CertArea
All NVIDIA exams

Free NCP-ADS practice test: 30 questions with answers

30 original questions for NVIDIA-Certified Professional: Accelerated Data Science, drawn from every topic area. Answer each one to see whether you got it right and why. No sign-up.

Free questions
30
Full bank
100 questions
Real exam
120 min
Passing score
Pass/fail

NCP-ADS exam guide: topics, cost and format

All 30 NCP-ADS questions with answers

The same questions as the test above, grouped by exam topic. Open any question to see the answer and explanation.

Accelerated Data Processing (4 questions)

  1. Question 1

    A data scientist has a pandas workflow that is slow on a 50 GB dataset. Which NVIDIA library provides a GPU-accelerated dataframe API with a similar interface?

    1. AcuDF
    2. BTriton Inference Server
    3. CcuGraph
    4. DcuOpt
    Show answer and explanation
    Correct answer: A

    cuDF provides a GPU dataframe API that mirrors much of pandas for acceleration. cuGraph handles graph analytics, cuOpt solves optimization problems, and Triton serves models.

  2. Question 2

    Which library in the RAPIDS suite provides GPU-accelerated machine learning algorithms similar to scikit-learn?

    1. AcuML
    2. BcuDF only
    3. CcuGraph
    4. DNCCL
    Show answer and explanation
    Correct answer: A

    cuML offers GPU-accelerated versions of many scikit-learn style algorithms. cuDF handles dataframes, cuGraph handles graphs, and NCCL is a communication library.

  3. Question 3

    A pipeline reads a large CSV file and must parse it quickly on the GPU. Which approach is most efficient?

    1. ALoading the file into a Python list on the CPU
    2. BConverting the CSV into image files before parsing rows
    3. CReading one row at a time through a serial shell loop
    4. DReading the file directly into GPU memory with cuDF readers
    Show answer and explanation
    Correct answer: D

    cuDF readers parse files into GPU memory, avoiding CPU bottlenecks. CPU lists, image conversion and per-row shell loops are slow.

  4. Question 4

    A GPU dataframe operation runs out of memory on a large dataset. Which strategy helps most?

    1. AConverting the numeric columns to strings so the dataframe stores them as text
    2. BDisabling the GPU and running the full dataframe workload on CPU with pandas instead
    3. CAdding more CPU cores and system RAM to the workstation
    4. DProcessing the data in partitions or chunks that fit within GPU memory
    Show answer and explanation
    Correct answer: D

    Partitioned processing keeps working sets within GPU memory. CPU cores do not help GPU memory limits, string conversion increases size, and disabling the GPU removes acceleration.

GPU Machine Learning (4 questions)

  1. Question 5

    A team trains a gradient boosted tree model on GPU using a RAPIDS-compatible library. Which benefit is most expected compared with CPU training on a large dataset?

    1. AAutomatic accuracy improvement without hyperparameter tuning
    2. BLower GPU memory need because training data is compressed on the device
    3. CElimination of the need for training data via automatic feature engineering
    4. DFaster training through parallel GPU computation on large data
    Show answer and explanation
    Correct answer: D

    GPU parallelism speeds training for large data. Accuracy does not improve automatically, data is still required, and GPUs do not reduce memory needs by default.

  2. Question 6

    A GPU-accelerated model trains quickly but overfits. Which control helps most?

    1. AIncreasing the learning rate sharply at every epoch
    2. BRemoving the validation set to enlarge training
    3. CRegularization and early stopping based on a validation set
    4. DTraining for more epochs without monitoring loss curves
    Show answer and explanation
    Correct answer: C

    Regularization and early stopping on validation data reduce overfitting. Higher learning rates, removed validation and unmonitored training make overfitting worse.

  3. Question 7

    Hyperparameter tuning must cover many configurations on a GPU cluster. Which approach uses resources efficiently?

    1. AParallel trials with early termination of poor configurations
    2. BRandomly choosing one configuration per day
    3. CRunning configurations one after another with no early stopping
    4. DTesting only the default configuration
    Show answer and explanation
    Correct answer: A

    Parallel trials with early termination use cluster capacity well and stop poor runs. Sequential runs without stopping, default-only testing and daily random picks are inefficient.

