Loading demo links...
What to study for each NCP-ADS domain
Data manipulation and software literacy
Covers GPU-accelerated data manipulation with cuDF and Dask-cuDF, including joins, groupby operations, string and datetime handling, categorical encoding, and zero-copy interchange across RAPIDS libraries. It also addresses software literacy practices such as pinned reproducible environments, profiling, and collaborative notebook workflows.
Key skills
Advanced cuDF and Dask-cuDF operations, complex joins and groupbys, string processing on GPU, and writing efficient, vectorised data code at scale.
MLOps
Covers production machine learning operations for GPU workloads, including experiment tracking, model registries, and deployment to inference servers such as NVIDIA Triton. It also addresses drift monitoring, CI/CD quality gates, feature-store consistency, orchestration, rollback, and data validation in pipelines.
Key skills
Production ML pipelines for GPU workloads, experiment tracking and model registries, CI/CD for models, deployment patterns, and monitoring for drift and performance decay.
Data preparation
Covers preparing large datasets on the GPU, including missing-value treatment, outlier and anomaly detection, encoding high-cardinality categories without leakage, and stratified splitting for imbalanced data. It also addresses scaling, distributed deduplication and joins under skew, text feature preparation, temporal leakage, and label noise.
Key skills
Treating missing values and outliers in large GPU datasets, anomaly detection, feature engineering at scale, and validating data quality before training.
GPU and cloud computing
Covers running accelerated workloads on GPU and cloud infrastructure, including CUDA memory management, Dask clusters with one worker per GPU, MIG partitioning, NVLink, and Kubernetes GPU scheduling. It also addresses container and driver compatibility, checkpointing on preemptible instances, DCGM monitoring, and autoscaling for cost control.
Key skills
CUDA memory management, Dask cluster deployment, running RAPIDS on cloud GPU instances and Kubernetes, and cost/performance trade-offs of GPU infrastructure.
Machine learning
Covers GPU-accelerated machine learning with cuML and GPU XGBoost, including tree ensembles, clustering, dimensionality reduction, and linear models. It also addresses leakage-safe cross-validation, distributed hyperparameter tuning with pruning, early stopping, nearest-neighbour search at scale, ranking metrics, and mixed-precision training.
Key skills
cuML algorithms at depth, GPU XGBoost, clustering and dimensionality reduction (UMAP, HDBSCAN), hyperparameter tuning, and model evaluation at production scale.
Data analysis
Covers production-grade analysis of large GPU-resident datasets, including full-data aggregation, hypothesis testing with multiple-comparison control, and distinguishing correlation from causation. It also addresses cross-filtered exploration with cuxfilter, cohort and time series analysis, A/B test interpretation, drift measurement with PSI, geospatial analysis, and reproducible reporting.
Key skills
Full-data aggregation over GPU-resident datasets, hypothesis testing, statistical validation of results, and communicating findings from large-scale analysis.