NCA-AIIO Exam Domains: Complete Breakdown of All 3 Domains (2026)

The NCA-AIIO blueprint is almost perfectly split between two big domains: AI Infrastructure at 40% and Essential AI Knowledge at 38%, with AI Operations filling the remaining 22%. Together the first two are 78% of the exam, so a candidate who is solid on hardware and AI fundamentals is already most of the way to a pass. This article walks all three domains topic by topic at the depth a 50-question associate exam expects.

New to the certification? Start with the complete NCA-AIIO guide. Ready to schedule the work? The 4-week study plan sequences everything below.

Domain 1: AI Infrastructure (40%)

The largest domain, and the reason this certification exists: can you build and run the physical infrastructure AI workloads need?

GPU hardware platforms (DGX, HGX). Know the NVIDIA system families and where each fits. DGX is NVIDIA's turnkey, fully integrated AI system (compute, networking, and software in one box); HGX is the baseboard building block that OEMs and cloud providers use to build their own servers. The exam expects you to match a scenario ("a turnkey system for an enterprise's first AI deployment" vs "a hyperscaler building custom servers at scale") to the right platform.

NVLink and InfiniBand networking. Two different fabrics at two different scales, and the exam tests the distinction directly:

• NVLink / NVSwitch connect GPUs inside a node at very high bandwidth (terabytes per second), so GPUs in one server share data almost as if they were one.
• InfiniBand (and Spectrum-X Ethernet) connect nodes to each other across the cluster at hundreds of gigabits per second.

The rule to remember: NVLink is intra-node, InfiniBand is inter-node. A question about GPU-to-GPU communication within a single DGX wants NVLink; one about scaling training across many nodes wants InfiniBand.

Power and cooling. AI hardware draws far more than traditional servers (a single high-end system can pull well over 10 kW, and a dense rack far more), which breaks assumptions built for 5-10 kW racks. The exam tests awareness that GPU density forces facility-level planning: rack power budgets, and the shift to direct liquid cooling when air cooling can no longer remove the heat.

On-premises vs cloud. The trade-offs: on-prem gives control and predictable cost at scale but requires capital and facilities; cloud (including DGX Cloud) gives elasticity and no facilities burden at a higher per-hour cost. Scenario questions weigh these against a stated requirement.

Storage and data center design. AI training is data-hungry, so high-throughput parallel storage feeds the GPUs; the exam expects awareness that storage and network design must keep GPUs fed rather than starved.

Reference architectures (BasePOD, SuperPOD). NVIDIA's validated cluster blueprints. DGX BasePOD is the smaller-scale reference design; DGX SuperPOD is the large-scale one for the biggest training jobs. Know that these exist as proven, pre-validated designs teams deploy rather than architecting from scratch.

Domain 2: Essential AI Knowledge (38%)

Nearly as large, and the conceptual foundation: what AI is, and why it needs this hardware.

AI vs ML vs deep learning. The nesting the exam expects you to state cleanly: AI is the broad field, machine learning is a subset that learns from data, and deep learning is a subset of ML using multi-layer neural networks. Deep learning is what drives the GPU demand, because training large neural networks is massively parallel.

GPU vs CPU architecture for AI. Why GPUs win at AI: a CPU has a few powerful cores optimized for sequential work; a GPU has thousands of simpler cores optimized for the parallel math (matrix multiplication) that neural networks are built from. The exam tests this reasoning, not just the conclusion.

The NVIDIA software stack. Know what each layer does:

• CUDA: the parallel-computing platform and programming model that lets software use the GPU
• cuDNN: GPU-accelerated primitives for deep neural networks
• TensorRT: an inference optimizer and runtime that speeds up trained models for deployment
• NCCL: the collective-communications library that coordinates multi-GPU and multi-node training (the all-reduce that syncs gradients)

A question naming "optimizing a trained model for fast inference" points to TensorRT; "coordinating gradient updates across many GPUs" points to NCCL.

Training vs inference. Two different workloads with different demands. Training builds the model: compute-heavy, memory-heavy, often multi-GPU and multi-node, run once (or periodically). Inference uses the model: latency-sensitive, runs constantly in production, and often needs less compute per request. The exam tests recognizing which workload a scenario describes, because the infrastructure choices differ.

AI use cases and recent trends. Broad awareness of where AI is applied (vision, language, recommendation, generative AI) and current directions, including the generative-AI and large-language-model wave driving today's infrastructure build-out.

Domain 3: AI Operations (22%)

The smallest domain, and the one most rewarded by hands-on experience: keeping the infrastructure running.

GPU monitoring with DCGM. DCGM (Data Center GPU Manager) is NVIDIA's tool for monitoring GPU health and utilization at fleet scale: temperature, power, memory, utilization, ECC errors, and throttling. The exam expects you to know DCGM is the standard for GPU observability and health checks, distinct from the single-GPU nvidia-smi snapshot.

Cluster orchestration and job scheduling. How work gets placed on GPUs: Kubernetes for containerized and inference workloads, and batch schedulers like Slurm for training jobs. The exam tests awareness that a shared GPU cluster needs a scheduler to allocate GPUs to jobs fairly.

GPU virtualization (MIG, vGPU). Two ways to share a GPU:

• MIG (Multi-Instance GPU) partitions a single GPU into up to seven hardware-isolated instances, each with dedicated memory and compute. Hard isolation, predictable performance.
• vGPU is software-mediated virtualization, typical for VDI and shared virtual desktops.

Know that MIG gives hardware isolation (the answer for multi-tenant inference needing guaranteed performance) while vGPU is the virtualization-platform approach.

Data center management practices. Operational routine: capacity planning, health monitoring, and keeping the cluster utilized rather than idle (idle GPUs are expensive waste).

Driver and firmware management. Keeping GPU drivers, firmware, and the software stack at compatible, qualified versions across the fleet, because version skew is a common source of subtle failures.

Turning the Breakdown into a Score

Two habits convert this map into points. First, respect the weighting: AI Infrastructure and Essential AI Knowledge are 78% of the exam, so they earn the bulk of your study time, but do not skip the 22% Operations domain, whose questions are easy points once you have seen DCGM and MIG in action. Second, verify each domain with weighted practice questions: Preporato's NCA-AIIO practice exams mirror this three-domain split across 7 full-length tests and 420 explained questions, with per-domain tracking that shows where you still owe points.

For the schedule, continue with the 4-week study plan, and keep the cheat sheet open for review passes.

Sources:

• NVIDIA NCA-AIIO Official Certification Page
• NVIDIA DGX Platform Documentation
• NVIDIA DCGM Documentation

Last updated: July 11, 2026