  4. Question 8

    A GPU training job spends most time waiting for data. Which change is most likely to help?

    1. ADisabling batching so samples are fed one at a time
    2. BPrefetching and parallel data loading to keep the GPU busy
    3. CUsing a smaller GPU with fewer cores and less memory bandwidth
    4. DReducing the number of training samples per epoch
    Show answer and explanation
    Correct answer: B

    Prefetching and parallel loading address input bottlenecks. A smaller GPU, fewer samples and disabled batching do not solve input-bound training.

Deployment and Operations (4 questions)

  1. Question 9

    A GPU-accelerated feature pipeline runs nightly and must be monitored for failures and runtime changes. Which practice fits best?

    1. ADeleting logs after each run to free disk space on the GPU nodes
    2. BScheduled runs with logging, alerts on failure, and runtime tracking over time
    3. CRunning the pipeline manually when someone remembers, reviewing outputs later
    4. DIgnoring failures that happen at night, since jobs run unattended
    Show answer and explanation
    Correct answer: B

    Scheduled runs with logs, alerts and runtime tracking support reliable operations. Manual runs, ignored failures and deleted logs weaken operations.

  2. Question 10

    A model trained on GPU must be served with low latency. Which deployment consideration is most important?

    1. AServing on the same hardware used for data exploration during development sessions with notebook kernels attached
    2. BDeploying the training notebook directly to production, with its cell outputs and hard-coded paths intact
    3. CMatching the inference hardware and optimizing the model for it, such as conversion to an optimized engine
    4. DServing predictions from a shared spreadsheet that is refreshed by hand and emailed to stakeholders every week
    Show answer and explanation
    Correct answer: C

    Optimized inference and suitable hardware reduce latency. Exploration hardware, notebooks and spreadsheets are not production serving approaches.

  3. Question 11

    A data science team stores its models and datasets in versioned locations with metadata. Which benefit is most important?

    1. AReproducibility and the ability to roll back to a known version
    2. BEliminating the need for documentation, since tags replace written notes
    3. CMaking the files harder to find by spreading them across folders
    4. DStoring duplicates to confuse competitors and auditors
    Show answer and explanation
    Correct answer: A

    Versioned storage with metadata enables reproducibility and rollback. Harder discovery, no documentation and duplicates are harmful.

  4. Question 12

    A GPU cluster is shared by several teams. Which practice ensures fair and efficient use?

    1. AResource quotas and scheduling policies that allocate GPUs by priority and need
    2. BTurning off scheduling so jobs start in arrival order on each GPU node
    3. CAllowing one team to use every GPU in the cluster without quota limits during peak hours
    4. DAssigning GPUs at random each morning based on a shuffled queue
    Show answer and explanation
    Correct answer: A

    Quotas and scheduling policies distribute resources fairly and efficiently. Unlimited single-team use, random assignment and disabled scheduling cause conflicts and waste.

Model Evaluation and Metrics (3 questions)

  1. Question 13

    A binary classifier has an AUC close to 0.5 on a test set. What does this indicate?

    1. AThe model has no training data to learn from
    2. BThe model performs close to random ranking
    3. CThe model is perfect with flawless separation
    4. DThe model is too large to fit the GPU memory
    Show answer and explanation
    Correct answer: B

    AUC near 0.5 indicates little ability to rank positives above negatives. A perfect model is near 1.0, and data or size explanations are not implied by AUC.

  2. Question 14

    A regression model reports a low mean squared error on training data but a much higher error on new data. Which issue is most likely?

    1. AOverfitting to the training data
    2. BA GPU driver bug in the CUDA runtime
    3. CPerfect generalization to val
    4. DMissing labels in test data
    Show answer and explanation
    Correct answer: A

    A gap between training and new-data error suggests overfitting. Perfect generalization would show similar errors, and missing labels or driver bugs are not implied.

  3. Question 15

    Which metric is best for evaluating a ranking model where the top results matter most?

    1. AMean squared error between predicted and actual values
    2. BCount of rows processed per second by the loader
    3. CRoot mean squared error weighted equally across rows
    4. DNormalized discounted cumulative gain (NDCG)
    Show answer and explanation
    Correct answer: D

    NDCG weights top-ranked results more heavily, which matches ranking goals. Regression errors and row counts do not evaluate ranking quality.

