NVIDIA Certified Associate - Accelerated Data Science practice questions

Question 1

A team is rolling out a cuDF refactor of a pandas pipeline and the lead asks which Pandas patterns they can keep using essentially unchanged. According to the cuDF comparison guide, which set of operations is supported?

Accepted Answer

Series, DataFrame, and Index data structures with familiar binary and unary operations.

Accepted Answer

Indexing, filtering, concatenating, joining, groupby, and window operations on those structures.

Answer

Built-in cross-cluster checkpointing that automatically restarts failed Dask schedulers.

Answer

Hand-written CUDA kernels emitted from arbitrary Python decorators at runtime.

Question 2

A developer is evaluating how to keep a single pipeline that mixes cuML and scikit-learn estimators while still being able to run it on machines without cuML. Which cuML serialization feature is designed for this hybrid workflow?

Accepted Answer

as_sklearn()/from_sklearn() conversion, which lets the pipeline mix both libraries without installing cuML on deployment hosts.

Answer

The cuGraph interoperability layer, which exports the trained pipeline out to NetworkX graphs so it can be replayed anywhere.

Answer

The cudf.pandas accelerator, which lets existing pandas code run completely unchanged on either a CPU or a GPU.

Answer

Setting n_streams to 1 on every estimator, which forces a single-stream execution that any CPU can later replay.

Question 3

An MLOps team is configuring distributed XGBoost training on GPUs with Dask and reviewing the official docs to confirm two behaviors before production. Which TWO statements are correct? (Select two.)

Accepted Answer

Setting the device parameter to cuda enables GPU acceleration for training

Accepted Answer

Constructing a DaskDMatrix forces earlier lazy Dask computations to be evaluated

Answer

The cuda:<ordinal> syntax can only ever target GPU 0 and ignores any other device

Answer

Dask computations run eagerly by default, so calling compute() is never required

Question 4

A developer is migrating a pandas exploration script onto the GPU and is worried about rewriting code. Given the constraints described in the RAPIDS cuDF guide, what is the key adoption advantage of cuDF over scaling out to Dask or Spark?

Accepted Answer

cuDF exposes a pandas-like API on the GPU, so the script ports with little rewriting, unlike Dask or Spark.

Answer

cuDF removes the need to load any data into memory, streaming every row directly from disk during each operation.

Answer

cuDF automatically shards the workload across a multi-node cluster, so no single GPU ever holds the full frame.

Answer

cuDF guarantees identical floating-point results to Spark by forcing every column into 64-bit precision on load.

Question 5

An architect is reviewing how cuDF implements the Apache Arrow columnar format internally so the team can reason about zero-copy interchange. According to the cuDF library design, which construct encapsulates cuDF's implementation of that data format?

Accepted Answer

The Column class, which encapsulates cuDF's own implementation of the Apache Arrow data format under the hood.

Answer

The DataFrame object, which stores all its rows contiguously and exposes the Arrow buffers through its row index.

Answer

The RMM pool, which lays out each Arrow buffer in device memory and hands columns to estimators on demand.

Answer

The Series wrapper, which converts Arrow record batches into NumPy arrays before any compute can be run.

Question 6

A platform team is rolling out monitoring for a Triton-served fraud model and wants Prometheus-style GPU and request metrics. According to the Triton metrics guide, how are those metrics exposed by default?

Accepted Answer

Triton provides Prometheus metrics on a default endpoint at http://localhost:8002/metrics in plain text format.

Answer

Triton pushes metrics to a remote SaaS service every five minutes by default.

Answer

Triton requires installing NIM Enterprise before any metrics endpoint is enabled.

Answer

Triton emits only OpenTelemetry traces; Prometheus scraping is not supported.

Question 7

A platform team is migrating a Pandas pipeline to cuDF and needs columns for exact decimal values, variable-length nested lists, and key/value records, beyond the dtypes cuDF shares with NumPy and Pandas. According to the cuDF supported data types guide, which special data types does cuDF additionally provide for this?

Accepted Answer

Special data types for working with decimals, list-like data, and dictionary-like data

Answer

Sparse, interval, and period data types inherited directly from the standard Pandas library

Answer

Tensor, ragged, and adjacency-list data types exposed through the cuGraph library

Answer

Fixed-size geometry, raster, and trajectory data types from the cuSpatial library

Question 8

A team is installing RAPIDS with conda when the solver reports that the machine's CUDA driver (11.8.0) is incompatible with the requested cuda-version (12.0), even though the team is confident the driver is actually new enough. According to the RAPIDS installation guide, what is the documented way to override conda's incorrect driver detection?

