

• NVIDIA NCP-AIN Official Certification Page
• NVIDIA Networking Documentation
• NVIDIA Spectrum-X Platform
• NVIDIA DOCA Framework

Last updated: July 9, 2026