NVIDIA Certified Professional - Accelerated Data Science practice questions

Question 1

An architect is evaluating how to bring GPU acceleration to a large existing pandas codebase and must choose between cuDF's two usage modes. Which statement correctly distinguishes them?

Accepted Answer

cudf.pandas accelerates existing pandas with zero code changes, alongside cuDF's existing GPU-only API for native code.

Answer

core cuDF and cudf.pandas are completely identical, so both require rewriting every single pandas call into cuDF syntax.

Answer

core cuDF runs only on the CPU, while cudf.pandas is the single mode that ever touches the GPU at all here.

Answer

cudf.pandas requires manually porting each function to cupy, whereas core cuDF runs unmodified pandas as-is.

Question 2

An MLOps team is moving a distributed cuML model trained on a multi-GPU Dask cluster into production and wants inference to fan out across the same Dask cluster rather than a single GPU. According to the cuML persistence workflow, which approach distributes the inference step?

Accepted Answer

Load the saved combined model and wrap it with ParallelPostFit from Dask-ML for distribution

Answer

Load the combined model on one GPU and replicate the Dask process group across the cluster by hand

Answer

Serve the cuml.dask wrapper through Triton's sequence batcher to spread load across the GPUs

Answer

Re-extract the model with get_combined_model() on every node before each prediction request

Question 3

A company is cleansing a staging cuDF DataFrame and uses a .empty check to skip frames with no usable data, but a frame holding only null values is still being processed. Per the cuDF pandas compatibility notes, why is that?

Accepted Answer

A cuDF DataFrame or Series that contains only null values is still not considered empty by the .empty property.

Answer

The .empty property in cuDF only returns True after the index is explicitly reset, which the cleansing step skipped.

Answer

cuDF treats null-only frames as empty, so the real cause must be a duplicate frame slipping past the .empty filter here.

Answer

The .empty check is unsupported in cuDF and always returns False, so it can never be used to skip any frame at all.

Question 4

A production team is enabling UCX in Dask-CUDA to get NVLink or InfiniBand transport for shuffles. According to the Dask-CUDA UCX integration docs, what hardware prerequisite must be in place?

Accepted Answer

GPUs connected with NVLink bridges or NVIDIA Mellanox InfiniBand Adapters respectively.

Answer

A USB-3 link between every pair of GPUs to enable UCX shuffles across the cluster.

Answer

Plain Wi-Fi 6 between worker nodes is sufficient; UCX requires no specialised interconnect.

Answer

GPUs virtualised across two AWS regions before UCX can transport any shuffle data.

Question 5

A production team has four GPUs on a training node and must pin XGBoost training to the second GPU only. According to the XGBoost GPU documentation, what is the documented syntax for selecting that specific device ordinal?

Accepted Answer

Use the cuda:<ordinal> syntax, so the second GPU is selected by setting device='cuda:1'.

Answer

Set device='gpu_1' as the boolean device selector; the ordinal is then implicit on call.

Answer

Set device='cuda' and additionally invoke nvidia-smi -i 1 inside the training script.

Answer

Set device=1 as a bare integer; the XGBoost docs do not mention any cuda:<ordinal> form.

Question 6

A developer is preparing a multi-gigabyte dataframe of string-valued labels before saving it for a Datashader pipeline and needs a more memory-efficient column representation. According to the Performance guide, how should the label column be stored?

Accepted Answer

Convert the column to a categorical dtype, which uses a more memory-efficient representation than object strings

Answer

Hash each label to a 64-bit integer and drop the strings, rebuilding the text lazily only when plotted

Answer

Keep the labels as object strings and simply compress the saved file, since the column layout will not matter here

Answer

Cast the labels to 32-bit floats, which halves the memory of the default 64-bit numeric column storage

Question 7

A company is deciding at what data scale to adopt cudf.pandas for its analytics jobs and wants a documented rule of thumb for when GPU acceleration starts to pay off. Which guidance matches the cuDF FAQ?

Accepted Answer

It shines on workflows above about 10,000 to 100,000 rows

Answer

Benefits appear only past roughly one billion rows of data

Answer

Acceleration is guaranteed for any workload at any row count

Answer

Speedups need CPU fallback disabled so all ops run on GPU

Question 8

A team has decided to serve an XGBoost gradient-boosted tree model on Triton. They also have LightGBM and cuML random forest models in the pipeline. Which Triton backend is specifically designed for these tree-based models?

Accepted Answer

The FIL backend, which serves tree models from XGBoost, LightGBM, sklearn, and cuML.

