NCP-ADS問題集、NVIDIA実際の試験問題

質問 1

You are working on a large-scale graph analysis problem that involves computing the shortest paths between nodes in a massive social network dataset. You decide to leverage NVIDIA RAPIDS cuGraph for accelerated computation.
Which of the following cuGraph functions should you use?

A. cugraph.label_propagation()

B. cugraph.sssp()

C. cugraph.pagerank()

D. cugraph.k_truss()

正解: B

質問 2

You are training a large-scale random forest model on a dataset with millions of rows and hundreds of features. The training time is significantly high when using traditional CPU-based machine learning frameworks.
Which NVIDIA technology should you use to accelerate training while maintaining compatibility with common ML frameworks like scikit-learn?

A. NVIDIA DeepStream to preprocess tabular data and optimize random forest model execution.

B. NVIDIA TensorRT to accelerate random forest model training by optimizing tree-based algorithms.

C. NVIDIA Triton Inference Server to distribute random forest model training across multiple GPUs.

D. NVIDIA RAPIDS cuML to accelerate random forest training using GPU-optimized implementations.

正解: D

質問 3

You are training a machine learning model using NVIDIA RAPIDS cuML and notice that the training process is significantly slower than expected. You suspect that there are bottlenecks in data movement and computation.
Which of the following techniques can best help you diagnose and resolve these bottlenecks?

A. Move all data from GPU memory to CPU memory before training the model.

B. Use cudf.DataFrame.to_pandas() to convert the dataset to a pandas DataFrame for analysis.

C. Reduce the number of features used in training without profiling the actual bottlenecks.

D. Use cuml.common.device_auto_mem_size() to check GPU memory usage and adjust batch sizes accordingly.

正解: D

質問 4

You are processing a large dataset using RAPIDS cuDF and Dask-cuDF on an NVIDIA GPU. Your profiling indicates that data transfer times between CPU and GPU are significantly slowing down your pipeline.
What is the most effective way to reduce this bottleneck?

A. Convert the dataset into a CSV format before transferring it to the GPU

B. Use cudf.read_parquet() instead of Pandas to load data directly into GPU memory

C. Increase the CPU RAM allocation to store more data before transferring to the GPU

D. Transfer data in multiple smaller chunks to the GPU instead of larger batches

正解: B

質問 5

A data scientist is processing a dataset that is too large to fit into the memory of a single GPU. They decide to use Dask with cuDF to leverage multiple GPUs for accelerated computation.
Which of the following approaches ensures efficient parallelism when working with dask_cudf?

A. Use Dask futures to asynchronously submit GPU tasks and collect results when completed.

B. Use a single-threaded scheduler to avoid contention when running on multiple GPUs.

C. Assign explicit chunks of data to each worker by manually dividing the dataset before passing it to Dask.

D. Convert the dataset into a Dask cuDF dataframe and use persist() to keep it in GPU memory.

正解: D

質問 6

You are designing an ETL workflow to process large-scale financial transaction data using GPU acceleration. The dataset is stored in a Parquet file and contains millions of records.
Which of the following approaches is the most efficient for performing extract, transform, and load (ETL) operations using NVIDIA RAPIDS technologies?

A. Use Pandas DataFrame for transformation, and then convert the dataset to cuDF before writing to storage.

B. Load the Parquet file directly into a cuDF DataFrame and use cuDF's built-in functions for transformations before writing the results back to storage.

C. Use Apache Spark with CPU-based processing for ETL, then convert the results into cuDF for accelerated analytics.

D. Store all data as CSV files and perform ETL operations using traditional row-based processing.

正解: B

質問 7

In the context of cloud computing, what are the key benefits of using GPUs for data science tasks?
(Select two)

A. Lower cost of cloud infrastructure

B. Faster parallel processing for large datasets

C. Lower energy consumption compared to CPUs

D. Efficient handling of matrix operations in machine learning models

E. Better for memory-intensive workloads

正解: B,D

質問 8

You are working on a data science project that requires augmenting a dataset using synthetic data.
You are utilizing cuDF and NVIDIA RAPIDS to speed up the data generation process.
Which of the following methods is the most effective way to generate synthetic data using cuDF in a RAPIDS workflow?

