NVIDIA Certified Professional: AI Infrastructure Study Guide

How the Exam Is Structured

NVIDIA Certified Professional: AI Infrastructure (NCP-AII) validates system bring-up, hardware management, and control plane installation, gpu configuration, partitioning, and lifecycle management, cluster scheduling, containers, and ai workload runtime, network fabric, infiniband, and distributed communication performance. The ExamPal practice bank includes 281 premium questions and 40 free questions mapped across the official blueprint.

18% of exam

Domain 2: GPU Configuration, Partitioning, and Lifecycle Management

Covers GPU operational state, persistence, MIG-enabled environments, Fabric Manager, and compatibility across firmware, driver, CUDA, and management tools. It also includes interpreting GPU error and event conditions to support lifecycle management and troubleshooting.

22% of exam

Domain 4: Network Fabric, InfiniBand, and Distributed Communication Performance

Covers validation of InfiniBand and high-speed network configuration, diagnostic tools, NCCL communication topology and behavior, communication performance measurement, distributed training troubleshooting, and cluster-level communication readiness. It emphasizes fabric health, topology selection, and performance verification for AI/HPC communication paths.

Key Terms to Know

These terms are loaded from the shared terminology pack and appear across the question explanations.

--gpus flag

A container runtime option used to assign GPU resources to a container.

CUDA_ERROR_NO_DEVICE

A CUDA runtime error indicating that no visible or usable GPU device is available to the application.

Communication hang

A condition where distributed communication stalls and processes stop making progress, often due to transport or synchronization issues.

Container runtime GPU configuration

Settings that control whether and how GPUs are exposed to containers, including runtime flags and environment variables.

DCGM

NVIDIA Data Center GPU Manager, a toolset for discovering, monitoring, diagnosing, and managing GPUs in data center environments.

DCGM diagnostic Level 2

The minimum DCGM diagnostic level that includes memory stress testing in addition to basic health checks.

ECC page retirement

The permanent removal of faulty GPU memory pages from use after ECC detects repeated or severe errors.

Enroot

An unprivileged container runtime commonly used on HPC systems to run containerized workloads without Docker or root access.

GPU discovery

The process of detecting and enumerating GPU devices available on a system.

GRES

Generic RESources in Slurm, a mechanism for scheduling specialized resources such as GPUs.

Graphics engine exception

A fault reported by the GPU graphics or compute engine when executing invalid or problematic workload instructions.

H100 SXM

An NVIDIA Hopper-generation SXM-form-factor GPU accelerator designed for high-performance AI and HPC workloads.

HBM3

High Bandwidth Memory generation 3, a stacked memory technology providing very high throughput for GPUs.

InfiniBand

A high-performance networking technology commonly used in HPC and AI clusters for low-latency, high-throughput communication.

Memory bandwidth

The rate at which data can be read from or written to GPU memory, typically expressed in GB/s or TB/s.

Memory stress testing

A diagnostic procedure that exercises GPU memory heavily to detect stability or reliability issues.

NCCL

NVIDIA Collective Communications Library used for multi-GPU and multi-node communication primitives such as all-reduce and broadcast.

NCCL debug environment variables

Configuration variables such as NCCL_DEBUG and related settings used to troubleshoot communication failures and hangs.