Answer

The TensorRT-LLM backend, designed for LLM serving with KV cache optimization paths.

Answer

The Python backend, where each tree must be hand-coded as a Python function per request.

Answer

The OpenVINO backend, which only accepts Intel OpenVINO IR-format model files for trees.

Question 9

A platform team is deciding whether to install Magnum IO GPUDirect Storage manually for a RAPIDS deployment. According to the cuDF I/O docs, what is the practical packaging note the team should know?

Accepted Answer

GDS is also included in CUDA Toolkit alongside the standalone SDK download path.

Answer

GDS only ships as source code and must be built from a private NVIDIA build farm.

Answer

GDS bundles exclusively with Triton Inference Server containers and ships nowhere else.

Answer

GDS is a cloud-only managed service available only on AWS Inferentia accelerators.

Question 10

An MLOps team is migrating its RAPIDS model training to Amazon SageMaker and wants to avoid paying for an idle GPU on its notebook instance. Using SageMaker Estimators as described in the RAPIDS deployment docs, how is GPU billing handled?

Accepted Answer

Estimators run on ephemeral managed VMs, so you are billed only for the GPU instances for the time the training job runs.

Answer

The notebook instance itself must be a GPU instance, and it is billed continuously the entire time the notebook server stays running.

Answer

SageMaker charges a flat monthly fee for the training cluster, regardless of how long each individual job actually runs.

Answer

GPU charges accrue from the moment the Docker image is pushed to the registry until that image is manually deleted later.

Question 11

A platform team is rolling out cuml.accel across a shared cluster by exporting an environment variable in the base image, but some nodes keep running on CPU with no error. According to the Zero Code Change Acceleration documentation, what explains the silent CPU fallback?

Accepted Answer

If cuML is not installed properly, CUML_ACCEL_ENABLED is silently ignored and normal CPU execution occurs.

Answer

The accelerator only loads inside Jupyter kernels, so non-interactive batch nodes fall back to CPU execution.

Answer

cuml.accel requires a 2D PyTorch tensor input, and NumPy arrays on those nodes force a CPU code path entirely.

Answer

Setting the variable disables the GPU until cuda:0 is also exported, which those nodes are currently missing.

Question 12

A developer is building line plots in Datashader with Canvas.line() and needs to describe, for a design review, what geometric representation the library uses for each line by default.

Accepted Answer

A polyline: a connected set of straight line segments, the default representation Canvas.line() uses

Answer

A cubic Bezier spline fitted through the data points, which Canvas.line() smooths before it rasterizes them

Answer

A filled polygon bounded by the curve and the x-axis, so the shaded area under each line is aggregated

Answer

A kernel-density surface estimated from the points, which Canvas.line() blurs to approximate the line

Question 13

An engineer is designing a custom row-level transform on a cuDF Series that the built-in API cannot express. Beyond Series.apply, what does the cuDF UDF guide say can be used to operate directly on Series objects?

Accepted Answer

Pass Series objects directly into CUDA kernels written with Numba.

Answer

Push the Series back to a pandas DataFrame on the CPU and call .map with a lambda.

Answer

Train an XGBoost model on the Series and use its predictions as the transform output.

Answer

Serialise each row to JSON and call an LLM API to compute the transformed value.

Question 14

A team is running variable-length sequence inputs through Triton and wants dynamic batching, but the inputs rarely share the same shape so they are currently padding every tensor on the client. Which Triton feature lets them batch without that explicit padding?

Accepted Answer

Ragged batching, which avoids explicit padding by specifying which inputs vary

Answer

The response cache, which stores padded tensors and replays them across requests

Answer

Model control mode POLL, which reshapes inputs as the repository changes

Answer

Sequence batching with Model Analyzer auto-tuning the maximum batch dimension

Question 15

An mlops team is designing a target-encoding step for several high-cardinality categorical columns and is worried about label leakage inflating offline metrics. According to the cuML TargetEncoder documentation, how does the default configuration mitigate this?

Accepted Answer

It splits the training data into n_folds (default 4) and encodes each fold using the target values of the remaining folds.

Answer

It encodes every row using the global mean of the target across the entire dataset, ignoring fold structure entirely here.

Answer

It drops any category whose target mean exceeds the dataset average, removing the rows most likely to leak the signal.

Answer

It defers all encoding to inference time so the training set never sees any target-derived values during fitting at all.

NVIDIA Certified Professional - Accelerated Data Science

Sample questions

Sources

Similar exams