Free NCA-AIIO practice test: 30 questions with answers

GPUs contain thousands of smaller cores designed for data-parallel arithmetic, which suits the matrix multiplications used in neural network training. CPUs are optimised for sequential, branchy workloads, and GPUs do not provide persistent storage.

Inference uses a trained model with frozen weights to compute outputs for new inputs, which needs only the forward pass. Backpropagation and weight updates happen in training, and labelling and cleaning are data preparation steps.

The batch size is the number of samples processed together before one weight update. The learning rate sets the step size, the epoch count is the number of full passes over the dataset, and dropout is a regularisation rate.

A large gap between training and validation performance is the classic sign of overfitting. Underfitting would show poor performance on both sets, and memory limits would stop training rather than change the gap.

Transfer learning reuses knowledge from a pretrained model and adapts it with fine-tuning on a smaller dataset. Training from scratch on a small set usually overfits, while quantization and batch size search do not adapt the model to new data.

Mixed precision performs most operations in FP16 or BF16 on Tensor Cores for speed and memory savings, while keeping FP32 master weights for stability. Full FP64 is slower, and integer-only training without scaling is unstable.

The F1 score is the harmonic mean of precision and recall, so it penalises a model that achieves high accuracy only by predicting the majority class. Mean squared error is a regression loss, while learning rate and batch size are training settings rather than evaluation metrics.

Supervised learning uses labelled examples to learn a mapping from inputs to known outputs. It needs training data, it is used across data types, and it covers regression as well as classification.

LLMs are well suited to summarising and drafting natural-language text. Motor control and in-camera video compression rely on different control and signal-processing systems rather than language models.

Gradient descent and its variants compute the gradient of the loss with respect to the weights and step in the negative gradient direction. PCA is a dimensionality reduction method, K-means is a clustering algorithm, and beam search is a decoding strategy used at inference time.

The NVIDIA GPU Operator automates the lifecycle of the GPU software stack on Kubernetes nodes, including drivers, the container toolkit and the device plugin. Triton serves models, RAPIDS accelerates data science, and NeMo is a framework for building generative models.

DCGM collects and exports GPU health and telemetry data, and it integrates with monitoring systems such as Prometheus. TensorRT optimises inference engines, the NGC catalog hosts software, and cuDNN is a deep learning primitives library.

Triton Inference Server hosts models from multiple frameworks, supports dynamic batching and concurrent execution, and exposes standard inference endpoints. cuDF is a dataframe library, Nsight Systems is a profiler, and DGX Cloud is a hosted service.

The device plugin exposes GPUs as the extended resource nvidia.com/gpu, and requesting it under resource limits schedules the pod onto a node with a free GPU. An environment variable alone does not reserve hardware, and host paths and CPU selectors do not allocate GPUs.

NVIDIA Run:ai provides GPU orchestration and scheduling for AI workloads, including fractional GPU allocation and quota management. Nsight Compute profiles kernels, Riva is a speech AI toolkit, and NeMo Guardrails enforces conversational rules.

NGC containers bundle validated CUDA and framework stacks for NVIDIA GPUs, which keeps environments consistent across nodes. Manual installs drift over time, and Git LFS and spreadsheets do not provide runnable validated environments.

High GPU utilization with a growing queue while other resources are idle points to the GPU being the constraint, so more GPU capacity, batching, or a smaller model is needed. Disk, DNS and client-side network speed are not indicated by these symptoms.

Containers give reproducible environments and bundle application dependencies so the same image runs across nodes. The host still needs a compatible GPU driver, an orchestrator is still needed at scale, and containers do not tune hyperparameters.

Uncorrectable ECC errors indicate memory faults that can corrupt results, and GPU telemetry such as DCGM exposes them. The other items are unrelated to GPU hardware health.

A model registry stores versioned models with their metadata and lifecycle stages, which supports promotion and rollback. A PDF, a laptop folder and firewall rules do not provide versioned model governance.

NVLink with NVSwitch provides high-bandwidth GPU-to-GPU links inside a system, avoiding the slower PCIe path through the CPU. PCIe is lower bandwidth, and wireless and USB connections are not used for GPU-to-GPU traffic.

BlueField data processing units run networking, storage and security services on the adapter, freeing host CPU cycles. Tensor Cores are compute units inside the GPU, NVLink bridges connect GPUs, and NVMe drives store data.

A parallel file system spreads data across many storage servers and delivers high aggregate throughput to many clients at once, which keeps GPUs fed. A single export on a slow link becomes a bottleneck, and removable or ad hoc copies do not scale.

MIG partitions a supported GPU into isolated instances, each with dedicated memory and compute resources. NVLink connects GPUs, GPUDirect Storage moves data between storage and GPUs, and the container toolkit exposes GPUs to containers without partitioning them.

Direct liquid cooling carries heat away from high-power components more efficiently than air, which allows higher rack densities. Removing GPUs defeats the purpose, extra fans in the same design give limited gains, and disabling thermal management risks hardware damage.

InfiniBand and RDMA over Converged Ethernet provide the low-latency, high-bandwidth communication that distributed training depends on. Bluetooth, dial-up and mail relays do not meet these requirements.

GPUDirect RDMA lets network adapters move data directly to and from GPU memory, bypassing CPU copies on the path. cuDF and NVML do not transfer data over the network, and the Container Toolkit exposes GPUs to containers.

Dataset size and checkpoint frequency determine capacity, and aggregate throughput determines whether the GPUs wait for data. Physical rack appearance and administrator peripherals do not affect storage sizing.

DGX systems and DGX SuperPOD architectures integrate GPUs, networking and software for large-scale AI workloads. A gaming card, a Jetson embedded module and a DPU are components or products aimed at different use cases.

PUE is total facility power divided by IT equipment power, where values closer to 1.0 indicate less overhead for cooling and distribution. The other metrics measure reliability, caching and network performance respectively.