Data Preparation for Machine Learning (3 questions)

  1. Question 16

    A dataset has missing values in a numeric column. Which GPU-accelerated preprocessing step handles this before modeling?

    1. AReplacing missing values with the literal text 'unknown' in a numeric column
    2. BIgnoring missing values because the GPU kernels handle them natively
    3. CDeleting the entire dataset whenever a column has a missing value
    4. DImputing missing values using a strategy such as the median computed on the GPU
    Show answer and explanation
    Correct answer: D

    Imputation with a suitable strategy keeps numeric features usable. Deleting the dataset, text placeholders in numeric columns and ignoring missing values are incorrect or harmful.

  2. Question 17

    Categorical features must be encoded for a GPU model. Which approach keeps the feature space manageable for high-cardinality columns?

    1. ATarget encoding or hashing with cross-validation to prevent leakage
    2. BDropping every categorical feature during preprocessing
    3. COne-hot encoding every unique value without a cap on width
    4. DAssigning random numbers to categories with an unseeded generator on each run
    Show answer and explanation
    Correct answer: A

    Target encoding with cross-validation or hashing manages high cardinality while reducing leakage. Unlimited one-hot encoding explodes dimensions, random numbers vary between runs, and dropping features loses information.

  3. Question 18

    A feature engineering step computes statistics using the full dataset before splitting into training and validation sets. What problem does this create?

    1. AFaster training with identical validation results across folds
    2. BNo problem because the statistics are usually neutral
    3. CData leakage that inflates validation performance
    4. DSmaller final model size and fewer learned parameters
    Show answer and explanation
    Correct answer: C

    Computing statistics on the full dataset leaks validation information into training, which inflates results. The other statements are incorrect.

Graph Analytics and Optimization (3 questions)

  1. Question 19

    A team wants to find the shortest paths across a network of millions of nodes on the GPU. Which library fits best?

    1. AcuDF only
    2. BcuGraph
    3. CcuML for clustering
    4. DRiva
    Show answer and explanation
    Correct answer: B

    cuGraph provides GPU-accelerated graph algorithms including shortest paths. cuDF handles dataframes, cuML focuses on machine learning, and Riva is for speech.

  2. Question 20

    A fraud detection workflow identifies communities of accounts that transact with each other. Which GPU algorithm family applies?

    1. ASpeech recognition with RNNs
    2. BRay tracing for realistic rendering
    3. CImage segmentation of medical scans
    4. DCommunity detection in cuGraph
    Show answer and explanation
    Correct answer: D

    Community detection finds densely connected groups, which fits this scenario. Image segmentation, speech recognition and ray tracing are unrelated.

  3. Question 21

    A logistics team needs to solve a vehicle routing problem with many constraints quickly. Which NVIDIA library is designed for this?

    1. AcuGraph only
    2. BNeMo Curator
    3. CcuDF
    4. DcuOpt
    Show answer and explanation
    Correct answer: D

    cuOpt is designed for optimization problems such as routing with constraints. cuDF handles dataframes, cuGraph does graph analysis, and Curator prepares text data.

Dask and Distributed Computing (3 questions)

  1. Question 22

    A Dask cluster runs on several GPU nodes. Which component coordinates the tasks and tracks where data lives?

    1. AThe Dask scheduler
    2. BThe browser cache
    3. CThe GPU driver
    4. DThe file system's mount table
    Show answer and explanation
    Correct answer: A

    The Dask scheduler coordinates tasks and worker data placement. Drivers, mount tables and browser caches do not coordinate computation.

  2. Question 23

    A Dask computation is slow because a single very large partition is processed by one worker. What is the best fix?

    1. AAdding more memory to the scheduler process, while workers stay unchanged
    2. BRepartitioning the data into more balanced partitions across workers
    3. CConverting the data to a single gzip-compressed text file
    4. DRunning the computation on the client, collecting partitions locally first
    Show answer and explanation
    Correct answer: B

    Balanced partitions let all workers contribute. Scheduler memory, text files and client-only execution do not parallelize the large partition.

  3. Question 24

    A workflow must use lazy evaluation so that Dask can optimize the full graph before running. Which behavior is expected?

