This plan is for someone who already works with pandas and scikit-learn and can study about six to eight hours a week. It covers NVIDIA's eight NCA-ADS sections in three weeks, grouped so that topics that depend on each other sit together, and leaves the fourth week for timed practice.
The reading comes from the suggested reading NVIDIA lists for each section in the NCA-ADS study guide, and the courses from the learning path on the certification page.
Before week 1: find where you stand
Answer a set of questions cold, before any reading. NVIDIA publishes no sample questions for NCA-ADS (as of October 2026), so use the 20 NCA-ADS practice questions, or take one of the practice tests under time if you already have them.
Write down the sections where you missed questions. The plan below gives each group of sections a week, and your list tells you where to spend the extra hours inside each week.
Try 15 NCA-ADS practice questions free
They come from the NCA-ADS practice tests, and each option shows its explanation once you answer.
- 6 timed practice exams
- Every answer explained
- A score for each domain
- $19.99 one-time
Week 1: data preparation and the foundations (35% of the exam)
Data Manipulation and Preparation (23%) and Foundations of Accelerated Data Science (12%) belong together: the foundations explain why cuDF, Dask and the GPU behave as they do, and data preparation is where that behavior shows.
Foundations. Start with the Python data stack: NumPy vectorization and broadcasting, pandas and Jupyter. Then learn what accelerated computing is, which workloads a GPU speeds up (large, parallel operations on whole columns) and which ones it does not (small tables, row-by-row loops, frequent copies between host and device), how cudf.pandas falls back to the CPU, and when one GPU is enough and when Dask or Spark spreads the work. Finish with supervised versus unsupervised learning and model parameters versus hyperparameters.
- Data Science With Python Tutorial (GeeksforGeeks)
- Comparison of cuDF and pandas
- What Is Accelerated Computing?
- Supervised vs. unsupervised learning
Data preparation. Practice joins, reshaping and cleaning in pandas and cuDF, including missing values, duplicates and types. Learn GPU ETL with Dask and the RAPIDS Accelerator for Apache Spark, feature engineering for numerical and categorical variables, class imbalance, resampling and synthetic data, dimensionality reduction with PCA and UMAP, sampling, and Parquet: reading only the columns and row groups you need, and compression.
- NVIDIA cuDF: GPU-accelerated DataFrames
- GPU-accelerated DataFrames for pandas users
- RAPIDS cuDF for feature engineering (KDnuggets)
- Best practices for multi-GPU data analysis using RAPIDS with Dask
- 10 techniques to deal with class imbalance (Analytics Vidhya)
The study guide recommends two courses for this section: Accelerating End-to-End Data Science Workflows and Best Practices in Feature Engineering for Tabular Data With GPU Acceleration.
Clean data, engineer features and move a pipeline to the GPU
Pick one or two in place of a reading session. The RAPIDS lab runs on an NVIDIA GPU pod; the other three run in the browser without a GPU. All four come with Pro.
- Open labGPU-Accelerated Data Science with RAPIDSintermediate 40 minGPU sandbox
- Open labData Cleaning and Feature Engineering: Predict Rents From a Messy Exportbeginner 50 minHosted
- Open labImbalanced Data and Anomaly Detection: Average Precision, Rebalancing, Alert Budgets and Unseen Failuresintermediate 55 minHosted
- Open labCustomer Segmentation with k-means and PCA: Scaling, Choosing k, Stability and Deploymentintermediate 50 minHosted
End of the week: answer questions 1 to 5, 15 and 16 of the practice questions article. Reread the explanation of every miss.
Week 2: machine learning and pipelines (29%)
Machine Learning With RAPIDS (16%) and Data Science Pipelines and Workflow Automation (13%) share most of their vocabulary: splits, folds, preprocessing steps and the leaks between them.
Machine learning. Train models with cuML and with XGBoost on the GPU, and learn how cuml.accel runs unchanged scikit-learn code. Cover regression, classification and clustering, and the tools for judging them: cross-validation splitters for shuffled, grouped and time-ordered data, hyperparameter search (grid, random, successive halving), and the metrics read from a confusion matrix, from precision and recall to the ones that hold up when one class is rare.
- Machine learning with GPU-accelerated pandas and scikit-learn
- NVIDIA cuML brings zero-code-change acceleration to scikit-learn
- XGBoost and random forest GPU hyperparameter tuning with Dask
- scikit-learn: common pitfalls and recommended practices
- The confusion matrix, for example in scikit-learn's documentation
Pipelines. Keep every fitted step inside a pipeline so cross-validation refits it on each fold, and learn feature selection and transformation, the fixes for underfitting and overfitting, integrating external data without leaking the future, scaling work out with Dask-CUDA, and making runs reproducible with seeds, pinned versions and environments. The study guide also adds the data flywheel, where production data keeps improving deployed models.
