NCA Foundations study plan

Diagram the NVIDIA platform stack from hardware through drivers and libraries, labeling where CUDA, cuDNN, TensorRT, and NGC sit.

Compare GPU and CPU execution models by explaining latency hiding, throughput orientation, and why SIMT suits dense math.

Identify the major GPU generations and describe what changed in tensor cores, memory technology, and interconnect across them.

Compute arithmetic intensity for a sample workload and classify it as memory-bound or compute-bound.

Explain the CUDA execution hierarchy from thread to grid and map it onto streaming multiprocessors and warps.

Trace how a CUDA program moves through compilation, host-device transfer, kernel launch, and synchronization.

Distinguish the roles of CUDA Runtime, cuBLAS, cuDNN, NCCL, and CUDA driver APIs in a training run.

Describe warp divergence and shared memory bank conflicts, and propose one fix for each.

Outline the end-to-end deep learning workflow from data preparation through training, evaluation, and deployment.

Explain forward and backward propagation and how gradients flow through a simple network.

Compare data-parallel and model-parallel training and identify which libraries support each.

Set up a containerized environment using NGC images and verify GPU visibility with a diagnostic query.

Describe how batch size, learning rate, and mixed precision interact during training.

Explain the purpose of data loaders, augmentation pipelines, and GPU-accelerated preprocessing.

Compare CPU-only pandas workflows with GPU-accelerated dataframe libraries on the same dataset.

Identify where data movement bottlenecks occur between storage, host memory, and device memory.

Apply GPU-accelerated libraries to a tabular and a graph or array workload and measure speedup.

Explain the difference between training and inference requirements in latency, throughput, and precision.

Describe quantization, pruning, and layer fusion and state the tradeoff each makes.

Outline how a trained model is converted and optimized for deployment with TensorRT.

Compare batch and real-time inference serving patterns and their scaling characteristics.

List the key GPU health metrics to monitor and what each indicates about workload behavior.

Use NVIDIA management and monitoring tooling to capture utilization, memory, temperature, and power data.

Diagnose three common failure modes such as out-of-memory errors, thermal throttling, and driver mismatches.

Explain multi-GPU and multi-node communication and why interconnect bandwidth matters for scaling.

Describe common deployment topologies for inference at the edge, in the data center, and in the cloud.

Match NVIDIA solution families to representative use cases in healthcare, autonomous systems, and enterprise AI.

Summarize professional practices for reproducible experiments, versioning, and responsible AI review.

Consolidate all domains into a single concept map linking architecture, software stack, workflow, and operations.

Complete a timed practice assessment and categorize every miss by domain.

Re-teach the three weakest domains in your own words without notes.

Review flagged uncertain facts and confirm each against primary documentation.