CareerExamPro

NVIDIA AI Infrastructure (NCP-AII) & AI Power & Cooling (NCA) practice test

11 practice questions for this exam, each with a full explanation and cited rule or blueprint reference.

Start practising

NVIDIA NCP-AAI, MS AI-103, GenAI Leader: practice drill

10 realistic exam questions, graded as you go, with the exact rule or statute citation for every answer. Start free, no sign-up.

Timed mock exam

11 questions in 13 minutes, drawn from every topic. No answers shown until the end, then a full score breakdown, your weakest topics and a review of what you missed.

Exam Snapshot & Official Blueprint

Official Exam Fee
$125 (NCA) / $200 (NCP-AII) via NVIDIA
Format & Time Limit
60–70 MCQs · 90–120 minutes
Passing Standard
Scaled passing score
Typical Prep Incumbents
Brand-new 2026 GPU cluster & liquid cooling track — virtually zero Q-banks

Official Domain Breakdown

Exam DomainWeight
DGX/HGX Architecture, NVLink/NVSwitch & InfiniBand / Spectrum-X RDMA32%
Cluster Orchestration: Slurm, Kubernetes GPU Operator, MIG & DCGM28%
High-Density Liquid Cooling (Direct-to-Chip, CDUs, ASHRAE W1–W5)24%
AI Data Center Power Density, PUE Math & Thermal/Dew-Point Safety16%

High-Yield Formulas, Statutes & Rules Cheat Sheet

Frequently Asked Questions

Why did NVIDIA launch an AI Power & Cooling certification?

Modern 40–120+ kW liquid-cooled GPU racks exceed legacy air-cooling physics, creating an acute shortage of engineers who understand CDUs, ASHRAE water classes, and DGX cluster thermal limits.

Practice tests that prepare you for this exam

Practise by topic

About this exam

Agency / Issuer
NVIDIA
Region
Global / Online

Exact question counts, passing scores, and allowed reference materials vary by testing window and state/agency. Always confirm current rules with the official testing provider.

Sample questions

During large-scale distributed training over an Ethernet AI fabric, standard ECMP hashing causes severe tail-latency spikes and underutilized links due to elephant-flow hash collisions and receiver incast during NCCL all-reduce bursts. How does NVIDIA Spectrum-X with RoCEv2 mitigate these bottlenecks?

  1. By routing all NCCL collectives through the host Linux TCP/IP kernel stack so standard TCP Reno slow-start throttles GPU transmit queues smoothly.
  2. Statically pinning each GPU pair to a dedicated VLAN and disabling RDMA so switches buffer tensors in host RAM.
  3. By compressing model weights with lossless gzip inside the top-of-rack switch ASIC before forwarding packets across standard single-path links.
  4. Hardware adaptive routing (packet spraying), RoCEv2 ECN/PFC congestion control, and BlueField-3 out-of-order GPU placement.
Show the answer

D. Hardware adaptive routing (packet spraying), RoCEv2 ECN/PFC congestion control, and BlueField-3 out-of-order GPU placement.

Standard ECMP hashes long-lived, high-bandwidth RDMA elephant flows onto single paths, creating hash collisions and tail latency. NVIDIA Spectrum-X pairs Spectrum-4 switches with BlueField-3 SuperNICs to spray packets evenly across fabric paths (adaptive routing), reassemble out-of-order packets directly into GPU memory via GPUDirect RDMA, and apply ECN/telemetry-based congestion control plus PFC to keep RoCEv2 lossless and congestion-free.

Reference: NVIDIA Spectrum-X Ethernet Networking Platform Architecture / NCP-AII Blueprint

A 64-node DGX cluster managed by NVIDIA Base Command Manager (BCM) and Slurm experiences sudden slowdowns and job crashes during multi-day pre-training runs caused by a single node accumulating uncorrectable ECC memory errors and NVLink CRC replay spikes. What is the recommended operational workflow to prevent degraded GPUs from stalling jobs?

  1. Couple DCGM health diagnostics (dcgmi diag) with Slurm prolog/epilog checks and BCM alerts to auto-drain unhealthy nodes.
  2. Disable ECC memory scrubbing and ignore XID error interrupts in the NVIDIA kernel driver so training jobs continue running uninterrupted on degraded HBM banks.
  3. Schedule Slurm cron jobs to reboot all 64 compute nodes every 6 hours instead of tracking per-GPU health metrics.
  4. Replace BCM bare-metal provisioning with manual SSH scripts that check nvidia-smi output once per week during planned maintenance windows.
Show the answer

A. Couple DCGM health diagnostics (dcgmi diag) with Slurm prolog/epilog checks and BCM alerts to auto-drain unhealthy nodes.

In large synchronous training clusters, a single degraded GPU drags down or crashes the entire job. NVIDIA DCGM continuously tracks XID errors, uncorrectable ECC/row-remapping events, thermal throttling, and NVLink errors; coupling DCGM diagnostics with Slurm prolog/epilog health checks and BCM automated alerting drains unhealthy nodes immediately so healthy jobs continue without interruption.

Reference: NVIDIA Base Command Manager (BCM) & DCGM Administration Guide / NCP-AII Blueprint

Under ASHRAE TC 9.9 liquid cooling classifications (W1 through W5), why do modern warm-water direct-to-chip AI clusters target W4-class facility water supply (up to 45°C) rather than legacy W1/W2 chilled-water ranges (up to 17°C or 27°C)?

  1. W4 water pushes GPUs into a superconducting regime that doubles HBM3e bandwidth at constant rack power.
  2. W4 facility water eliminates the need for secondary CDUs, filtration, or corrosion inhibitors because warm water cannot harbor microbial growth.
  3. Warm supply temperatures enable chillerless heat rejection via dry coolers or towers across most climates, lowering PUE.
  4. ASHRAE W1 and W2 classes are restricted exclusively to two-phase fluorocarbon immersion tanks and cannot legally circulate through copper cold plates.
Show the answer

C. Warm supply temperatures enable chillerless heat rejection via dry coolers or towers across most climates, lowering PUE.

ASHRAE TC 9.9 defines liquid cooling classes by allowable facility supply water temperature: W1 (2–17°C), W2 (2–27°C), W3 (2–32°C), W4 (2–45°C), and W5 (>45°C). Operating in the W4 warm-water regime (e.g., 30–45°C supply) lets facilities reject heat directly to ambient air using dry coolers or evaporative towers for most or all of the year without energy-intensive mechanical chillers, dramatically lowering PUE and enabling district waste-heat reuse.

Reference: ASHRAE TC 9.9 Thermal Guidelines for Liquid-Cooled Data Processing Environments (W1–W5)

Taking this exam soon?

Tell us which exam domains or calculation drills you want more practice on.

Other 2026 NVIDIA, Agentic & Cloud AI exams