Accepted Answer

Set the CONDA_OVERRIDE_CUDA environment variable to the supported version

Answer

Downgrade the requested cuda-version so it matches the installed driver

Answer

Set the CONDA_OVERRIDE_GLIBC environment variable and then re-run the conda solve

Answer

Reinstall the CUDA driver from the runfile package before installing it

Question 9

An analytics team has decided to rewrite a pandas script as cuDF to reduce wall-clock time on a 200 GB CSV. Before deployment, they want to confirm which RAPIDS library actually owns the GPU DataFrame data structure that hosts the rows.

Accepted Answer

cuDF, a Python GPU DataFrame library built on the Apache Arrow columnar memory format.

Answer

Dask-CUDA, which extends Dask Worker classes but does not provide a DataFrame implementation.

Answer

cuGraph, a graph algorithm library that does not provide a tabular DataFrame structure.

Answer

cuxfilter, a cross-filter visualization framework rather than a dataframe engine.

Question 10

An engineer is tuning cuML RandomForestClassifier training and wants to raise the number of parallel streams used while the forest is being built. Which parameter controls this, according to the cuML documentation?

Accepted Answer

n_streams, which sets the number of parallel streams used while building the forest.

Answer

n_bins, which sets the maximum number of bins the split algorithm uses per feature.

Answer

min_samples_leaf, which sets the minimum number of samples allowed in each leaf node.

Answer

n_estimators, which sets how many individual decision trees are grown inside the forest.

Question 11

A team is being designed around an existing pandas batch pipeline and considering rewriting some steps with cuDF. Which set of pandas-compatible features does the cuDF/Pandas comparison guide promise will carry over to the GPU step?

Accepted Answer

Series, DataFrame, Index, plus operations like indexing, filtering, concatenating, joining, groupby, and window operations.

Answer

All pandas plotting and styling APIs ship with cuDF unchanged and run on the GPU.

Answer

Only DataFrame.head() and DataFrame.tail() are guaranteed; everything else must be rewritten manually.

Answer

Only Triton-served inference functions are accelerated; pandas operations stay on CPU forever.

Question 12

A startup is evaluating the original browser-based Crossfilter library and finds it bogs down on very large datasets. According to the cuxfilter documentation, what limitation of Crossfilter does cuxfilter overcome, and how?

Accepted Answer

Crossfilter keeps data in-memory in the client browser; cuxfilter moves it onto the GPU via the rapids.ai cuDF stack.

Answer

Crossfilter cannot draw any line charts; cuxfilter adds Canvas.line() with the ds.count() aggregator to fix the gap.

Answer

Crossfilter lacks a layout engine; cuxfilter's only improvement is the layout_array for arranging charts on a page.

Answer

Crossfilter uses 64-bit floats; cuxfilter halves memory by forcing every dimension down to 32-bit precision first.

Question 13

A team is evaluating the RAPIDS stack and wants to understand which Apache project's columnar memory format underpins cuDF. According to the cuDF docs, which foundation does cuDF build on?

Accepted Answer

cuDF is built on the Apache Arrow columnar memory format.

Answer

cuDF is built on the Apache Avro row format and not on any columnar memory layout.

Answer

cuDF is built directly on Apache Spark in-memory tables with no Arrow involvement.

Answer

cuDF is built on Apache Kafka brokered messages instead of an in-memory format.

Question 14

An MLOps team is being asked to choose a Triton model control mode that lets them roll out new model versions by changing files in the repository, and have the server scan for those changes on a schedule rather than via explicit API calls. According to the model management guide, which control mode fits?

Accepted Answer

POLL mode, one of NONE, EXPLICIT, or POLL listed in Triton's model control modes.

Answer

STREAM mode, a hidden mode that polls the repository and is enabled by default.

Answer

OPTIMIZE mode, which auto-promotes any new model regardless of validation.

Answer

NIM_ONLY mode, the only mode Triton actually exposes for production environments.

Question 15

A developer is rolling out copy-on-write in a cuDF pipeline and is reading the cuDF copy-on-write guide. According to that guide, when should the team enable the option for it to behave consistently across all objects?

Accepted Answer

Set copy-on-write at the beginning of the script so all objects share the new behavior consistently.

Answer

Toggle copy-on-write in the middle of the script so older objects keep their original semantics.

Answer

Leave copy-on-write enabled by default; the guide says it is on by default in cuDF.

Answer

Disable copy-on-write only after writing the final output, because the guide forbids it earlier.

NVIDIA Certified Associate - Accelerated Data Science

Sample questions

Sources

Similar exams