Pass the NVIDIA NCA-AIIO — the First Time

1. Why are GPUs generally preferred over CPUs for training deep learning models?

   • A. GPUs run at considerably higher clock speeds than server CPUs do
   • B. GPUs run operating systems far more efficiently than CPUs do
   • C. GPUs have far more system memory available to them than the host CPUs have
   • D. GPUs have thousands of cores running the same operation in parallel

   Answer: D. Deep learning training is dominated by matrix and tensor operations that can be split across thousands of identical calculations. A GPU's massively parallel architecture executes these simultaneously, while a CPU's smaller number of complex cores is optimized for sequential, branching work: not raw parallel throughput. The other options are false: GPUs trade lower single-thread performance for much greater total throughput, so clock speed is not their advantage; the operating system runs on the host CPU; and GPU memory is usually smaller than host memory (a DGX H100 pairs 80 GB per GPU with 2 TB of system memory).

   Exam domain: AI Infrastructure

2. What is the primary purpose of NVLink in a multi-GPU server?

   • A. To manage the GPUs' power delivery and their thermal limits
   • B. To encrypt all traffic between the CPU and the system memory
   • C. To connect the server directly to the external storage arrays in the rack
   • D. To give GPUs a high-bandwidth direct link, much faster than PCIe

   Answer: D. NVLink is NVIDIA's direct GPU-to-GPU interconnect. It lets GPUs in the same system exchange data at far higher bandwidth and lower latency than the PCIe bus allows, which matters when a model or its activations are spread across multiple GPUs. NVLink does not manage power or thermals, encrypt traffic between the CPU and system memory, or attach the server to external storage arrays.

   Exam domain: AI Infrastructure

3. A team is building a cluster where a single training job will run across dozens of GPU nodes. Which network characteristic matters most for the fabric connecting those nodes?

   • A. Support for legacy 1 GbE management traffic between the nodes
   • B. Wireless connectivity between the racks for flexible cabling
   • C. Deep packet inspection carried out at the edge of the network
   • D. High bandwidth with very low latency, via InfiniBand or RoCE

   Answer: D. Distributed training synchronizes gradients between nodes many times per second, so the inter-node fabric must deliver very high bandwidth at very low latency. That is why AI clusters use InfiniBand or RDMA-capable Ethernet (RoCE) rather than ordinary TCP/IP networking.

   Exam domain: AI Infrastructure

4. What is the main role of a DPU (Data Processing Unit), such as NVIDIA BlueField, in an AI data center?

   • A. To replace the GPUs entirely when running smaller training jobs
   • B. To compile the CUDA code faster than the host CPU is able to
   • C. To act as a backup power controller for the entire server rack
   • D. To offload networking, storage, and security from the host CPU

   Answer: D. A DPU is a programmable processor on the network card that takes over infrastructure workloads (software-defined networking, storage access, and security functions) freeing host CPU cycles for applications and providing isolation between infrastructure and tenant workloads.

   Exam domain: AI Infrastructure

5. A startup needs GPU capacity for a 3-week training experiment and has no existing data center. Which deployment approach best fits this situation?

   • A. Purchase a multi-node GPU cluster for on-premises installation
   • B. Wait for GPU prices to fall before starting the whole project
   • C. Rent GPU instances from a cloud provider for the experiment
   • D. Build a new colocation facility fitted with liquid cooling

   Answer: C. Cloud infrastructure is elastic: you can provision GPU capacity in minutes, pay only for the weeks you use it, and release it afterward. Buying hardware makes more sense for sustained, predictable workloads where utilization stays high over years, not a short experiment.

   Exam domain: AI Infrastructure

6. Why are liquid-cooling solutions increasingly used for racks of modern AI servers?

   • A. Liquid cooling is required before any of the GPUs will even boot
   • B. Liquid cooling eliminates the need for electrical power distribution
   • C. Dense GPU racks exceed the heat air cooling can remove effectively
   • D. Air cooling interferes with the high-speed networking signalling

   Answer: C. Modern GPU servers concentrate tens of kilowatts of power in a single rack: far beyond typical enterprise racks. At those densities, air cooling becomes inefficient or insufficient, so facilities turn to liquid cooling (direct-to-chip or rear-door heat exchangers) to remove heat reliably.

   Exam domain: AI Infrastructure

7. In a typical AI training cluster, what is the role of the login (head) node?

   • A. It translates GPU instructions into CPU instructions for each job
   • B. It stores all of the cluster's training datasets on its own local NVMe disks
   • C. It is where users connect, prepare jobs, and submit them to the scheduler
   • D. It runs all of the model training computation on behalf of the cluster's users

   Answer: C. Login/head nodes are where users land: they hold development tools and the job scheduler interface. Actual training runs on GPU compute nodes; datasets live on shared storage. Keeping heavy work off the login node keeps the cluster responsive for everyone. No node translates GPU instructions into CPU instructions: device code runs on the GPUs themselves.

   Exam domain: AI Infrastructure

8. Which statement correctly describes the relationship between AI, machine learning (ML), and deep learning (DL)?

   • A. ML is a subset of DL, which is in turn a subset of the much broader field of AI
   • B. AI is a subset of ML, while DL is a separate field unrelated to both
   • C. AI, ML, and DL are interchangeable names for the same technique
   • D. DL is a subset of ML, which is itself a subset of the broader field of AI

   Answer: D. AI is the broad field of making machines act intelligently. Machine learning is the subset where systems learn patterns from data instead of being explicitly programmed. Deep learning is a further subset of ML that uses many-layered neural networks.

   Exam domain: Essential AI Knowledge

9. Which statement best contrasts AI training with AI inference?

   • A. Training requires no data; inference requires the full training dataset
   • B. Training answers user requests in real time; inference builds the model
   • C. Training happens on CPUs only, whereas inference happens on GPUs only
   • D. Training is long-running and throughput-bound; inference is latency-bound

   Answer: D. Training iterates over large datasets for hours to weeks to produce model weights: throughput is what matters. Inference applies the finished model to new inputs, where latency and cost-per-prediction usually drive the architecture. Both can use GPUs, but they are sized and optimized differently.

   Exam domain: Essential AI Knowledge

10. Which combination of factors is most credited with the rapid progress of AI over the last decade?

    • A. Quantum computers replacing classical servers in data centers
    • B. Cheaper hard drives, faster web browsers, and office software
    • C. Massive datasets, affordable GPU compute, and better algorithms
    • D. One single breakthrough algorithm that was discovered in 2020

    Answer: C. Modern AI took off when three curves crossed: huge digitized datasets to learn from, GPU computing making large-scale parallel training affordable, and algorithmic advances (deep learning architectures) able to exploit both. No single factor explains it alone.

    Exam domain: Essential AI Knowledge

Want more? Finish the quiz above to unlock 10 more free questions, or open the full NCA-AIIO practice app.