Skip to content

NCP-ADS Accelerated Data Science Professional Study Strategy

Study Approach

Phase 1: Foundation (1-2 weeks)

  1. RAPIDS Ecosystem - Components, installation, CPU equivalents
  2. cuDF - Data loading, transformations, pandas differences
  3. cuML - Algorithms, API patterns, model evaluation
  4. 📖 RAPIDS Docs

Phase 2: Advanced (2-3 weeks)

  1. cuGraph - Graph construction, algorithms, integration with cuDF
  2. Dask-cuDF - Multi-GPU scaling patterns
  3. Spark RAPIDS - Configuration, supported operations, tuning
  4. 📖 Spark RAPIDS Docs

Phase 3: Exam Prep (1-2 weeks)

  1. Practice code-based questions
  2. Review API patterns and algorithm selection
  3. Focus on performance optimization scenarios

Exam Tactics

Keywords

  • "DataFrame" or "tabular" - cuDF
  • "Classification/clustering/regression" - cuML
  • "Graph" or "network" - cuGraph
  • "Large dataset" or "multi-GPU" - Dask-cuDF
  • "Spark" or "existing pipeline" - RAPIDS Accelerator
  • "Memory" - RMM, managed memory, chunking
  • "Fastest format" - Parquet

Common Pitfalls

  • cuDF is not 100% pandas-compatible - custom .apply() needs Numba
  • cuML API mirrors scikit-learn but input should be cuDF/cuPy
  • cuGraph requires integer vertex IDs for best performance
  • Dask-cuDF operations are lazy - need .compute() to execute
  • Spark RAPIDS is a plugin, not a code rewrite
  • Parquet is always faster than CSV for GPU loading

Self-Assessment Questions

  • Can I map each data science task to the correct RAPIDS component?
  • Do I know cuDF API for loading, filtering, groupby, and merge?
  • Can I select the right cuML algorithm for a given problem?
  • Do I understand cuGraph construction and key algorithms?
  • Can I configure Spark RAPIDS for GPU acceleration?
  • Do I know when to use cuDF vs Dask-cuDF vs Spark RAPIDS?