NCP-ADS 시험 - NVIDIA실제시험문제와 답 - 300문항

Question No : 1

You are working on a data science project where you need to process a large dataset containing 500 million records. You want to determine whether GPU acceleration would significantly improve performance.
Which of the following factors best indicates that you should use an accelerated computing solution like RAPIDS?

A.The dataset is a structured table with less than 100,000 records and can be handled efficiently with a Pandas DataFrame.
B.The dataset has high-dimensional sparse features and requires complex operations such as nearest neighbor search and clustering.
C.The dataset is heavily structured but mainly requires text-based analysis using regex-based search and manipulation.
D.The dataset consists of simple arithmetic operations on a few columns and can be processed using vectorized NumPy operations.

정답:

Question No : 2

You are training a deep learning model on a large dataset. Initially, you train the model on a single GPU and achieve a training time of 10 hours. To speed up training, you switch to a multi-GPU setup with four GPUs. However, after testing, you notice that the training time is only reduced to 3.5 hours instead of the expected 2.5 hours (a linear speedup).
What is the most likely reason for this sublinear speedup?

A.The optimizer is not designed for multi-GPU training, causing inefficiency.
B.The learning rate is too low, causing slower convergence despite multiple GPUs.
C.Increased memory bandwidth usage causes a bottleneck in the system.
D.Data transfer and communication overhead between GPUs limit scalability.

정답:

Question No : 3

You have a pandas DataFrame with a column containing floating-point numbers, but it takes up too much memory. You want to convert it into a lower-precision type using CuDF or pandas while ensuring computational efficiency.
Which function would you use?

A.df.to_float16()
B.df.astype('float16')
C.df['col'].apply(lambda x: np.float16(x))
D.df.convert_dtypes()

정답:

Question No : 4

You are tasked with designing an ETL workflow for a large-scale data processing pipeline using NVIDIA technologies. You need to ensure that the extraction, transformation, and loading phases are optimized for performance using hardware acceleration.
Which of the following NVIDIA technologies would be most suitable for accelerating the ETL process?

A.TensorFlow
B.CUDA
C.TensorRT
D.RAPIDS

정답:

Question No : 5

You are tasked with profiling a deep learning model using NVIDIA’s DLProf to identify performance bottlenecks and optimize resource utilization.
Which of the following statements correctly describes the capabilities of DLProf?

A.DLProf requires significant modifications to the source code to collect profiling data.
B.DLProf is primarily designed for debugging model accuracy rather than performance analysis.
C.DLProf can generate detailed reports that highlight kernel-level execution times and GPU utilization trends.
D.DLProf only works with TensorFlow models and does not support PyTorch-based workloads.

정답:

Question No : 6

A machine learning team is handling large-scale datasets that need to be efficiently stored and accessed within an NVIDIA RAPIDS workflow.
Which of the following storage formats and techniques provides the best performance for GPU-based data science pipelines?

A.Store data in the Apache Parquet format and load it directly into cuDF using cuDF’s read_parquet() method.
B.Store data in SQLite databases and query it into pandas before converting it to cuD
C.Use CSV files for storage, as they are widely compatible and can be read quickly using cuDF’s read_csv() method.
D.Save data in JSON format to ensure hierarchical relationships and flexibility before loading it into cuD

정답:

Question No : 7

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

A.Lower energy consumption compared to CPUs
B.Better for memory-intensive workloads
C.Faster parallel processing for large datasets
D.Lower cost of cloud infrastructure
E.Efficient handling of matrix operations in machine learning models

정답:

Question No : 8

df['category'] = df['category'].astype('category')

정답: D

Question No : 9

You need to determine the optimal data processing library for a small dataset of 500,000 records that will be processed on a multi-core CPU machine with no GPU access.
Which of the following libraries would be the most efficient for this task?

A.cuDF
B.Dask
C.pandas
D.PyTorch

정답:

Question No : 10

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.

정답:

Question No : 11

You are using NVIDIA DLProf to analyze the performance of a deep learning model deployed on an A100 GPU. The report indicates that compute-bound operations are dominating execution time, and kernel execution efficiency is below 50%.
What is the best action to take based on this insight?

A.Increase the batch size to fully utilize available GPU memory and reduce per-sample processing overhead.
B.Reduce the number of layers in the model to decrease computation time.
C.Use DLProf’s Tensor Core Analysis to check if the model is leveraging Tensor Cores effectively.
D.Enable mixed precision training to improve computational efficiency.

정답:

Question No : 12

You are tasked with optimizing the performance of an MLOps pipeline that uses GPU-accelerated workflows. After running initial benchmarks, you notice that the training time is higher than expected, despite the use of multiple GPUs.
What are the best strategies to optimize the GPU-accelerated workflow in this case? (Select two)

A.Reduce the number of GPUs used and focus on fine-tuning the hyperparameters for optimal performance on a single GP
B.Ensure that the model is distributed evenly across GPUs to prevent some GPUs from being underutilized.
C.Ensure efficient multi-GPU communication and synchronization strategies, such as using NCCL for distributed training.
D.Increase the batch size to better utilize the multiple GPUs and reduce the number of updates to the model during training.
E.Disable gradient accumulation when using multi-GPU setups to increase communication efficiency.

정답:

Question No : 13

You have a large-scale dataset consisting of IoT sensor readings collected at one-minute intervals across multiple locations. The dataset contains missing values and requires scaling before applying a machine learning model. You plan to use NVIDIA RAPIDS to preprocess and analyze the time-series data efficiently on GPUs.
Which of the following preprocessing steps is the most efficient approach using NVIDIA RAPIDS?

A.Use Dask for distributed missing value imputation and train a model using TensorFlow's CPU-based estimator.
B.Use pandas for missing value imputation, then normalize the data using NumPy before converting to cuD
C.Use pandas to fill missing values and scale the data, then convert it to cuDF for training.
D.Use cuDF to handle missing values with GPU-accelerated interpolation and apply cuML's StandardScaler for feature scaling.

정답:

Question No : 14

A data scientist is working with large-scale ETL (Extract, Transform, Load) pipelines on GPU-accelerated infrastructure using RAPIDS. The workload involves frequent shuffle operations, which significantly impact performance.
What is the best approach using NVIDIA technologies to reduce shuffle overhead and improve performance?

A.Use RAPIDS cuDF’s GPU memory caching to store intermediate DataFrames and avoid redundant shuffle operations.
B.Store all intermediate shuffle data in CPU memory using pandas to ensure persistence and reduce GPU load.
C.Use RAPIDS cuML to replace shuffle-intensive operations with an ML model that predicts data distribution.
D.Enable CUDA Unified Memory to automatically optimize shuffle performance without manual intervention.

정답:

Question No : 15

You are working on a dataset containing missing values, duplicate records, and inconsistent data types.
The dataset size is 15GB and you need to efficiently perform data cleansing operations such as:
- Handling missing values
- Dropping duplicates
- Converting data types
Which of the following approaches would be the most efficient way to perform these operations on an NVIDIA GPU?

A.Use Vaex to perform data cleansing, as it is optimized for large-scale datasets
B.Use pandas to load the dataset and perform the operations using standard pandas functions
C.Convert the dataset into a NumPy array and process it using CuPy before converting it back to a DataFrame
D.Load the dataset using cuDF, then use cuDF’s built-in .dropna(), .drop_duplicates(), and .astype() methods

정답:

NVIDIA NCP-ADS 시험