A. Use cudf.DataFrame.applymap() to create new synthetic features through complex mathematical functions.

B. Use cudf.DataFrame.sample() to duplicate random rows and create synthetic data.

C. Use cuDF to manipulate data distributions and generate new data points based on existing features.

正解: C

質問 9

Which of the following Nvidia technologies is primarily used for performing benchmarking and optimizing GPU-accelerated deep learning workflows, especially focusing on model training performance?

A. Nvidia DeepStream SDK

B. Nvidia Triton Inference Server

C. Nvidia CUDA

D. Nvidia Nsight Systems

正解: D

質問 10

You are working on a large dataset (several terabytes in size) and need to perform data preprocessing, filtering, and transformations before training a machine learning model.
Given the dataset size and the requirement to optimize for GPU acceleration using NVIDIA technologies, which of the following is the most appropriate data processing library to use?

A. NumPy with CuPy acceleration

B. pandas

C. Modin with Ray backend

D. Dask DataFrame with Dask-CUDA

正解: D

質問 11

You are building a real-time recommendation system that processes high-frequency transactional data from millions of users.
The system must:
- Ingest and preprocess data efficiently
- Perform similarity computations for user-item recommendations
- Scale to handle rapid incoming transactions
Which of the following NVIDIA technologies is the best choice for this use case?

A. RAPIDS cuGraph

B. CUDA Kernels with Custom C++ Code

C. NVIDIA Triton Inference Server

D. NVIDIA NVTabular

正解: D

質問 12

A data engineer is preparing a dataset for training a deep learning model. The dataset contains numerical features with missing values, outliers, and inconsistent units.
Which of the following strategies is the most appropriate for ensuring a standardized and clean dataset?

A. Standardize the dataset using the mean and standard deviation, but keep missing values and outliers unchanged to avoid data manipulation.

B. Remove all rows with missing values and outliers to ensure only clean data is used.

C. Replace missing values with the mean, apply z-score normalization, and clip extreme outliers based on a threshold (e.g., 3 standard deviations).

D. Use the median to fill missing values, convert all numerical values into categorical bins, and apply Min-Max scaling.

正解: C

質問 13

You are comparing the performance of NVIDIA RAPIDS cuML, TensorFlow, and PyTorch for training and inference on a dataset with millions of records.
To design a fair and effective benchmark, which approach should you take?

A. Run each framework on different GPUs to maximize available resources and compare execution times across different hardware configurations.

B. Ensure all frameworks run on the same GPU, use optimized batch sizes, and measure execution time and memory usage with NVIDIA Nsight Systems.

C. Use only a CPU baseline for comparison to demonstrate the benefits of GPU acceleration, ignoring GPU-specific optimizations.

正解: B

質問 14

You have a massive time-series dataset containing millions of records per day, and you need to perform forecasting at scale.
Which of the following techniques best utilizes NVIDIA technologies to optimize time-series forecasting?

A. Convert time-series data to tensors and train a deep learning model using PyTorch with mixed precision on an NVIDIA GPU.

B. Use RAPIDS cuDF and Dask-cuDF to preprocess the data and train an LSTM model with PyTorch Lightning on multiple GPUs.

C. Use RAPIDS cuML's XGBoost-GPU implementation to forecast time-series patterns with improved training efficiency.

D. Run Facebook Prophet on a multi-core CPU environment to leverage parallel forecasting.

正解: B

質問 15

After profiling a deep learning model using NVIDIA DLProf, you notice that a specific GEMM (General Matrix Multiplication) operation takes significantly longer than expected. The profiler output reveals that tensor cores are underutilized despite having an Ampere-based GPU with Tensor Cores enabled.
Which of the following actions is the MOST appropriate to improve performance?

A. Convert the model's data type to float16 or bfloat16 and re-run the training with automatic mixed precision (AMP).

B. Increase the batch size to maximize GPU memory usage and reduce kernel launch overhead.

C. Switch from stochastic gradient descent (SGD) to Adam optimizer, as Adam improves convergence and computational efficiency.

D. Disable CUDA graphs and enforce PyTorch's eager execution mode to improve kernel execution order.

正解: A

NVIDIA-Certified-Professional Accelerated Data Science - NCP-ADS 模擬練習