MLS-C01 Question 167
Single answerYou are tasked with training a deep learning model for image classification on a dataset containing millions of high-resolution images. The training requires significant computational resources and is expected to take several days. Additionally, you need to ensure that the training process can scale efficiently as your dataset grows. Which compute resources should you choose to optimize both performance and cost?
- A
Use CPU-based instances with multiple cores in an Auto Scaling group for distributed training
- B
Use a single GPU instance with a high memory-to-GPU ratio for local training
- C
Use GPU-based instances with distributed training across multiple instances using Amazon SageMaker
- D
Use CPU-based instances with Amazon EC2 Spot Instances to reduce cost
Show answer and explanation
Correct answer: C
Explanation
For deep learning tasks involving large-scale datasets, GPUs are the preferred compute resource due to their ability to handle parallel computations efficiently. Distributed training across multiple GPU instances further reduces training time and enhances scalability. Amazon SageMaker provides built-in support for distributed training, making it an ideal choice for this scenario. CPUs, even when scaled or cost-optimized with Spot Instances, are not suitable for such compute-intensive tasks.
- A. Incorrect.
CPU-based instances with multiple cores are not well-suited for deep learning tasks, especially for large image datasets, as GPUs are optimized for parallel matrix computations required in training neural networks.
- B. Incorrect.
A single GPU instance might not be sufficient for this large-scale dataset, as it will significantly increase training time and limit scalability. While it could work for smaller datasets, it is not cost-efficient or performant for this case.
- C. Correct.
GPU-based instances with distributed training across multiple instances using Amazon SageMaker is the optimal choice. GPUs are specifically designed for deep learning workloads, and SageMaker facilitates distributed training, allowing the workload to scale efficiently across multiple instances, reducing training time.
- D. Incorrect.
While using CPU-based Spot Instances might reduce cost, CPUs are not efficient for training deep learning models, especially for tasks like image classification with high-resolution datasets. This would lead to significant delays in training.