    1. AEvery operation executes immediately as soon as it is written, materializing data
    2. BOperations do not run, because execution is deferred indefinitely until the cluster is shut down
    3. COperations build a task graph and execute only when a result is requested with compute
    4. DOperations execute on the scheduler's laptop, with each worker's results streamed back to it
    Show answer and explanation
    Correct answer: C

    Lazy evaluation builds a graph and computes when needed, which allows optimization. Immediate execution, no execution and laptop-only execution do not match Dask's model.

Spark and Accelerated Analytics (3 questions)

  1. Question 25

    A Spark job should run faster by using GPUs for supported SQL and DataFrame operations. Which accelerator is designed for this?

    1. AA Spark cluster with only CPU executors
    2. BA text editor plugin
    3. CThe RAPIDS Accelerator for Apache Spark
    4. DA web browser extension
    Show answer and explanation
    Correct answer: C

    The RAPIDS Accelerator offloads supported Spark operations to GPUs. CPU-only clusters, editor plugins and browser extensions do not accelerate Spark workloads.

  2. Question 26

    A GPU-accelerated Spark job falls back to the CPU for some operations. Which tool helps identify which operations were not accelerated?

    1. AThe job's logo, the cluster's color theme, and the Spark web UI footer text
    2. BThe explain output or qualification report showing unsupported operations
    3. CThe cluster's time zone and the default locale setting used by the Python driver process
    4. DThe number of Spark notebooks open in the workspace last quarter
    Show answer and explanation
    Correct answer: B

    Qualification and explain outputs identify unsupported operations that fall back to the CPU. Logos, time zones and notebook counts do not identify fallbacks.

  3. Question 27

    A Spark pipeline spends much time shuffling data between executors. Which change most helps?

    1. AWriting every intermediate result to a text file on disk between stages, then reading each file back
    2. BDisabling partitioning and letting a single executor hold the entire dataset in memory
    3. CIncreasing the number of shuffle operations by splitting each stage further
    4. DReducing shuffle volume through better partitioning and filtering before wide transformations
    Show answer and explanation
    Correct answer: D

    Filtering early and partitioning well reduce shuffle volume. More shuffles, text files and disabled partitioning increase cost.

Data Visualization and Reporting (3 questions)

  1. Question 28

    A team needs to visualize millions of points from a GPU-processed dataset interactively. Which approach is most appropriate?

    1. AAggregating or sampling the data on the GPU before plotting a summarized view
    2. BPlotting every point in a static image of the whole dataset at once
    3. CPrinting the points as a table with one row per record and every column
    4. DOpening each point in a new window with its own detail panel and tooltip on click
    Show answer and explanation
    Correct answer: A

    Aggregation or sampling on the GPU produces a responsive, informative view. Full static plots, tables and per-point windows do not scale.

  2. Question 29

    A chart must reflect data that updates every minute from a GPU pipeline. Which practice keeps the visualization current without overloading the system?

    1. AReloading the full raw dataset from cold storage every second
    2. BFreezing the chart permanently after the first render so updates stop
    3. CShowing data from last year, using a snapshot taken during the previous fiscal cycle
    4. DRefreshing the aggregated results on a fixed schedule with incremental updates
    Show answer and explanation
    Correct answer: D

    Scheduled, incremental aggregation keeps the chart current efficiently. Full reloads every second overload the system, and frozen or outdated data fails the requirement.

  3. Question 30

    A dashboard shows model predictions and their accuracy over time. Which practice helps detect degradation early?

    1. AHiding accuracy metrics from stakeholders and business owners
    2. BTracking accuracy metrics on recent data against a baseline and alerting on drops
    3. CShowing the first day's predictions as the sole baseline for later weeks
    4. DUsing a chart type that hides trends in a cumulative stacked area view
    Show answer and explanation
    Correct answer: B

    Tracking accuracy against a baseline with alerts reveals degradation. Showing only early data, hiding metrics and obscuring trends prevent early detection.

Want more NCP-ADS questions?

The full bank has 100 questions with answers and explanations, as a PDF, an online practice exam, or both.

Full question bank · 100 questions
$89$124

Both formats together. The best value.

Instant access after payment, with 90 days of updates. Checkout is on our store, AIExamCerts.