- Scaling data pipelines: AT&T optimizes speed, cost and efficiency with GPUs
- Best practices for multi-GPU data analysis using RAPIDS with Dask
- RAPIDS: learn more
- NVIDIA AI Workbench data science quickstart
For both sections the study guide recommends Accelerating End-to-End Data Science Workflows, the course from week 1.
Train, cross-validate and tune, then automate the pipeline
All four run in the browser without a GPU, and you can try the first step of each before you subscribe to Pro.
- Open labYour First Machine Learning Model: Predict Churn with scikit-learnbeginner 50 minHosted
- Open labOverfitting and Cross-Validation: Find What Really Makes a Better Cupintermediate 45 minHosted
- Open labGradient Boosting with XGBoost: Predict Late Deliveries and Explain Every Predictionintermediate 50 minHosted
- Open labTraining Pipeline Orchestration: Ordering, Retries, Failure Isolation and Resumeintermediate 45 minHosted
End of the week: answer questions 6 to 11 of the practice questions article, then take your first full practice test under time. Sort the misses by section and keep the list for week 4.
Three questions from the NCA-ADS practice tests
cuML HDBSCAN on device telemetry returns 4,000 clusters, most with 5 to 10 members, and analysts want only substantial behavior groups of at least 500 devices, with smaller groupings treated as noise. The model uses default settings. Which parameter change fits?
[object Object]
Correct. min_cluster_size is the minimum number of samples for a group to count as a cluster; smaller groupings are left as noise. The default is 5, which explains the many tiny clusters.
A. A different distance changes neighborhoods, but it does not set a minimum cluster size.
B. HDBSCAN does not take a number of clusters. It finds clusters from density.
D. 0.0 is already the default for this setting, and it is not a size threshold.
Months after a model shipped, a team finds its metrics in a spreadsheet but cannot tell which version of the pipeline code, which parameters, or which output files produced them. It is adopting MLflow for future runs. Which MLflow component records these for each run?
[object Object]
Correct. MLflow Tracking is an API and UI for logging parameters, code versions, metrics and output files when running machine learning code, and for visualizing results later.
A. The registry versions models and their lineage. Logging code versions, parameters and metrics happens in Tracking.
C. Projects package code and environments to run. They do not replace tracking of results.
D. The MLmodel file describes flavors, signature and environment. Metrics are tracked per run.
An analyst loads a new 22-million-row weather table into cuDF and, before any statistics, wants one call that shows the number of rows, each column's name and dtype, non-null information and memory usage. Which call fits, as in NVIDIA's cuDF EDA walkthrough?
[object Object]
Correct. The walkthrough starts with .info, which works the same as in pandas and shows the entry count, the columns with their dtypes and the memory usage.
A. head shows the first rows of data. Dtypes and memory usage need info, which summarizes the frame.
C. describe reports statistics such as mean and quartiles for numeric columns. It does not show memory use.
D. shape returns only the row and column counts, without dtypes, non-null counts or memory use.
Full NCA-ADS set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsWeek 3: analysis, MLOps, time series, graphs and environments (36%)
Four smaller sections share this week: Descriptive Analysis and Visualization (13%), Introductory MLOps Practices (10%), Advanced Data Structures (7%) and Software and Environment Management (6%).
Descriptive analysis and visualization. Review describe, value counts, crosstabs and correlations, then choosing a chart for a goal and the GPU tools for large data: hvPlot, Datashader, cuxfilter and Plotly Dash. For statistics, know which test fits paired or independent samples and counts, how significance differs from the size of an effect, and how rolling windows show a trend.
- A guide to data visualization with RAPIDS
- Making a Plotly Dash census visualization powered by RAPIDS
- pandas documentation
- Data exploration and visualization tools (neptune.ai)
- Speed up data exploration with RAPIDS cuDF
MLOps. Learn the Prometheus metric types and how histograms give latency percentiles and averages, experiment tracking with MLflow and Weights & Biases, saving and loading models, data, prediction and concept drift when labels arrive late, versioning data, models and environments, and fair benchmarks for choosing hardware.
