Loading demo links...
What to study for each NCA-ADS domain
Data Manipulation and Preparation
Covers preparing tabular data on the GPU with cuDF, including reading CSV and Parquet, filtering, sorting, joins, groupby aggregations, handling missing values and duplicates, dtype and datetime handling, categorical data, and scaling preparation with Dask-cuDF. It emphasizes vectorized, pandas-compatible operations performed in parallel on the GPU.
Key skills
Reading CSV and Parquet data into cuDF, filtering, sorting, joining and grouping GPU DataFrames, handling missing values, and converting between pandas and cuDF.
Machine Learning With RAPIDS
Covers training and evaluating machine learning models with cuML using its scikit-learn-like API, including random forests, linear and logistic models, KMeans, DBSCAN, PCA and UMAP. It also addresses train/test splitting, scaling without leakage, cross-validation and handling imbalanced data on the GPU.
Key skills
Training cuML models with the scikit-learn-style API, random forests and gradient boosting, KMeans clustering, train/test evaluation, and GPU versus CPU model workflows.
Data Science Pipelines and Workflow Automation
Covers building repeatable GPU data pipelines as dependency-ordered workflows, including DAG concepts, Dask-based distributed execution, Parquet checkpoints, idempotent tasks and fail-fast validation. It also addresses triggers, scheduling, caching, alerting and reproducibility through version and parameter control.
Key skills
Building DAG-based GPU pipelines, scheduling Dask-backed workflows, parameterising repeatable runs, and automating data preparation and training steps.
Descriptive Analysis and Visualization
Covers summarizing and exploring data with descriptive statistics, value counts, correlation, histograms, box plots and pivot tables computed with cuDF. It also covers interactive and large-scale visualization with cuxfilter, Plotly and Datashader for GPU-accelerated exploration.
Key skills
Computing descriptive statistics and correlations in cuDF, building histograms and box plots, and visualising GPU-resident data with cuXfilter and plotting libraries.
Foundations of Accelerated Data Science
Covers why GPUs accelerate data science through massive parallelism and the CUDA SIMT model, the core RAPIDS libraries cuDF, cuML and cuGraph, and host-device memory transfer costs. It also addresses GPU memory capacity, data parallelism, and when small or sequential workloads gain little from acceleration.
Key skills
Explaining GPU parallelism and the CUDA SIMT model, the RAPIDS stack (cuDF, cuML, cuGraph), zero-copy interoperability via Apache Arrow, and when GPU acceleration pays off.
Introductory MLOps Practices
Covers introductory MLOps concepts including experiment tracking with MLflow, model registries and versioning, model serving with NVIDIA Triton, and monitoring for data and concept drift. It also addresses CI for ML projects, feature stores and recording dataset versions and seeds for reproducibility.
Key skills
Tracking experiments with MLflow, registering and versioning models, packaging models for deployment, and monitoring served models for drift.
Advance Data Structures
Covers data structures that underpin GPU analytics, including columnar memory layout in cuDF, the Apache Arrow format for zero-copy sharing, and hash tables for parallel joins and groupby operations. It also addresses dense versus sparse representations, dictionary encoding and graph representations in cuGraph.
Key skills
Columnar memory layout and Apache Arrow interoperability, GPU hash tables, string and datetime representations in cuDF, and memory-efficient data structures for analytics.
Software and Environment Management
Covers setting up and verifying a RAPIDS environment using conda environments or official container images, and checking driver, CUDA and RAPIDS version compatibility. It also covers using nvidia-smi to inspect GPUs and running a smoke test to confirm cuDF can use the GPU.
Key skills
Creating conda environments for RAPIDS, using official NVIDIA containers, verifying GPU/driver/CUDA versions, and resolving common environment issues.