nvidia-nca-aiio
Use when preparing for NVIDIA-Certified Associate AI Infrastructure and Operations (NCA-AIIO) — AI and accelerated-computing fundamentals, GPU servers and interconnects, data-centre networking and sto
SKILL.mdskills/certs/nvidia-nca-aiio/SKILL.md
GitHubNVIDIA-Certified Associate: AI Infrastructure and Operations (NCA-AIIO)
Bottom line: it is an infrastructure exam with an AI vocabulary. Know why AI workloads need GPUs, how GPUs are connected and fed with data, and how clusters are scheduled and monitored — at the level of "which component solves this bottleneck".
Exam shape (confirm on NVIDIA's certification page)
- Online, remotely proctored, 60 minutes, in English; multiple-choice questions; result reported as pass/fail; the credential is valid for a limited period and renewed by re-taking.
- Areas: essential AI knowledge (AI/ML/DL concepts, training vs inference, the GPU software stack), AI infrastructure (compute, networking, storage, data-centre considerations) and AI operations (management, monitoring, orchestration). Check the exam page for the current weights and objectives.
Primary sources: NVIDIA certification page (https://www.nvidia.com/en-us/learn/certification/), the exam study guide, NVIDIA documentation for the products named in the objectives.
Concept map
| Topic | Know |
|---|---|
| Why GPUs | Massive parallelism for matrix operations; CPU vs GPU roles; tensor cores and reduced precision (FP16/BF16/FP8, INT8) for training vs inference |
| Training vs inference | Training: large batch, high memory and interconnect bandwidth; inference: latency, throughput, model optimisation and serving |
| Software stack | CUDA, CUDA-X libraries, containers from NGC, frameworks, inference servers (e.g. Triton), NVIDIA AI Enterprise |
| Scale-up interconnect | NVLink / NVSwitch inside a server or rack |
| Scale-out networking | InfiniBand and Ethernet (RoCE) for multi-node training; RDMA; DPUs offloading networking, storage and security |
| Storage | High-throughput parallel or network file systems; GPUDirect Storage avoiding CPU copies |
| Data centre | Power density, liquid vs air cooling, reference architectures (DGX systems and SuperPOD-style designs) |
| Operations | Cluster schedulers (Slurm, Kubernetes with GPU operators), GPU monitoring (DCGM), utilisation and health metrics, Multi-Instance GPU (MIG) and virtual GPUs for sharing |
Study loop
- Turn each objective in the study guide into a one-line "problem → component" pair (e.g. "multi-node training stalls on gradient exchange → faster interconnect, RDMA").
- Read the NVIDIA documentation overview page for each named product; note its purpose, not its configuration flags.
- Practise scenario questions and justify every option.
- Revisit terms that sound alike (NVLink vs NVSwitch, MIG vs vGPU, DPU vs SmartNIC) until you can contrast them in one sentence.
Writing practice questions
- Ground each answer in NVIDIA documentation or the original papers for general AI concepts; verify any arithmetic (memory per parameter, bandwidth estimates) by recomputing.
- Use scenarios with different bottlenecks rather than definition lists.
Pitfalls
- Treating inference and training requirements as the same.
- Confusing scale-up (inside a node) with scale-out (across nodes) networking.
- Assuming more GPUs always helps; data loading and communication often become the bottleneck.
- Mixing up virtualisation options: MIG partitions one GPU in hardware; vGPU shares through a hypervisor.
- Product names change across generations; learn the role, then check current names.
Self-check
- A practice set within 60 minutes with clear margin.
- Draw a multi-node training cluster from storage to GPUs and label each interconnect.
- Explain three ways to share a GPU among workloads and when to use each.