- A guide to monitoring machine learning models in production
- Prometheus: histograms and summaries
- ML model monitoring in production: best practices (Datadog)
- MLflow: hyperparameter tuning and experiment tracking
- Mastering version control for ML models (DagsHub)
Time series and graphs. For time series: time-ordered splits, forecast metrics and their traps, lag and rolling features that never peek at the future, and resampling and interpolating irregular timestamps. For graphs: loading an edge list into cuGraph, ranking nodes by importance (degree, betweenness, PageRank, Katz), finding communities and shortest paths, and laying out a network to visualize it. NVIDIA's Accelerated Machine Learning for Time Series Forecasting course is part of the learning path for this section.
- Accelerating time series forecasting with RAPIDS cuML
- The difference between graph and relational databases (AWS)
- Beginner's guide to GPU-accelerated graph analytics in Python
Software and environment management. Practice environment files, lockfiles and pip requirements, Docker with the NVIDIA Container Toolkit, reading nvidia-smi (the CUDA version it reports is the newest one the driver supports), choosing which GPUs a process or container sees, solving a dependency conflict and the basic Git commands.
- Get started with GPU acceleration for data science
- nvidia-smi documentation
- Conda environments guide
- NVIDIA Container Toolkit: specialized configurations with Docker
- Basic Git commands
Track experiments, watch for drift, forecast and ship a container
The MLflow and container labs run on NVIDIA GPU pods, and the drift and forecasting labs run in the browser. Descriptive statistics, visualization and graph analytics have no lab yet, so their reading carries that part of the week.
- Open labMLflow Experiment Tracking: From Single Run to Team Workflowintermediate 35 minGPU sandbox
- Open labML Model Monitoring: Data Drift with PSI, Delayed Labels, Alert Rules and Root Causeintermediate 50 minHosted
- Open labTime-Series Forecasting: Baselines, Backtesting, Calendar Features, Prediction Intervals and the Newsvendor Ruleintermediate 50 minHosted
- Open labGPU Container Lifecycle: Build, Test, Ship, Rollbackintermediate 40 minGPU sandbox
End of the week: answer questions 12 to 14 and 17 to 20 of the practice questions article, then take a second timed practice test and compare the section scores with the first.
Week 4: timed practice and booking the exam
Alternate days: a timed practice test, then a day on the weakest section from that result, using the explanations of your misses and the reading for that section. Expect scores to move between sections from one test to the next, since a single test samples each section with only a few questions.
Book the exam once two practice tests you have not taken before both come back at 80% or higher. NVIDIA does not publish a passing score, so the margin protects you from a hard draw of questions.
For the exam itself you need a Certiverse account. Read the certification page again in the last week for any changes.
If your background is different
GPUs are new to you. Add a week between weeks 1 and 2 for the foundations: run your own pandas notebooks under cudf.pandas, watch what falls back to the CPU, and try a Dask-CUDA cluster. NVIDIA's Accelerating End-to-End Data Science Workflows is self-paced and also runs as an instructor-led workshop.
Your statistics are rusty. Spend two extra sessions in week 3 on hypothesis tests: which test fits paired or independent samples and categorical counts, and how to read a p-value next to the size of an effect.
You know machine learning well but not RAPIDS. Keep the plan, and give week 1 more time on where cuDF differs from pandas (iteration, the order of results, unsupported types) and on the Dask settings for GPUs.
Frequently asked questions
Is four weeks enough to prepare for NCA-ADS?
This plan assumes it is for someone who already works with pandas and scikit-learn and studies six to eight hours a week. If GPUs or statistics are new to you, add a week where your first practice test shows the most misses.
Do I need NVIDIA's courses to pass NCA-ADS?
The study guide recommends Accelerating End-to-End Data Science Workflows for every section, but the exam does not require it. The suggested reading covers the same topics, and the course adds guided hands-on practice.
Are there hands-on labs for NCA-ADS?
Yes, with Preporato Pro: labs for data cleaning and features, class imbalance, a RAPIDS pipeline on a GPU, scikit-learn and XGBoost with cross-validation, pipeline automation, MLflow tracking, drift monitoring, time series backtesting and Docker, each checked step by step in the browser. The ones that fit each week are listed in the plan above. Descriptive statistics, visualization and graph analytics have no lab yet.
Which NCA-ADS section should I study first?
Data Manipulation and Preparation, the largest section at 23%, together with the foundations it builds on. That is why they open this plan.
How many practice tests should I take before NCA-ADS?
Take them until two tests you have not seen before both reach 80% or higher. In this plan that means one at the end of week 2, one at the end of week 3, and three or four in week 4.
Where can I find NCA-ADS practice tests?
The NCA-ADS practice tests overview lists what exists. The free questions on the certificate page come from the full preporato.com practice tests.
Sources: